Earthing switch knife switch state monitoring system and method based on AI image recognition

Abstract

The invention discloses a grounding switch disconnecting link state monitoring system and method based on AI image recognition, and relates to the technical field of power monitoring, and the system comprises an image collection module, an image preprocessing module, an AI image recognition module, a state judgment and early warning module, a communication module and a database. According to the system, images are collected through a high-definition camera in a timed or event triggering mode, after enhancement and denoising, the images are input into a YOLOV8 model for state recognition, and abnormal early warning is achieved in combination with a preset threshold value. Compared with a traditional detection scheme, according to the scheme, fine features of the disconnecting link are recognized through the AI image recognition technology, the misjudgment rate is remarkably reduced, the YOLOV8 algorithm responds rapidly, the rapid monitoring requirement of an electric power system is met, a multi-sensor scheme can be replaced with only a small number of cameras, hardware and operation and maintenance cost is greatly reduced, and the detection efficiency is improved. The problems that a traditional method is easily interfered by the environment and the hardware failure rate is high are effectively solved, and reliable guarantee is provided for safe operation of an electric power system.

Description

Technical Field

[0001] This invention relates to the field of power monitoring technology, specifically to a grounding switch status monitoring system and method based on AI image recognition. Background Technology

[0002] In the stable operation and efficient maintenance of modern power systems, grounding switches play a crucial role. As a key component for ensuring the safety of power equipment maintenance and maintaining the stability of the power system, the status monitoring of grounding switches is directly related to the reliability of power supply and the safety of personnel operations.

[0003] In existing technologies, the status of grounding switch disconnectors is commonly detected through manual inspection, auxiliary contact detection, or sensor detection. However, each of these methods has its own problems: Manual inspection relies on visual observation of the switch's position, which is time-consuming, labor-intensive, and inefficient. Furthermore, manual judgment is susceptible to environmental factors, leading to misjudgments. Auxiliary contact detection infers the status of the main disconnector by detecting the opening and closing of auxiliary contacts, but these contacts are prone to poor contact due to frequent operation and oxidation, resulting in signal transmission errors. Additionally, the installation and maintenance of auxiliary contacts are complex, increasing costs. Sensor detection uses multiple sensors, such as position sensors and current sensors, to detect the status of the grounding switch disconnector. However, position sensors may introduce measurement errors due to mechanical vibration and displacement, current sensors lack accuracy under low current signals, and multi-sensor integration is complex and costly. In summary, existing detection methods are insufficient in terms of accuracy, reliability, and cost-effectiveness, necessitating a practical and precise technical solution for automatic, real-time, and accurate determination of the grounding switch disconnector's status.

[0004] Based on this, we now provide a grounding switch status monitoring system and method based on AI image recognition, which can eliminate the drawbacks of existing technical solutions. Summary of the Invention

[0005] The purpose of this invention is to provide a grounding switch status monitoring system and method based on AI image recognition, so as to solve the problems of the existing detection methods in the background art in terms of accuracy, reliability and cost-effectiveness.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A grounding switch status monitoring system and method based on AI image recognition, comprising:

[0008] The image acquisition module is equipped with multiple high-definition cameras and is used to trigger image data acquisition operations based on preset time intervals and specific events. The preset time interval is set to 5 minutes, and the specific events include operation command issuance and power failure alarm. The image data is all stamped with timestamps and device identification information.

[0009] The image preprocessing module is used to perform image enhancement, noise reduction, and grayscale processing on the acquired image data.

[0010] The AI ​​image recognition module, based on the YOLOV8 model, is used to detect the position of the grounding switch knife switch based on the preprocessed image data and perform state classification operations. The YOLOV8 model includes a CSPDarknet53 backbone network, a spatial pyramid pooling module, and a path aggregation network structure.

[0011] The status judgment and early warning module is used to receive the switch status recognition results output by the AI ​​image recognition module, compare them with the thresholds and patterns preset in the normal state of the system, and trigger an early warning when there is an abnormality.

[0012] The communication module is used to enable data interaction and transmission between various modules;

[0013] The database is used to store image data, recognition results, and historical status records, supporting data traceability and analysis.

[0014] Preferably, the image acquisition module adopts a multi-view synchronous acquisition strategy, with multiple high-definition cameras respectively arranged at different views of the grounding switch. The views include front view, side view, top view, and bottom view. All high-definition cameras support acquisition modes triggered by timed events and triggered by specific events.

[0015] Preferably, the image preprocessing module includes:

[0016] The image enhancement unit employs a multi-scale Retinex and adaptive histogram equalization fusion algorithm to perform overall and local illumination optimization and contrast enhancement operations on image data.

[0017] The image denoising unit employs a fusion algorithm of bilateral filtering and nonlocal mean filtering to remove noise while preserving edge details of the image data.

[0018] The image grayscale unit executes a weighted conversion formula from color image to grayscale image, converting color image data into single-channel grayscale image data.

[0019] Preferably, the CSPDarknet53 backbone network is used to extract multi-scale features from the input grounding switch image. By progressively downsampling to generate feature maps of different resolutions, multi-level features of the switch are extracted. The CSPDarknet53 backbone network adopts a cross-stage partial connection structure. By dividing the feature map output from the basic convolutional layer into two parts, one part is subjected to deep feature extraction through dense stacking of residual blocks, and the other part is directly fused with the residual block output features across stages to form a multi-scale feature pyramid.

[0020] Preferably, the spatial pyramid pooling module is connected to the output of the CSPDarknet53 backbone network. It extracts global features of image data through multi-scale max pooling operations, enhancing the model's ability to recognize knife switches captured at different distances. The spatial pyramid pooling module consists of four parallel branches, including one direct-connection branch and three max pooling layers of different sizes. The feature maps output by the branches are stitched together in the channel dimension and compressed through a 1×1 convolutional layer to form multi-scale fused features.

[0021] Preferably, the path aggregation network structure is used to perform bidirectional feature fusion from top to bottom and from bottom to top on the feature map output by the spatial pyramid pooling module, thereby improving the small target detection capability. The path aggregation network structure adopts a bidirectional feature fusion path from top to bottom and from bottom to top, combining high-level semantic features with low-level detail features to improve the positioning accuracy of the knife switch. The top-down path fuses high-level semantic features with shallow high-resolution features step by step through upsampling, and the bottom-up path passes the fused features to the deeper layers again through downsampling.

[0022] Preferably, the YOLOv8 model employs a decoupled detection head to predict the state category, location coordinates, and confidence level of the grounding switch disconnector, including:

[0023] The classification branch uses the Sigmoid function to output the probability distribution of the disconnector status category, thereby determining the disconnector status, which includes closed, open, and abnormal status.

[0024] The regression branch uses the CIoU loss function to optimize the bounding box coordinate prediction results of the switch position;

[0025] The confidence branch uses a binary cross-entropy loss function to calculate the confidence level of the target's existence, thereby evaluating the detection reliability.

[0026] A method for monitoring the status of grounding switch disconnectors based on AI image recognition, applied to the AI ​​image recognition-based grounding switch disconnector status monitoring system, includes the following steps:

[0027] S1. The image data acquisition task is triggered by a preset time interval and a specific event. Multiple high-definition cameras are used to perform image acquisition operations on the grounding switch knife switch from multiple perspectives. Each acquired image data is accompanied by a timestamp and device identification information.

[0028] S2. Using the image preprocessing module, the original image data collected in step S1 is sequentially processed by image enhancement, noise reduction and grayscale conversion, and the processed image data is transmitted to the AI ​​image recognition module.

[0029] S3. Based on the YOLOv8 model, the processed image data is input and the output is the identification result of the grounding switch disconnector. The identification result includes the state category, location coordinates and confidence level.

[0030] S4. Compare the recognition result with the threshold and pattern under normal conditions. If the recognition result matches the threshold and pattern under normal conditions, record the relevant data in the database. If the recognition result does not match, send a fault warning message to the management personnel through the communication module and trigger the local audible and visual alarm operation. At the same time, store the relevant data in the database.

[0031] S5. Based on image data, recognition results, and historical status records in the database, a machine learning algorithm is used to build a predictive model of equipment status evolution, predict potential failure trends, and formulate maintenance plans based on the prediction results.

[0032] Preferably, the training process of the YOLOv8 model includes:

[0033] By searching and obtaining diverse image data of grounding switches and disconnectors in the database, the bounding boxes of grounding switches and disconnectors in the image data are marked and the status categories are recorded. The diverse image data includes data under different lighting conditions, weather conditions and equipment status.

[0034] Deploy the YOLOv8 training framework on the selected computing platform, load the training set, validation set and test set data, use the AdamW adaptive optimizer to adjust the learning rate, and iteratively update the model parameters;

[0035] The hyperparameters of the YOLOv8 model are tuned using the Bayesian optimization algorithm, which intelligently searches for the optimal parameter combination within the hyperparameter space.

[0036] When the system detects that new grounding switch disconnector status image data and model performance indicators are lower than the preset threshold, the new data is labeled and mixed with historical data to form a new training set, and the YOLOv8 model is incrementally trained to update the model parameters.

[0037] Preferably, the relevant data includes image data, recognition results, and time information.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. This AI image recognition-based grounding switch status monitoring system can accurately capture the subtle differences in the grounding switch status under different states, effectively avoiding misjudgments caused by environmental factors, equipment aging, and human factors. Compared with traditional judgment methods, it greatly improves the accuracy of status judgment and significantly reduces the misjudgment rate.

[0040] 2. The YOLOV8 algorithm itself has a high image processing speed, which can complete the entire process from image acquisition to status judgment in a very short time, realize real-time dynamic monitoring of the status of grounding switch disconnectors, meet the strict requirements of power system for rapid response to equipment status, and provide timely and accurate status information for the safe and stable operation of power system;

[0041] 3. This invention abandons the excessive reliance on easily damaged and faulty auxiliary contacts and sensors that are greatly affected by environmental interference in traditional methods. Instead, it focuses on stable and reliable image acquisition and AI image recognition technology based on deep learning, thereby reducing the risk of monitoring failure due to hardware failure and improving the overall reliability and stability of the system.

[0042] 4. Compared with traditional multi-sensor detection solutions, this invention only requires the installation of a small number of high-definition cameras at key locations, which significantly reduces the procurement, installation and maintenance costs of hardware equipment. At the same time, the automated status judgment function greatly reduces the cost of manual inspection. From a long-term operation perspective, it has significant cost-effectiveness advantages and can save power companies a lot of operation and maintenance funds.

[0043] 5. By deeply mining and analyzing the historical status data of grounding switches and disconnectors stored in the database, and using advanced algorithms such as machine learning, an accurate equipment status evolution model can be established to predict potential fault trends in advance. Based on the prediction results, managers can formulate targeted preventive maintenance plans, take timely maintenance measures, effectively reduce the probability of sudden faults, reduce equipment maintenance costs, extend equipment service life, and improve the overall operation and maintenance efficiency of the power system. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall framework structure of the system of the present invention.

[0045] Figure 2 This is a schematic diagram of the image preprocessing module of the present invention.

[0046] Figure 3 This is a schematic diagram of the YOLOV8 model architecture of the present invention.

[0047] Figure 4 This is a schematic diagram illustrating the steps of the method of the present invention.

[0048] Figure 5 This is a schematic diagram illustrating the boundary box annotation of the present invention.

[0049] Figure 6 This is a visualization chart of key metrics during the training process of the YOLOv8 model of this invention.

[0050] Figure 7 This is a visualization chart of key metrics during the training process of the YOLOv8 model of this invention.

[0051] Figure 8 This is a schematic diagram of diverse image data used for training the YOLOv8 model of the present invention.

[0052] Figure 9 This is a complete metrics monitoring chart of the YOLOv8 model training process of the present invention.

[0053] Figure label annotations: Image acquisition module 10, Image preprocessing module 20, Image enhancement unit 21, Image denoising unit 22, Image grayscale conversion unit 23, AI image recognition module 30, Status judgment and early warning module 40, Communication module 50, Database 60. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0055] In this embodiment, as Figures 1-9 As shown, a grounding switch status monitoring system and method based on AI image recognition includes:

[0056] Image acquisition module 10 is equipped with multiple high-definition cameras to trigger image data acquisition operations based on preset time intervals and specific events. The preset time interval is set to 5 minutes, and specific events include operation command issuance and power failure alarm. The image data is all stamped with timestamps and device identification information.

[0057] Specifically, the image acquisition module 10 adopts a multi-view synchronous acquisition strategy. Multiple high-definition cameras are arranged at different angles of the grounding switch disconnector. The camera side can be equipped with a dustproof and moisture-proof design to avoid affecting industrial practicality. The viewing angles include front view, side view, top view and bottom view. The high-definition cameras all support acquisition modes triggered by timed triggering and specific events, automatically start the image acquisition program, and acquire images of the same grounding switch disconnector from multiple different angles to obtain comprehensive and accurate equipment status information for subsequent data processing and traceability.

[0058] Image preprocessing module 20 is used to perform image enhancement, denoising and grayscale processing operations on the acquired image data;

[0059] Specifically, such as Figure 2 As shown, the image preprocessing module 20 includes:

[0060] Image enhancement unit 21 employs a multi-scale Retinex algorithm combined with adaptive histogram equalization to perform overall and local illumination optimization and contrast enhancement on image data. The multi-scale Retinex algorithm, by simulating the human visual system's adaptability to illumination, decomposes the image into a reflection component (essential target features) and an illumination component (environmental interference), as shown in the formula: S(x,y) = R(x,y) × L(x,y), where S(x,y) is the original image, R(x,y) is the reflection component (features to be retained), and L(x,y) is the illumination component (ambient light to be corrected). The illumination component is estimated and removed using multi-scale convolutional kernels to achieve illumination homogenization. The adaptive histogram equalization algorithm divides the image into multiple sub-blocks and performs histogram equalization independently on each sub-block, avoiding noise amplification caused by global equalization. The formula is: Where f max f min The minimum and maximum pixel values ​​within the sub-block are used. During image enhancement, the multi-scale Retinex algorithm is first used to remove uneven illumination, and then the adaptive histogram equalization algorithm is applied to the reflection component to enhance local contrast, so that the metal edges, contacts and other details of the knife switch are clearly visible under different lighting conditions.

[0061] Image denoising unit 22 employs a fusion algorithm of bilateral filtering and non-local mean filtering to remove noise while preserving edge details of the image data. Bilateral filtering needs to consider both pixel spatial distance and grayscale value differences simultaneously, and the formula is as follows: The weight w is composed of a spatial Gaussian kernel d(i,j,k,l) ​​and a gray-level Gaussian kernel r(i,j,k,l), which can preserve edges while removing noise. Non-local mean filtering utilizes recurring texture patterns in the image, taking a weighted average of similar regions across the entire image for each pixel. The formula is: Among them, the weight w(i,j) is calculated based on the Euclidean distance between pixel blocks, which is robust to noise. In the process of image denoising, Gaussian noise is first removed quickly by bilateral filtering, and then non-local mean filtering is used to process the noise, while preserving the fine structure of the knife switch contact.

[0062] Image grayscale unit 23 executes the weighted conversion formula from color image to grayscale image, converting color image data into single-channel grayscale image data. The grayscale processing uses the formula: Gray = 0.299*R + 0.587*G + 0.114*B. This weight is set based on the sensitivity of the human eye to different colors, ensuring that the grayscale image retains the most critical visual information, converting the color image into a single-channel grayscale image, and preparing data for subsequent AI image recognition processing.

[0063] AI image recognition module 30, based on YOLOV8 model, is used to detect the position of grounding switch disconnector based on preprocessed image data and perform state classification operations. YOLOV8 model includes CSPDarknet53 backbone network, spatial pyramid pooling module and path aggregation network structure.

[0064] Specifically, such as Figure 3 As shown, the CSPDarknet53 backbone network is used to extract multi-scale features from the input grounding switch image. It generates feature maps of different resolutions through progressive downsampling to extract multi-level features of the switch. The CSPDarknet53 backbone network adopts a cross-stage partial connection structure, consisting of residual blocks in 5 stages. Each stage halves the feature map resolution and doubles the number of channels through downsampling, ultimately generating 4 feature maps of different scales (128×128, 64×64, 32×32, 16×16). The feature map output from the basic convolutional layer is divided into two parts. One part is subjected to deep feature extraction through dense stacking of residual blocks, with each dense residual block containing 16 residual units. The other part is directly fused with the output features of the residual blocks across stages to form a multi-scale feature pyramid. The feature map output from each downsampling stage contains information at different levels. The 128×128 feature map retains detailed information such as the switch edges and contacts, while the 16×16 feature map is used to extract the overall shape of the switch and global semantic features of whether it is closed or open.

[0065] Specifically, such as Figure 3As shown, the Spatial Pyramid Pooling (SPP) module is connected to the output of the CSPDarknet53 backbone network. It extracts global features from image data through multi-scale max pooling operations, enhancing the model's ability to recognize knife switches captured at different distances. The SPP module consists of four parallel branches, including one direct-connect branch and three max-pooling layers of different sizes. The feature maps output by the branches are concatenated along the channel dimension and compressed using a 1×1 convolutional layer to form multi-scale fused features. The input feature map from the CSPDarknet53 output is assumed to have a size of H×W×C=13×13×1024. This feature map already contains high-level semantic information such as the overall structure and local components of the knife switch, but lacks multi-scale global contextual information. The SPP module captures global features from different receptive fields. The direct-connect branch directly retains the original feature map without pooling, and the output size is consistent with the input size, H×W×C=13×13×1024. The 5×5 branch performs 5×5 max pooling on the input feature map with stride = 1 and padding = 2, keeping the output size unchanged. The 9×9 branch performs 9×9 max pooling on the feature map output by the 5×5 branch with stride = 1 and padding = 4, keeping the output size unchanged. The 13×13 branch performs 13×13 max pooling on the feature map output by the 9×9 branch with stride = 1 and padding = 6, keeping the output size unchanged. The outputs of the four branches are concatenated in the channel dimension (C) to obtain the fused feature map with an output size of H×W×4C = 13×13×4096. A 1×1 convolutional layer is used to compress the number of channels from 4C back to C, and 4096 back to 1024. The convolutional kernel size is 1×1, stride = 1, no padding, and the number of output channels is C, so that the output size is consistent with the input size, thus preserving the key features after multi-scale fusion. The multi-scale fused features output by the SPP module are input into the path aggregation network structure for cross-layer feature fusion, further improving the model's ability to detect knife gates at different scales.

[0066] Specifically, such as Figure 3As shown, the path aggregation network structure is used to perform bidirectional feature fusion from top to bottom and bottom to top on the feature map output by the spatial pyramid pooling module, improving the small target detection capability. The path aggregation network structure adopts a bidirectional feature fusion path from top to bottom and bottom to top, combining high-level semantic features with low-level detail features to improve the positioning accuracy of the knife switch. The top-down path fuses high-level semantic features with shallow high-resolution features step by step through upsampling, while the bottom-up path passes the fused features to the deeper layers again through downsampling. The SPP module has extracted global context information through multi-scale pooling, but it needs to further fuse spatial details at different levels. The path aggregation network structure (PAN structure) takes the multi-scale fused feature map output by the SPP module as input, assuming a size of H×W×C=13×13×1024. The top-down path obtains a high semantic feature map of 26×26×512 from the deep layers of the CSPDarknet53 backbone network, and performs a 2x upsampling on the high-level feature map through bilinear interpolation or transposed convolution to make its size similar to that of the adjacent shallow features. Figure 1 The 13×13×1024 feature map is upsampled to 26×26×512. The upsampled high-level features are then concatenated (C) with the shallow high-resolution feature map (26×26×256) along the channel dimension, or added element-wise. A 1×1 convolution is used to adjust the number of channels and reduce computational cost. Then, a 3×3 convolution is used to refine the fused features, eliminating the aliasing effect of the upsampled map. The bottom-up path starts from the fused result of the top-down path (26×26×256) and is processed using convolutions with a stride of 2 or max pooling. Downsampling is performed to match the deep feature map size (13×13). The downsampled features are then concatenated or added to the original deep features (H×W×C=13×13×1024). Similarly, the feature representation is optimized through 1×1 and 3×3 convolutions. The output multi-level feature maps (52×52×256, 26×26×512, 13×13×1024) are then input into the decoupled detection head of the YOLOV8 model. The fused features are robust to remote / near-range shooting, occlusion, or partially visible knife switches.

[0067] Specifically, such as Figures 6-9 As shown, the training process of the YOLOv8 model includes:

[0068] By searching and obtaining diverse image data of grounding switches and disconnectors from database 60, the bounding boxes of grounding switches and disconnectors in the image data are marked and the status categories are recorded. The diverse image data includes data under different lighting conditions, weather conditions and equipment status.

[0069] Deploy the YOLOv8 training framework on the selected computing platform, load the training set, validation set and test set data, use the AdamW adaptive optimizer to adjust the learning rate, and iteratively update the model parameters;

[0070] The hyperparameters of the YOLOv8 model are tuned using the Bayesian optimization algorithm, which intelligently searches for the optimal parameter combination within the hyperparameter space.

[0071] When the system detects that new grounding switch disconnector status image data and model performance indicators are lower than the preset threshold, the new data is labeled and mixed with historical data to form a new training set, and the YOLOv8 model is incrementally trained to update the model parameters.

[0072] Specifically, through Figure 6 This feature displays the curves of key metrics during training as the number of iterations increases, such as loss function value, precision, recall, and mAP. This helps to evaluate model convergence in real time. By observing whether the loss is steadily decreasing or the precision is saturating, it helps determine if overfitting, underfitting, or improper learning rate settings have occurred. This provides a clear basis for hyperparameter optimization, adjusting the learning rate and batch size, and other related operations. Figure 8 This section presents samples from the training dataset under different lighting conditions (strong light / nighttime), weather (rain / fog), and switch states (closed / open / abnormal), verifying the comprehensiveness of the data coverage. It demonstrates that the model training data includes complex interferences from real-world scenarios (such as occlusion and corrosion), ensuring the model can adapt to actual environments. Samples with labeled bounding boxes and state categories are also displayed, reflecting the labeling standards. Figure 9 It demonstrates complete metric monitoring of YOLOv8 model training, and can compare the metrics of the training set and the validation set to evaluate the model's generalization ability. If the validation set metrics are significantly lower than those of the training set, a data augmentation strategy is introduced. After hyperparameter tuning, the optimal model version is selected by comparing metrics, and the optimal model version is used for image recognition operations.

[0073] The status judgment and early warning module 40 is used to receive the switch status recognition result output by the AI ​​image recognition module 30, compare it with the threshold and mode preset in the normal state of the system, and trigger an early warning when there is an abnormality.

[0074] Specifically, the status judgment and early warning module 40 receives structured data output by the AI ​​image recognition module 30, including the position coordinates (boundary box coordinates) of the disconnector, the status category (closed, open, abnormal) and the corresponding confidence level (0~1), timestamp and device identification information. The recognition results are verified by preset rules and thresholds. The threshold is selected according to the actual environmental requirements. Assuming the threshold is 0.5, if the confidence level is lower than the preset threshold, it is judged as "uncertain state" and triggers the manual review mechanism. When the confidence level reaches the standard, the mode judgment is entered. If the position of the disconnector meets the preset geometric conditions (such as the center point of the boundary box is located in the closed area), it meets the closed state. If the angle or position of the disconnector matches the open template, it meets the open state. If the disconnector is in a half-closed / half-open, obstructed, broken, etc., it meets the abnormal state and triggers the abnormal mark. In the normal state, the data is recorded to the database 60 and marked as "normal". In the abnormal state, the audible and visual alarm device is activated, and alarm information is pushed through SMS or email. The relevant data of the abnormal event is stored, including image data, recognition results, time, device ID, etc.

[0075] The communication module 50 is used to realize data interaction and transmission between modules. Internal communication uses lightweight protocols such as MQTT or ZeroMQ, which supports high-frequency, low-latency data transmission and is suitable for real-time interaction between modules. External communication uses HTTP or WebSocket, which is convenient for interfacing with external systems such as monitoring centers and mobile terminals.

[0076] Database 60 stores image data, recognition results, and historical status records, supporting data traceability and analysis. It collects disconnector images under different lighting conditions (strong light, backlight, nighttime), weather conditions (heavy rain, heavy fog), and equipment conditions (closed, open, abnormal corrosion) from power industry-related databases, research institution data sharing platforms, and internal image storage systems of major substations. These images are organized into samples according to data type and time sequence. Using an active learning-based annotation algorithm, a small number of representative image samples are first manually labeled. Then, the annotation algorithm selects the most representative images from a large number of unlabeled images based on the characteristics of the labeled samples. For images with certainty and representativeness, manual annotation is recommended. Use the LabelImg tool to annotate the boundary boxes of the disconnectors and mark the status category (e.g., closed = 0, open = 1, abnormal = 2). Further annotate the specific type of abnormal status (e.g., contact oxidation, mechanical deformation). Repeat this process to gradually expand the annotation sample set, thereby improving annotation efficiency and accuracy. The database 60 conforms to the IEC61850 standard and supports data interaction with power monitoring systems such as SCADA and EMS. Regular backup and off-site disaster recovery storage technology are adopted to prevent data loss and provide reliable data protection for the long-term stable operation of the system and data analysis.

[0077] Among them, such as Figure 3As shown, the YOLOv8 model uses a decoupled detection head to predict the state category, location coordinates, and confidence level of the grounding switch disconnector, including:

[0078] The classification branch uses the Sigmoid function to output the probability distribution of the disconnector status category, and then determines the disconnector status, which includes closed, open and abnormal status.

[0079] The regression branch uses the CIoU loss function to optimize the bounding box coordinate prediction results of the switch position;

[0080] The confidence branch uses a binary cross-entropy loss function to calculate the confidence of the target's existence, thereby evaluating the detection reliability.

[0081] Specifically, the input consists of multi-scale feature fusion maps (52×52×256, 26×26×512, 13×13×1024) from the path aggregation network structure (PAN structure). Each scale's feature map is input to one of the three branches of the decoupled detection head. The overall loss function is calculated using the formula: L total =λ1L CIoU +λ2L class +λ3L conf The weights λ1, λ2, and λ3 are adjusted according to the importance of the task.

[0082] The classification branch uses a 1×1 convolution to adjust the number of channels in the input feature map to three categories: closed, open, and abnormal states. The output size is H×W×C (C=3), and the output is the category label and probability value for each detection box. A Sigmoid function is applied to each channel to independently calculate the probability of each category, with a preset threshold of 0.5. If the probability of a category exceeds the threshold, it is determined to be in that state. The Sigmoid function calculation formula is: Allows the possibility of multiple states coexisting simultaneously for the disconnector;

[0083] Each detection point in the regression branch predicts four parameters, corresponding to the center offset and width / height scaling of the bounding box. These predicted values ​​are then converted to actual coordinates. The CIoU loss function is formulated as follows: Where ρ is the Euclidean distance between the center point of the predicted box and the center point of the ground truth box, and v is the aspect ratio consistency measure;

[0084] The confidence branch uses binary classification prediction, and the sigmoid function outputs the confidence score, which is within the interval [0,1]. The formula for the binary cross-entropy loss function is: L conf = -[ylog(p)+(1-p)log(1-p)], where y=1 indicates the existence of a target, and y=0 indicates the background and the absence of a target;

[0085] like Figure 4As shown, the specific steps for using this AI image recognition-based grounding switch status monitoring system are as follows:

[0086] S1. The image data acquisition task is triggered by a preset time interval and a specific event. Multiple high-definition cameras are used to perform image acquisition operations on the grounding switch knife switch from multiple perspectives. Each acquired image data is accompanied by a timestamp and device identification information.

[0087] S2. Using the image preprocessing module 20, the original image data acquired in step S1 is sequentially processed by image enhancement, noise reduction and grayscale conversion, and the processed image data is transmitted to the AI ​​image recognition module 30.

[0088] S3. Based on the YOLOv8 model, the processed image data is input and the output is the identification result of the grounding switch disconnector. The identification result includes the state category, location coordinates and confidence level.

[0089] S4. Compare the recognition result with the threshold and pattern under normal conditions. If the recognition result matches the threshold and pattern under normal conditions, record the relevant data in the database 60. If the recognition result does not match, send a fault warning message to the management personnel through the communication module 50 and trigger the local audible and visual alarm operation. At the same time, store the relevant data in the database 60. The relevant data includes image data, recognition result and time information.

[0090] S5. Based on the image data, recognition results and historical status records in database 60, a machine learning algorithm is used to build a prediction model for equipment status evolution, predict potential failure trends, and formulate a maintenance plan based on the prediction results.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A grounding switch / disconnector status monitoring system based on AI image recognition, characterized in that, include: The image acquisition module (10) is equipped with multiple high-definition cameras and is used to trigger image data acquisition operations based on a preset time interval and specific events. The preset time interval is set to 5 minutes, and the specific events include operation command issuance and power fault alarm. The image data all have timestamps and device identification information. The image preprocessing module (20) is used to perform image enhancement, denoising and grayscale processing on the acquired image data; AI image recognition module (30), based on YOLOV8 model, is used to detect the position of grounding switch knife switch according to preprocessed image data and perform state classification operation. The YOLOV8 model includes CSPDarknet53 backbone network, spatial pyramid pooling module and path aggregation network structure. The status judgment and early warning module (40) is used to receive the switch status recognition result output by the AI ​​image recognition module (30), compare it with the threshold and mode preset in the normal state of the system, and trigger an early warning when there is an abnormality. The communication module (50) is used to realize data interaction and transmission operations between the modules; The database (60) is used to store image data, recognition results and historical status records, and supports data traceability and analysis.

2. The grounding switch / disconnector status monitoring system based on AI image recognition according to claim 1, characterized in that, The image acquisition module (10) adopts a multi-view synchronous acquisition strategy. Multiple high-definition cameras are respectively arranged at different views of the grounding switch knife switch. The views include front view, side view, top view and bottom view. All high-definition cameras support acquisition modes triggered by timed events and triggered by specific events.

3. The grounding switch / disconnector status monitoring system based on AI image recognition according to claim 1, characterized in that, The image preprocessing module (20) includes: The image enhancement unit (21) adopts a multi-scale Retinex and adaptive histogram equalization fusion algorithm to perform overall and local illumination optimization and contrast enhancement operations on image data. The image denoising unit (22) uses a fusion algorithm of bilateral filtering and non-local mean filtering to remove noise and retain edge detail information of image data; The image grayscale unit (23) executes the weighted conversion formula from color image to grayscale image to convert color image data into single-channel grayscale image data.

4. The grounding switch / disconnector status monitoring system based on AI image recognition according to claim 1, characterized in that, The CSPDarknet53 backbone network is used to extract multi-scale features from the input grounding switch image. It generates feature maps of different resolutions through progressive downsampling to extract multi-level features of the switch. The CSPDarknet53 backbone network adopts a cross-stage partial connection structure. It divides the feature map output from the basic convolutional layer into two parts. One part is processed by dense stacking of residual blocks for deep feature extraction, and the other part is directly fused with the residual block output features across stages to form a multi-scale feature pyramid.

5. The grounding switch / disconnector status monitoring system based on AI image recognition according to claim 4, characterized in that, The spatial pyramid pooling module is connected to the output of the CSPDarknet53 backbone network. It extracts global features from image data through multi-scale max pooling operations, enhancing the model's ability to recognize knife switches captured at different distances. The spatial pyramid pooling module consists of four parallel branches, including one direct-connection branch and three max pooling layers of different sizes. The feature maps output by the branches are spliced ​​in the channel dimension and compressed through a 1×1 convolutional layer to form multi-scale fused features.

6. The grounding switch / disconnector status monitoring system based on AI image recognition according to claim 5, characterized in that, The path aggregation network structure is used to perform bidirectional feature fusion from top to bottom and bottom to top on the feature map output by the spatial pyramid pooling module, thereby improving the small target detection capability. The path aggregation network structure adopts a bidirectional feature fusion path from top to bottom and bottom to top, combining high-level semantic features with low-level detailed features to improve the positioning accuracy of the knife switch. The top-down path fuses high-level semantic features with shallow high-resolution features step by step through upsampling, and the bottom-up path passes the fused features to the deeper layers again through downsampling.

7. The grounding switch / disconnector status monitoring system based on AI image recognition according to claim 1, characterized in that, The YOLOV8 model employs a decoupled detection head to predict the state category, location coordinates, and confidence level of the grounding switch disconnector, including: The classification branch uses the Sigmoid function to output the probability distribution of the disconnector status category, thereby determining the disconnector status, which includes closed, open, and abnormal status. The regression branch uses the CIoU loss function to optimize the bounding box coordinate prediction results of the switch position; The confidence branch uses a binary cross-entropy loss function to calculate the confidence level of the target's existence, thereby evaluating the detection reliability.

8. A method for monitoring the status of grounding switch disconnectors based on AI image recognition, applied to the grounding switch disconnector status monitoring system based on AI image recognition as described in any one of claims 1-7, characterized in that, The specific steps include: S1. The image data acquisition task is triggered by a preset time interval and a specific event. Multiple high-definition cameras are used to perform image acquisition operations on the grounding switch knife switch from multiple perspectives. Each acquired image data is accompanied by a timestamp and device identification information. S2. Using the image preprocessing module (20), the original image data collected in step S1 is sequentially processed by image enhancement, noise reduction and grayscale conversion, and the processed image data is transmitted to the AI ​​image recognition module (30). S3. Based on the YOLOv8 model, the processed image data is input and the output is the identification result of the grounding switch disconnector. The identification result includes the state category, location coordinates and confidence level. S4. Compare the identification result with the threshold and pattern under normal conditions. If the identification result matches the threshold and pattern under normal conditions, record the relevant data in the database (60). If the identification result does not match, send a fault warning message to the management personnel through the communication module (50) and trigger the local sound and light alarm operation. At the same time, store the relevant data in the database (60). S5. Based on the image data, recognition results and historical status records in the database (60), a machine learning algorithm is used to construct a prediction model of equipment status evolution, predict potential fault trends, and formulate a maintenance plan based on the prediction results.

9. The method for monitoring the status of grounding switch disconnectors based on AI image recognition according to claim 1, characterized in that, The training process of the YOLOv8 model includes: By searching, diverse image data of grounding switch disconnectors in database (60) are obtained, and the bounding boxes of grounding switch disconnectors in the image data are marked and the status categories are recorded. The diverse image data includes data under different lighting conditions, weather conditions and equipment status. Deploy the YOLOv8 training framework on the selected computing platform, load the training set, validation set and test set data, use the AdamW adaptive optimizer to adjust the learning rate, and iteratively update the model parameters; The hyperparameters of the YOLOv8 model are tuned using the Bayesian optimization algorithm, which intelligently searches for the optimal parameter combination within the hyperparameter space. When the system detects that new grounding switch disconnector status image data and model performance indicators are lower than the preset threshold, the new data is labeled and mixed with historical data to form a new training set, and the YOLOv8 model is incrementally trained to update the model parameters.

10. The method for monitoring the status of grounding switch disconnectors based on AI image recognition according to claim 1, characterized in that, The relevant data includes image data, recognition results, and time information.

Images