Modeling method for medium and low voltage power distribution equipment

Abstract

The invention provides a medium and low voltage power distribution equipment modeling method, and belongs to the technical field of power supply. The mobile acquisition device is used for acquiring multi-source information of the medium and low voltage power distribution equipment; careful cleaning and screening are carried out, abnormal and irrelevant data are removed, high-quality pictures are screened out, and accurate data positioning is carried out; generating a medium and low voltage power distribution equipment sample set through an image preprocessing technology; designing a customized medium and low voltage power distribution equipment detection model according to the characteristics of the medium and low voltage equipment and the environment of the medium and low voltage equipment, and training by using the designed model; hyper-parameters are adjusted, training is repeated for multiple times, results are reserved, and an optimal model is selected as a final model; and performing post-processing on a result output by the model, establishing a middle and low voltage power distribution equipment edge layout, and providing a middle and low voltage power distribution equipment visual display, management and auxiliary decision-making platform. According to the invention, the accuracy and efficiency of data acquisition are improved, high-precision dynamic positioning and real-time drawing are realized, and the accuracy and timeliness of data are improved.

Description

Technical Field

[0001] The present invention provides a modeling method for medium and low voltage power distribution equipment, belonging to the technical field of power supply. Background Art

[0002] Currently, in the field of modeling technology for medium and low voltage power distribution equipment, traditional practices often rely on manual inspections, manually recording equipment information, and then using computers for data entry and drawing layout maps. This method has many shortcomings in practical applications. On the one hand, manual information entry is not only time-consuming, but also prone to errors. Especially in complex environments, the accuracy of equipment information cannot be guaranteed, and missed detections and false detections are prone to occur. On the other hand, the process of manually drawing layout maps is cumbersome and prone to data lags, which in turn affects the accuracy and timeliness of the data. In addition, existing data management methods are inefficient, and data verification and updates often require a lot of human resources. Especially when faced with a huge amount of existing data, data rectification work is extremely difficult.

[0003] Regarding positioning technology, the currently widely used mobile phone positioning technology lacks accuracy in complex environments, especially in areas with dense buildings or signal interference. GPS signals can be obstructed, resulting in poor positioning and prone to position drift. Furthermore, positioning and mapping tasks are often separated, requiring personnel to complete positioning first and then return to headquarters for data processing and mapping. This work model not only increases workload but can also lead to inconsistent data, affecting the precise positioning of low-voltage equipment and the quality of layout mapping. Summary of the Invention

[0004] In response to the above problems, the present invention aims to overcome the key difficulties in the current modeling of medium and low voltage distribution equipment through technological innovation. First, by developing a lightweight medium and low voltage distribution equipment identification system assisted by deep learning, combined with point cloud information, the recognition speed and accuracy are improved. Using deep learning-based target detection models such as the real-time self-attention model (RTDETR), feature extraction and recognition are performed on equipment images. At the same time, methods such as image denoising, dark light enhancement, image deblurring, image registration, three-dimensional point cloud modeling, and super-resolution image reconstruction are added to improve recognition results. Secondly, three-dimensional point cloud data is used to reconstruct the distribution equipment and use it as an auxiliary data source to enhance the detection results of medium and low voltage distribution equipment, improve detection accuracy and recall rate, and make equipment information identification more accurate. Thirdly, 5G and RTK technology are used to achieve dynamic positioning of operators and real-time mapping synchronization, thereby innovating geographic information management. Finally, an integrated management system is constructed to automatically verify and update data, ensure data integrity and timeliness, achieve transparent management of medium and low voltage distribution data, reduce human resource consumption, and improve work efficiency.

[0005] Through the above-mentioned technological innovations, the present invention solves the inefficient and error-prone manual information entry problem of traditional methods, thereby improving the accuracy and efficiency of data collection. It also addresses the limitations of positioning technology, achieving high-precision dynamic positioning and real-time mapping, improving the accuracy and timeliness of data. It also addresses the difficulty of data rectification by reducing the workload of manual review through automated data management, ensuring data consistency and accuracy. The application of the present invention can greatly improve the safety and efficiency of power grid operations and maintenance, and promote the transformation of the power industry towards intelligence.

[0006] Technical specific steps:

[0007] The method for modeling medium and low voltage power distribution equipment provided by the present invention comprises the following steps:

[0008] A. Acquire multi-source data through acquisition devices: First, a mobile acquisition device that integrates an infrared camera, a visible light camera, a 3D point cloud camera, and 5G and RTK positioning equipment is used to collect multi-source information on medium and low voltage distribution equipment.

[0009] B. Data cleaning and screening: Detailed cleaning and screening of collected data to eliminate abnormal and irrelevant data, and select high-quality images and accurate data positioning;

[0010] C. Data preprocessing: Generate medium and low voltage distribution equipment sample sets through image preprocessing steps such as image white balance, image denoising, image deblurring, multi-source image registration and multi-source image fusion;

[0011] D. Design and train models: Design customized detection models for medium and low voltage distribution equipment based on the characteristics of the equipment and its environment, and use the designed models for training.

[0012] E. Parameter tuning: Adjust hyperparameters, repeat training multiple times, retain the results, and select the optimal model as the final model;

[0013] F. Data Post-Processing: Post-process the model output results, relocate the actual position of the medium and low voltage power distribution equipment based on the positioning data of the mobile acquisition device and the relative position of the medium and low voltage power distribution equipment obtained based on the point cloud data, and associate the actual position information with the target medium and low voltage power distribution equipment to generate an electronic handover form;

[0014] G. Result visualization and information management: Establish a layout of medium and low voltage power distribution equipment, and provide a platform for visual display, management and decision support of medium and low voltage power distribution equipment.

[0015] Step A specifically involves creating a mobile data acquisition device that integrates a point cloud camera, an infrared camera, a visible light camera, a 5G+RTK hybrid positioning system, a built-in power supply, a high-density storage device, a SOC control module, and an IMU nine-axis sensor. This device is designed to collect multiple data from medium and low voltage power distribution equipment. It uses a point cloud camera to obtain three-dimensional spatial information about the equipment, generating high-density point cloud data to construct a three-dimensional model of the equipment. The visible light camera has a resolution of 2560×1440 and a frame rate of up to 30 frames per second, capable of capturing high-definition images of the equipment under natural lighting conditions, supporting device appearance feature recognition and environmental background analysis. The infrared camera, configured with a resolution of 640×512 and a frame rate of 30 frames per second, is used to detect the thermal distribution of the equipment under different temperature conditions and detect potential overheating hazards in power equipment. Furthermore, the device combines centimeter-level 5G high-precision positioning services with real-time dynamic measurement technology to record the location information of on-site inspectors in real time. Given the device's offline nature, a built-in power supply is designed to ensure continuous power supply. Given the large amount of data transmission and storage required, high-density, high-speed storage devices are used to store multi-source images, raw video, and positioning data. The SOC control module manages device startup and data read / write, ensuring data synchronization across all devices on the device. Meanwhile, the IMU nine-axis sensor monitors the acquisition device's posture and motion status in real time, including acceleration, rotation, and changes in direction. This data not only accurately captures the acquisition device's trajectory, but also ensures that the point cloud camera, visible light camera, and infrared camera remain stable and reliable during the acquisition process, thereby ensuring the quality of the acquired data. The system also works in conjunction with the 5G+RTK hybrid positioning system to achieve high-precision mapping of medium and low voltage power distribution equipment along the layout.

[0016] The specific steps of step B are as follows: in the data screening and cleaning stage, the collected raw data is first subjected to an automated data cleaning task, and low-quality images and positioning drift points are screened and removed through algorithms. Specifically, an image processing algorithm is used to automatically detect and remove blurred and distorted images caused by insufficient lighting, occlusion, or device movement. At the same time, the positioning data quality control mechanism is used to identify and correct positioning drift points caused by signal interference or environmental factors. In addition, the data is also checked for consistency to ensure that the data collected by different sensors match each other, thereby ensuring the high accuracy and reliability of the final data set, and providing a high-quality data foundation for subsequent intelligent analysis and three-dimensional reconstruction. Finally, the screened samples are labeled using manual labeling methods and made into a multi-source data set for training.

[0017] Among them, the algorithm used to automatically screen image quality is an adaptive image quality assessment and screening algorithm, and its implementation steps are as follows:

[0018] 1. Read image: Read each original image to be processed from the dataset.

[0019] 2. Preprocessing: Convert the image to grayscale, as most image analysis operations are more efficient on grayscale images. Apply a median filter to the grayscale image to reduce the effects of noise while preserving edge characteristics.

[0020] Evaluate image clarity: Calculate the gradient magnitude of the image. Use the Canny edge detection method to obtain the gradient information of the image.

[0021] Statistically analyze the distribution of gradient strengths across the entire image or within key regions. A clear image should have a high average gradient strength and a large standard deviation. A threshold, T1, is set to determine whether the image is sufficiently clear. If the overall gradient strength of the image is lower than T1, the image is marked as blurry. In this method, T1 is set to 0.52. This value is based on experimental observations of this dataset. If the overall average gradient strength of an image is lower than T1, the image is usually blurry or has insufficient lighting.

[0022] Checking image integrity: Using template matching techniques, the current image is compared with known good samples. These good samples are selected from a selection of manually selected high-quality standard images. A threshold T2 is set based on the matching results to determine the integrity of the image content. If the matching score falls below T2, the image is considered to have occlusion or other forms of damage. In this method, T2 is set to 0.74. This value is based on experimental observations of this dataset. If the matching score or feature point matching ratio falls below T2, it generally indicates that the image has occlusion or other forms of damage, resulting in incomplete image content.

[0023] Comprehensive decision: If an image is judged to be both blurry and lacking sufficient content completeness, it is marked as a low-quality image.

[0024] Otherwise, the image is retained for subsequent processing.

[0025] Output: Based on the results of the above steps, a new dataset is generated containing all qualified images and those marked as low quality are removed.

[0026] Parameter Tuning and Optimization: Training a classification model to automatically generate these thresholds improves automation and accuracy.

[0027] Step C specifically includes: performing image white balance, image denoising, image deblurring, image registration, and image fusion processing during the data preprocessing phase. First, image white balance is performed to eliminate color deviations caused by changes in lighting conditions. The gray world hypothesis algorithm is used to calculate the average brightness of the image and adjust the RGB channels to make them tend to the standard white point. The formula is: Where Ref is the reference brightness, Img is the average brightness of the image, R is the pixel value of the red channel in the original image, and R′ is the pixel value of the red channel after white balance adjustment. Next, image denoising is performed, and a bilateral filtering algorithm is used to remove random noise in the image. This algorithm retains edge details while removing noise. The formula for bilateral filtering is:

[0028]

[0029] Where I(x) represents the value of the center pixel, I′(x) is the filtered center pixel value, y represents all pixels traversed in the neighborhood centered on x, and f d is the distance weight function, f r is a range weight function; then, image deblurring is performed, using an iterative deblurring method based on the Lucy-Richardson algorithm to gradually restore image clarity by estimating the blur kernel; in terms of image registration, the feature point matching algorithm SIFT is used to find corresponding points between visible light and infrared images, and the transformation matrix is ​​estimated by the least squares method to achieve image alignment between infrared and visible light. The point cloud data comes from a binocular camera, so the point cloud image has been registered with the visible light image; finally, SeAFusion is used in combination with an advanced vision image fusion framework to merge image information acquired by multiple sensors into a single image, that is, multi-scale feature extraction and self-attention mechanism are used to select and fuse data from different sensors. The specific fusion steps are as follows:

[0030] (1) Multi-scale feature extraction: Use deep convolutional neural network ResNet to extract multi-scale features from images of different sensors. Feature extraction can be expressed as:

[0031] f(x)=Conv(x)

[0032] Where f is the feature extraction function, x is the input image, and Conv is the convolutional layer.

[0033] (2) Self-attention mechanism: The extracted features are weighted through the self-attention mechanism to emphasize those parts that contribute most to the fusion results. The self-attention mechanism can be expressed as:

[0034]

[0035] Where Q, K, and V are query, key, and value matrices respectively, d k is the square root of the dimension of the key vector.

[0036] (3) Feature fusion: The weighted feature maps are fused to generate a fused image. The fusion process can be performed by simple averaging or more complex feature concatenation, skip connection, etc. Feature fusion can be expressed as:

[0037] Fused=αf1+βf2+...+γf n

[0038] where αf1, βf2, ..., γf n are the feature maps of different sensors, and α, β, ..., γ are weighting coefficients.

[0039] (4) Residual learning: By introducing residual blocks to learn the residual between input features and fusion features, important information is preserved. Residual learning can be expressed as:

[0040] Residual=Fused+Input

[0041] Through the above steps, a single image is finally generated that integrates multi-sensor information. This image contains multi-source information from different sensors and can be used for subsequent tasks such as device identification and detection.

[0042] The step D: designing and training the model: designing a customized model according to the characteristics of the medium and low voltage equipment and its surrounding environment, and using the designed model for training. The specific operations are divided into steps D1 designing the model and D2 training the model.

[0043] The step D1 specifically includes: performing detection of medium and low voltage power distribution equipment by improving RTDETR according to the characteristics of the medium and low voltage power distribution equipment and its surrounding environment. This method is based on RT-DETR (Real-Time Detection TRansformer) and performs targeted optimization to adapt to the characteristics of medium and low voltage power distribution equipment and its surrounding environment.

[0044] First, this method introduces a more powerful backbone network in YOLOv8 to replace the original backbone network ResNet50 of RTDETR to enhance the ability to capture device-specific features such as shape, texture, and color.

[0045] Secondly, this method designs an efficient multi-scale feature fusion strategy that combines feature maps of different scales to ensure that the model can handle objects of various scales while maintaining high accuracy. To further improve the robustness of the model, this method also incorporates data augmentation techniques such as random cropping, rotation, scaling, and translation to increase the diversity of the training data.

[0046] Finally, when designing the loss function for the medium and low voltage distribution equipment detection model, this method uses classification loss (Classification Loss), distributed focal loss (DFL), and bounding box regression loss (Bounding Box Regression Loss). Among them, classification loss is used to measure the performance of the model on the classification task and determine whether the target category is correct. The regression box loss function of this method is cross-entropy loss (Cross-Entropy Loss), which is used to evaluate the accuracy of the model in locating the bounding box of the target object; distributed focal loss is an improved loss function designed for bounding box regression. It transforms the bounding box regression problem into a distribution prediction problem, that is, predicting the discrete distribution of the four parameters of the bounding box (center point coordinates, width and height).

[0047] The total training loss function is the weighted sum of the above three loss functions. When processing medium and low voltage distribution equipment detection, the model can not only accurately identify the equipment category, but also accurately locate the equipment and maintain high precision.

[0048] Step D2 specifically involves: During model training, a multi-source dataset of annotated medium and low voltage power distribution equipment detection models (including infrared images, visible light images, and 3D point cloud data) is used. This method uses the PyTorch deep learning library to build an improved RT-DETR model and initialize model parameters. During training, the AdamW optimizer and a dynamic learning rate adjustment strategy are used, with an initial learning rate set to 1e-4. The learning rate is automatically adjusted based on performance indicators on the validation set.

[0049] Specifically, the training process of the medium and low voltage distribution equipment detection model is as follows:

[0050] (1) Read m images X = {x1, x2, ..., xm} and their corresponding category labels and ground truth coordinate labels Y = {y1, y2, ..., ym} from the medium and low voltage distribution equipment dataset. m represents the batch size, which defaults to 8. Preprocess the read images.

[0051] (2) The image data X is passed into the backbone network Backbone to extract features f1, f2, and f3 of different scales in the image.

[0052] (3) These features are passed to the hybrid high-efficiency encoder, which outputs multi-scale fusion features η1, η2, and η3. The hybrid encoder effectively reduces the consumption of computing resources by decoupling internal scale interactions and cross-scale fusion, while improving detection accuracy. In the modeling project of medium and low voltage distribution equipment, the design of the hybrid encoder uses an attention mechanism to perform internal scale interactions of high-level features. When processing multi-scale data from different sensors, it can improve the accuracy of device recognition and three-dimensional reconstruction through efficient feature extraction and fusion, while reducing computing latency and supporting real-time data processing needs.

[0053] (4) The multi-scale fusion features η1, η2, and η3 are passed to the decoder. The decoder, through its flexible inference speed adjustment feature, allows the use of different decoding layers to change the inference speed without retraining, which greatly promotes the application practice in real-time scenarios.

[0054] (5) Based on the coordinates and categories of the predicted box and the true box, the classification loss (ClassificationLoss), distributed focal loss (DFL), and bounding box regression loss (Bounding BoxRegression Loss) are calculated. Among them, the classification loss is used to measure the gap between the predicted category and the true category, the bounding box regression loss is used to optimize the positioning performance of the model to ensure that the model can accurately mark the location of the object, and the distributed focal loss is used to predict the discrete distribution of the bounding box parameters to improve the positioning accuracy of the model.

[0055] (6) Multiply the classification loss and regression box loss by the dynamic weight w dfl 、w cls 、w iou , we get the total training loss L, and calculate the gradient of the model parameters based on L through back propagation. This gradient represents the direction and speed of change of the loss function under the current parameter value; then we use the stochastic gradient descent optimization function to update the model parameters. The dynamic weight is used to balance the impact of each loss on the model back propagation process. The initial w iou 2, w cls is 5, w dfl is 1, and the loss weight is adjusted according to the progress of training, and finally w iou 6.85, w cls is 1.0, w dfl is 1.25, which can be expressed as:

[0056] L=w cls L cls +w dfl L dfl +w iou Liou

[0057] L cls =-w n [l n logσ(s n )+(1-l n )log(1-σ(s n ))]

[0058]

[0059]

[0060] In L iou In, L IoU represents the intersection-over-union loss, A and B represent the candidate box and the real box, x and y represent the center coordinates of the candidate box, gt and y gt Represents the center coordinates of the true box, W and H are the smallest rectangles that can frame A and B, α and σ are two preset hyperparameters, with the default α = 1.6 and σ = 4. Represents the exponential sliding average of momentum. It is a dynamic value that combines historical gradient information with the current gradient to achieve a smoother and faster optimization process. Represents the monotonic focusing coefficient, which is used to measure the focusing effect.

[0061] In L dfl In, p t and p t+1 It is the probability of the two positions closest to the label y in the predicted distribution, and β is an adjustment factor, usually 2. This loss function allows the network to focus on values ​​near the target faster, increasing their probability, thereby improving the quality and accuracy of the bounding box.

[0062] In L cls In it, σ is the sigmoid activation function, Q is the weight of the n-th sample, l is the true label, and s is the output of the model.

[0063] (7) Repeat steps (1) to (6) until the preset number of iterations is reached or the early stopping termination condition is met.

[0064] (8) Filter the detection boxes according to the confidence level and delete the detection boxes that are less than the confidence threshold.

[0065] (9) Based on the detection boxes and the true annotations in the dataset, calculate a series of accuracy indicators, including precision, recall, and average precision, evaluate the model performance, and record the results and model hyperparameters.

[0066] (10) Adjust the hyperparameters, repeat steps (1) to (9), and compare the results multiple times until the optimal result is obtained.

[0067] The specific steps of step E are as follows: first, determine the hyperparameters that need to be adjusted during model training, including but not limited to learning rate, batch size, momentum factor, weight decay, etc. Use the Grid Search algorithm to enumerate all candidate value combinations of hyperparameters, and train the model one by one to evaluate the performance. In order to speed up the search process, the Random Search algorithm is used to randomly select hyperparameter combinations for testing. During the training process, a learning rate decay strategy is adopted, using the exponential decay method:

[0068]

[0069] Among them, η0 is the initial learning rate, ρ is the decay ratio, and t decay is the decay period. During training, this method regularly evaluates the model's performance on the validation set using metrics such as mAP (mean Average Precision) and Intersection over Union (IoU), and records the results multiple times. To prevent overfitting and improve training efficiency, early stopping is introduced during training. If performance on the validation set does not improve after several consecutive epochs, training is terminated early. Based on the evaluation results, the model is fine-tuned, which may include adjusting hyperparameters, modifying the network structure, or adding regularization terms.

[0070] After comparing the results of multiple experiments, the model with the best performance on the validation set is selected as the final model. To verify the reliability of the final model, it should also be evaluated on an independent test set to ensure good generalization. Throughout this process, hyperparameters are continuously adjusted until the optimal model configuration is found.

[0071] The specific steps of step F are as follows: In order to obtain the true geographic location of the detection target, this method needs to comprehensively consider the positioning information and orientation of the acquisition device, as well as the location of the target medium and low voltage power distribution equipment in the image. This method designs a relocation technology for the target medium and low voltage power distribution equipment. Through the deflection angle captured by the acquisition device, the detection target and image parameters, the latitude and longitude of the detection target are calculated. The steps to obtain the true geographic location of the detection target are as follows:

[0072] First, obtain the longitude X and latitude Y of the acquisition device;

[0073] (1) Obtain the xyz-axis deflection angle of the acquisition device through the nine-axis sensor, thereby obtaining the orientation of the acquisition device, Figure 5 It is assumed that the deflection angle of the acquisition device relative to the true north is α;

[0074] (2) The depth information of the target medium and low voltage equipment detected in the visible light image is obtained after registration in the point cloud map. The point cloud beyond the detection range is removed (the depth information exceeds the point cloud detection range, that is, the distant background is removed). The median of the remaining depth points is taken to obtain the depth d of the acquisition device from the detection target.

[0075] (3) The horizontal field of view angle of the visible light image in the acquisition device is β, and the deflection angle of the detection target relative to the true north is γ;

[0076] (4) The deflection angle of the detection target relative to the acquisition device is γ = α + [cx / width - 0.5] * β; where w is the image width and cx is the x-axis coordinate of the detection target in the image;

[0077] (5) Calculate the longitude of the detection target bx = X + d*sinγ, and the latitude of the detection target by = Y + d*cosγ

[0078] Then generate an electronic handover sheet and integrate the processed detection target information, including location, type, etc., into a unified format to facilitate subsequent data visualization, management and archiving.

[0079] Through the above steps, the location accuracy and information integrity of the model output results are ensured, providing a reliable basis for operation and maintenance decisions of medium and low voltage distribution equipment.

[0080] The specific step G is as follows: in the result visualization and information management stage, it is first necessary to combine the collected medium and low voltage power distribution equipment information with the geographic information system (GIS), use the three-dimensional reconstruction algorithm to generate a three-dimensional point cloud model of the equipment, and embed it into the geographic environment to create an intuitive three-dimensional layout map, so that users can freely browse and query the medium and low voltage power distribution equipment in the layout map on the Web-based intelligent medium and low voltage layout management platform. The Web-based intelligent medium and low voltage layout management platform in this method uses the Neo4j relational database, Alibaba Cloud GPU cloud server and front-end technology to build it, where the Neo4j database is used to store multi-source information of medium and low voltage power distribution equipment, and uses the graph database to build a connection relationship network between devices, which is convenient for tracking logical connections and physical connections between devices. In addition, the user interface includes functions such as equipment viewing, equipment management, and decision support, so that operation and maintenance personnel can easily view equipment status, location information, historical records, etc., and provide auxiliary functions such as data integration, intelligent analysis, and decision support, such as Figure 6 As shown, in this way, transparent and integrated management of layout data of medium and low voltage distribution equipment is achieved.

[0081] Compared with the prior art, the present invention has the following technical effects:

[0082] (1) Innovation of operating procedures

[0083] This method changes the traditional manual medium and low voltage distribution equipment layout drawing mode through a new method and operation process for modeling and positioning medium and low voltage distribution equipment, and proposes a new method of intelligent collection and drawing along the layout. First, a mobile acquisition device is used, which integrates a point cloud camera, an infrared camera, a visible light camera, a 5G+RTK hybrid positioning system, a built-in power supply, a high-density storage device, an SOC control module and an IMU nine-axis sensor. The device collects multi-source data of the target medium and low voltage distribution equipment and stores the data on a high-density device; secondly, the collected data is cleaned and filtered to filter out high-quality images and positioning data; then, a medium and low voltage distribution equipment training sample set is generated through data preprocessing; a detection model for medium and low voltage distribution equipment is designed, which can be used to detect the target medium and low voltage distribution equipment. The detection model is trained using the above training sample set and its parameters are continuously adjusted and optimized to select the optimal model as the final model. The model is then used to detect medium and low voltage distribution equipment. Based on the aligned depth map and the positioning information of the acquisition device, an electronic handover form is generated and uploaded to the intelligent medium and low voltage distribution equipment layout management platform. Finally, based on the electronic handover form, a medium and low voltage distribution equipment layout is created on the intelligent medium and low voltage distribution equipment layout management platform. This platform has functions including equipment viewing, equipment management, and decision support, and provides auxiliary functions such as three-dimensional information visualization, data integration, intelligent analysis, and decision support.

[0084] (2) Innovation of medium and low voltage distribution detection model structure

[0085] In current application scenarios, traditional detection models face multiple technical limitations. For example, they struggle with inaccurate device recognition in complex environments, particularly when devices vary in size and shape, are susceptible to occlusion, and face complex backgrounds. Traditional models suffer from low recognition accuracy and recall. Furthermore, existing models suffer from high computational resource consumption and inability to achieve real-time data processing when deployed on lightweight devices. Therefore, this method addresses these issues during model design. First, this method introduces the backbone network from YOLOv8 to replace the traditional ResNet50, enhancing the model's ability to capture underlying specific features (such as shape, texture, and color), thereby improving the model's ability to identify subtle line faults. Second, this method designs an efficient multi-scale feature fusion strategy, combining feature maps at different scales to ensure the model can handle medium and low voltage distribution equipment targets of various scales. This strategy also enhances the model's global perception capabilities, reduces interference from the environment in which the equipment resides, and improves the model's robustness. To further improve the model's generalization, this method incorporates data augmentation techniques such as random cropping, rotation, scaling, and translation to increase the diversity of the training data. These improvements enable the model to not only accurately identify device categories, but also reduce the deviation between the device prediction box and the target, thereby effectively addressing the multi-scale target recognition challenges of traditional models in complex environments.

[0086] (3) Relocation technology of target medium and low voltage distribution equipment

[0087] Existing methods rely on manual hand-drawing of maps, which can lead to errors and record the location of the data collector, rather than the true location of the medium and low voltage distribution equipment. This method calculates the deflection angle and image parameters captured by the data collection device and uses a medium and low voltage distribution equipment relocation algorithm to determine the true location of the target medium and low voltage distribution equipment. This combination of technologies makes the positioning of medium and low voltage distribution equipment more accurate, effectively avoiding the errors and waste of human resources caused by manual operation and achieving an automated process of simultaneous data collection and mapping.

[0088] In summary, the technological innovation of the "Medium and Low Voltage Distribution Equipment Detection Model" not only solves the inefficient and error-prone manual data entry problem of traditional methods, improving the accuracy and efficiency of data collection, but also enables high-precision dynamic positioning and real-time mapping, enhancing the accuracy and timeliness of data. More importantly, it provides a reliable basis for operation and maintenance decisions for medium and low voltage distribution equipment, greatly improving the safety and efficiency of power grid operations and maintenance, and promoting the transformation of the power industry towards intelligent systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 is a flow chart of the present invention;

[0090] Figure 2It is a structural diagram of the backbone network in the low-voltage power distribution equipment detection model of the present invention;

[0091] Figure 3 It is a structural diagram of a hybrid high-efficiency encoder in a low-voltage power distribution equipment model of the present invention;

[0092] Figure 4 It is a structural diagram of the decoder in the low-voltage distribution equipment detection model of the present invention;

[0093] Figure 5 is a schematic diagram of the deflection angle captured by the acquisition device of the present invention;

[0094] Figure 6 It is a schematic diagram of the low-voltage layout management platform of the present invention. DETAILED DESCRIPTION

[0095] The specific technical solutions of the present invention are described with reference to the embodiments.

[0096] Example 1:

[0097] In practical applications, such as Figure 1 As shown in the figure, the operation process of the intelligent detection model for medium and low voltage distribution equipment is as follows:

[0098] First, a mobile data acquisition device was used for on-site data collection. This device integrates an infrared camera (640×512 resolution, 30 frames per second) to detect the device's thermal distribution, a visible light camera (2560×1440 resolution, 30 frames per second) to capture the device's appearance features, and a 3D point cloud camera to generate high-density point cloud data for constructing a 3D model. The device is also equipped with a 5G+RTK hybrid positioning system (centimeter-level accuracy) and an IMU nine-axis sensor to record the device's attitude and position information in real time. A built-in power supply and high-density storage device ensure continuous operation in offline environments.

[0099] After acquisition, an adaptive image quality assessment algorithm (combining gradient intensity analysis and template matching) automatically selects high-quality images and removes blurred, occluded, or drifted data. Data preprocessing then follows, including a gray-world white balance algorithm to correct color deviations, bilateral filtering for denoising, a Lucy-Richardson iterative deblurring algorithm to enhance clarity, and multi-source image registration based on SIFT feature point matching. The SeAFusion framework is used in the fusion stage to extract multi-scale features through ResNet. A self-attention mechanism is then used to weightedly fuse infrared, visible light, and point cloud data to generate a comprehensive image that combines thermal distribution, texture, and spatial information.

[0100] The model training is based on the improved RT-DETR architecture, replacing the backbone network with the backbone network of YOLOv7-M to enhance the feature extraction capability, such as Figure 2As shown in Figure 2, a multi-scale feature fusion strategy was introduced to optimize detection performance for devices of different sizes. During training, a dynamically weighted loss function (classification loss, bounding box regression loss, and distributed focal loss) was used, combined with the AdamW optimizer and an exponentially decaying learning rate (initial 1e-4, decay factor ρ = 0.95). Hyperparameters were adjusted through grid search and random search, and early stopping was used to prevent overfitting. The final model achieved a mean average performance (MAP) of 52.3% on the validation set and demonstrated good generalization performance on the test set. Figure 3 4 is a structural diagram of a hybrid high-efficiency encoder in a medium- and low-voltage power distribution equipment model of a medium- and low-voltage power distribution equipment detection method according to an embodiment of the present invention. Figure 4 It is a structural diagram of a decoder in a medium and low voltage power distribution equipment detection model of a medium and low voltage power distribution equipment detection method according to an embodiment of the present method.

[0101] Target relocation phase, such as Figure 5 Based on the longitude and latitude (X, Y) of the acquisition device, the IMU sensor deflection angle α, and the target coordinates (cx) in the image, combined with the depth information d and the horizontal field of view angle β, the true geographic location of the device is calculated using the geometric formula (bx = X + d·sinγ, by = Y + d·cosγ), with an error control within ±0.1 meters. The results are integrated into the device type, location, and thermal status through an electronic handover form and uploaded to the web management platform based on the Neo4j graph database, such as Figure 6 The platform supports 3D layout visualization (embedded point cloud models), real-time monitoring of equipment status, and intelligent analysis. Operations and maintenance personnel can quickly locate overheating equipment or abnormal nodes through the interface to assist in making maintenance decisions.

[0102] In a pilot test at the Information Center of the Huizhou Power Supply Bureau in Guangdong Province, the model increased the efficiency of traditional manual inspections by four times, reduced the missed detection rate from 8.7% to 1.2%, improved positioning accuracy by 90% compared to GPS, and shortened data update delays from hours to minutes, significantly reducing operation and maintenance costs and improving grid security, verifying the practicality and advancement of the technical solution.

[0103] Example 2:

[0104] In the medium and low voltage power distribution systems of industrial parks, equipment is densely distributed and the environment is complex (such as high temperature, electromagnetic interference, equipment obstruction, etc.). Traditional detection methods are difficult to meet the needs of efficient operation and maintenance. This embodiment implements intelligent detection through the following process:

[0105] First, a customized mobile data acquisition device was deployed. Its infrared camera was upgraded to 1024×768 resolution (25 frames per second) to enhance its ability to capture subtle temperature rises in equipment. The visible light camera uses HDR mode (3840×2160 resolution) to provide clear imaging of highly reflective metal surfaces. The point cloud camera uses a solid-state lidar, increasing the point cloud density to 2 million points per second, ensuring accurate 3D reconstruction in complex obstructed environments. The device integrates dual-frequency RTK (supporting L1 / L5 bands) and a 5G private network, achieving centimeter-level dynamic positioning (±5cm positioning error) in areas of electromagnetic interference. A nine-axis IMU sensor also compensates for attitude deviations caused by mechanical vibration in real time.

[0106] After data collection, an improved adaptive image quality assessment algorithm is used to address common oil and dust interference in industrial scenarios. Based on gradient intensity analysis, a local contrast enhancement (CLAHE algorithm) preprocessing algorithm is introduced to improve the screening accuracy of low-light or high-noise images. During the data cleaning phase, an abnormal point cloud filtering module is added to eliminate outliers through the DBSCAN clustering algorithm to retain valid equipment structure information. During the preprocessing phase, the gray world white balance algorithm is dynamically adjusted in combination with the equipment material library (such as reflectivity parameters of copper busbars and insulators) to avoid color casts caused by metal reflections. The deblurring algorithm uses end-to-end restoration based on the generative adversarial network (DeblurGAN-v2) to significantly improve the clarity of motion-blurred images.

[0107] In terms of model training, in view of the large differences in equipment size in industrial parks, the multi-scale feature fusion module of RT-DETR was upgraded to a cross-modal attention mechanism (Cross-Modality Attention), which simultaneously integrates infrared thermal distribution, visible light texture and point cloud spatial features. The backbone network adopts the YOLOv8-P6 structure, supports higher-resolution input (1280×1280 pixels), and introduces a dynamic label assignment strategy (Task-Aligned Assigner) to improve the detection recall rate of small target devices (such as fuses and terminal blocks). Training data enhancement has added a synthetic algorithm that simulates industrial smoke and oil pollution to enhance model robustness. The optimized model achieved a mAP of 94.1% on the test set, an increase of 6.3% over the baseline, and the inference speed was maintained at 45FPS (NVIDIA Jetson AGX Orin platform).

[0108] In the target relocation phase, a multi-sensor collaborative calibration algorithm was designed to address positioning interference issues in densely populated equipment clusters. This algorithm verifies visible light image detection results using the spatial topology of point cloud data to eliminate mismatches between adjacent devices. It also combines RTK positioning trajectories with SLAM (Simultaneous Localization and Mapping) technology to construct a high-precision semantic map, limiting absolute positioning errors within ±0.15 meters. Electronic handover forms automatically link historical equipment inspection records with real-time thermal imaging data. The management platform's AI diagnostic module (based on LSTM time series models) predicts equipment aging trends and generates maintenance recommendations in advance.

[0109] During a trial at the Huaiji Power Supply Bureau in Zhaoqing, the system achieved 48-hour, full-coverage inspection of 1,852 medium- and low-voltage equipment across the district. This system increased efficiency sixfold compared to traditional manual inspections, reduced the false detection rate from 12.4% to 0.8%, and successfully issued warnings of overheating hazards at three cable joints. After data was synchronized to the cloud-based management platform, operational and maintenance response time was reduced to under 10 minutes, demonstrating the technology's reliability and advancement in complex industrial scenarios.

Claims

1. A modeling method for medium and low voltage power distribution equipment, characterized in that: The following steps are involved: A. Acquire multi-source data through acquisition devices: First, use mobile acquisition devices to collect multi-source information of medium and low voltage distribution equipment; B. Data cleaning and screening: Detailed cleaning and screening of collected data to eliminate abnormal and irrelevant data, and select high-quality images and accurate data positioning; C. Data preprocessing: Generate medium and low voltage distribution equipment sample sets through image preprocessing technology; D. Design and train models: Design customized detection models for medium and low voltage distribution equipment based on the characteristics of the equipment and its environment, and use the designed models for training. E. Parameter tuning: Adjust hyperparameters, repeat training multiple times, retain the results, and select the optimal model as the final model; F. Data Post-Processing: Post-process the model output results, relocate the actual position of the medium and low voltage power distribution equipment based on the positioning data of the mobile acquisition device and the relative position of the medium and low voltage power distribution equipment obtained based on the point cloud data, and associate the actual position information with the target medium and low voltage power distribution equipment to generate an electronic handover form; G. Result visualization and information management: Establish a layout of medium and low voltage power distribution equipment, and provide a platform for visual display, management and decision support of medium and low voltage power distribution equipment.

2. The method for modeling medium and low voltage power distribution equipment according to claim 1, characterized in that: The mobile acquisition device in step A integrates a point cloud camera, an infrared camera, a visible light camera, a 5G+RTK hybrid positioning system, a built-in power supply, a high-density storage device, a SOC control module and an IMU nine-axis sensor.

3. The method for modeling medium and low voltage power distribution equipment according to claim 1, characterized in that: Specifically, step B includes: in the data screening and cleaning stage, firstly, performing an automated data cleaning task on the collected raw data, screening and removing low-quality images and positioning drift points through algorithms; using image processing algorithms to automatically detect and remove blurred and distorted images caused by insufficient lighting, occlusion, or device movement; and identifying and correcting positioning drift points caused by signal interference or environmental factors through a positioning data quality control mechanism; performing a data consistency check to ensure that data collected by different sensors match each other; finally, manually labeling the screened samples and creating a multi-source dataset for training; Among them, the algorithm used to automatically screen image quality is an adaptive image quality assessment and screening algorithm, and its implementation steps are as follows: (1) Read image: read each original image to be processed from the dataset; (2) Preprocessing: convert the image into grayscale image; Evaluate image clarity: calculate the gradient amplitude of the image; use the Canny edge detection method to obtain the gradient information of the image; Count the distribution of gradient intensity in the entire image or key area; set a threshold T1 to determine whether the image is clear enough; if the overall gradient intensity of the image is lower than T1, mark the image as blurred; Check image integrity: Use template matching technology to compare the current image with known good samples. Set a threshold T2 based on the matching results to determine the integrity of the image content. If the matching score is lower than T2, the image is considered to be occluded or otherwise damaged. Comprehensive decision: If the image is judged to be both blurry and lacking sufficient content integrity, it is marked as a low-quality image; otherwise, the image is retained for subsequent processing; Output: Based on the above results, a new dataset containing all qualified images is generated, and those images marked as low quality are removed; Parameter Tuning and Optimization: Training a classification model to automatically generate these thresholds improves automation and accuracy.

4. The method for modeling medium and low voltage power distribution equipment according to claim 1, characterized in that: The step C specifically includes: performing image white balance, image denoising, image deblurring, image registration and image fusion processing in the data preprocessing stage; First, image white balance is performed to eliminate color deviation caused by changes in lighting conditions. A gray world hypothesis algorithm is used to calculate the average brightness of the image and adjust the RGB channels to make them tend to the standard white point. Next, image denoising is performed, using a bilateral filtering algorithm to remove random noise from the image. This algorithm removes noise while retaining edge details. Then, the image is deblurred, and the image clarity is gradually restored by estimating the blur kernel using an iterative deblurring method based on the Lucy-Richardson algorithm. In terms of image registration, the SIFT algorithm is used to find corresponding points between the visible light and infrared images, and the transformation matrix is ​​estimated through the least squares method to achieve image alignment between the infrared and visible light images. Since the point cloud data comes from the binocular camera, the point cloud image has already been registered with the visible light image. Finally, SeAFusion is combined with an advanced vision image fusion framework to merge the image information acquired by multiple sensors into a single image, that is, multi-scale feature extraction and self-attention mechanism are used to select and fuse data from different sensors.

5. The method for modeling medium and low voltage power distribution equipment according to claim 4, characterized in that: The fusion steps are as follows: (1) Multi-scale feature extraction: Use deep convolutional neural network ResNet to extract multi-scale features from images of different sensors; feature extraction is expressed as: f(x)=Conv(x) Where f is the feature extraction function, x is the input image, and Conv is the convolutional layer; (2) Self-attention mechanism: The extracted features are weighted through the self-attention mechanism to emphasize those parts that contribute most to the fusion results; (3) Feature fusion: The weighted feature maps are fused to generate a fused image. Feature fusion is expressed as: Fused=αf1+βf2+...+γf n where f1, f2, ..., f n are the characteristic maps of different sensors, α, β, ..., γ are weighting coefficients; (4) Residual learning: Residual blocks are introduced to learn the residual between input features and fusion features to ensure that important information is preserved. Residual learning is expressed as: Residual=Fused+Input Input is the input image. The final result is a single image that fuses multi-sensor information; this image contains multi-source information from different sensors.

6. The method for modeling medium and low voltage power distribution equipment according to claim 1, characterized in that: Step D: Designing and training a model: designing a customized model based on the characteristics of the medium and low voltage equipment and its surrounding environment, and using the designed model for training; The specific operation is divided into steps D1 design model and D2 training model; The step D1 is specifically as follows: according to the characteristics of the medium and low voltage power distribution equipment and its surrounding environment, the medium and low voltage power distribution equipment is detected by improving RTDETR: First, we introduce a more powerful backbone network in YOLOv8 to replace the original backbone network ResNet50 of RTDETR to enhance the ability to capture device-specific features; Secondly, an efficient multi-scale feature fusion strategy is adopted to combine feature maps of different scales to ensure that the model can handle objects of various scales while maintaining high accuracy. Data augmentation technology will also be incorporated to increase the diversity of training data. Finally, when designing the loss function for the medium and low voltage distribution equipment detection model, classification loss, distributed focus loss, and bounding box regression loss were used; The total training loss function is the weighted sum of the above three loss functions; The step D2 specifically includes: during the model training process, using the labeled medium and low voltage distribution equipment detection model multi-source dataset, using the PyTorch deep learning library to build an improved RT-DETR model, and initializing the model parameters; during the training phase, using the AdamW optimizer and a dynamic learning rate adjustment strategy, the initial learning rate is set to 1e-4, and the learning rate is automatically adjusted according to the performance indicators on the validation set; Specifically, the training process of the medium and low voltage distribution equipment detection model is as follows: (1) Read m images X = {x1, x2, ..., xm} and their corresponding category labels and ground truth coordinate labels Y = {y1, y2, ..., ym} from the medium and low voltage distribution equipment dataset, where m represents the batch size and defaults to 8; preprocess the read images; (2) The image data X is passed into the backbone network to extract features f1, f2, and f3 of different scales in the image; (3) These features are passed into a hybrid efficient encoder to output multi-scale fusion features η1, η2, η3; the design of the hybrid encoder uses an attention mechanism to perform internal scale interactions of high-level features; (4) The multi-scale fusion features η1, η2, η3 are passed to the decoder; (5) According to the coordinates and categories of the predicted box and the true box, the classification loss, distributed focus loss, and bounding box regression loss are calculated; among them, the classification loss is used to measure the gap between the predicted category and the true category, the bounding box regression loss is used to optimize the positioning performance of the model to ensure that the model can accurately mark the location of the object, and the distributed focus loss is used to predict the discrete distribution of the bounding box parameters to improve the positioning accuracy of the model; (6) Multiply the classification loss, regression box loss, and intersection-over-union loss by their respective dynamic weights w dfl 、w cls 、w iou , get the total training loss L, and calculate the gradient of the model parameters based on L through back propagation. This gradient represents the direction and speed of change of the loss function under the current parameter value; then use the stochastic gradient descent optimization function to update the model parameters; dynamic weights are used to balance the impact of each loss on the model back propagation process. The initial w iou 2, w cls is 5, w dfl is 1, and the loss weight is adjusted according to the progress of training, and finally w iou 6.85, w cls is 1.0, w dfl is 1.25, which can be expressed as: L=w cls L cls +w dfl L dfl +w iou L iou L cls =-w n [l n logσ(s n )+(1-l n )log(1-σ(s n ))] In L iou In, L IoU represents the intersection-over-union loss, A and B represent the candidate box and the real box, x and y represent the center coordinates of the candidate box, gt and y gt represents the center coordinates of the ground-truth box, W and H are the smallest rectangles that can enclose A and B, α and σ are two preset hyperparameters, with the default values ​​of α = 1.6 and σ = 4; The exponential moving average of driving momentum is a dynamic value; Represents the monotonic focusing coefficient, which is used to measure the focusing effect; in L dfl In, p t and p t+1 is the probability of the two positions closest to the label y in the predicted distribution, and β is an adjustment factor; in L cls In it, σ is the sigmoid activation function, Q is the weight of the nth sample, l is the true label, and s is the output of the model; (7) Repeat steps (1) to (6) until the preset number of iterations is reached or the early stopping termination condition is met; (8) Filter the detection boxes according to the confidence level and delete the detection boxes that are less than the confidence threshold; (9) Based on the detection boxes and the real annotations in the dataset, calculate a series of accuracy indicators, including precision, recall, and average precision, evaluate the model performance, and record the results and model hyperparameters; (10) Adjust the hyperparameters, repeat steps (1) to (9), and compare the results multiple times until the optimal result is obtained.

7. The method for modeling medium and low voltage power distribution equipment according to claim 1, characterized in that: The step E specifically includes: first determining the hyperparameters that need to be adjusted during model training, enumerating all candidate value combinations of the hyperparameters, and training the model one by one to evaluate the performance; using a random search algorithm to randomly select hyperparameter combinations for testing; during the training process, adopting a learning rate decay strategy, using an exponential decay method; During training, the model's performance on the validation set is regularly evaluated, and the results are recorded multiple times. Early stopping is introduced during training, terminating training early if performance on the validation set does not improve for several consecutive epochs. The model is then fine-tuned based on the evaluation results. By comparing the results of multiple experiments, the model with the best performance on the validation set is selected as the final model. To verify the reliability of the final model, it should also be evaluated on an independent test set to ensure that the model has good generalization ability. Throughout the process, hyperparameters are continuously adjusted until the optimal model configuration is found.

8. The method for modeling medium and low voltage power distribution equipment according to claim 1, characterized in that: The step F is specifically as follows: to design a relocation technology for target medium and low voltage power distribution equipment, the steps of detecting the target and image parameters by collecting the deflection angle captured by the device, calculating the longitude and latitude of the detected target, and obtaining the real geographical location of the detected target are as follows: First, obtain the longitude X and latitude Y of the acquisition device; (1) Obtain the xyz-axis deflection angle of the acquisition device through the nine-axis sensor, thereby obtaining the orientation of the acquisition device, and let the deflection angle of the acquisition device relative to the true north be α; (2) Obtain the depth information of the point cloud after registration by detecting the target medium and low voltage equipment in the visible light image, remove the point cloud beyond the detection range, and take the median of the remaining depth points to obtain the depth d of the acquisition device from the detection target; (3) The horizontal field of view angle of the visible light image in the acquisition device is β, and the deflection angle of the detection target relative to the true north is γ; (4) The deflection angle of the detection target relative to the acquisition device is γ = α + [cx / width - 0.5] * β; where w is the image width and cx is the x-axis coordinate of the detection target in the image; (5) Calculate the longitude of the detection target bx = X + d*sinγ, and the latitude of the detection target by = Y + d*cosγ; Then generate an electronic handover sheet to integrate the processed inspection target information into a unified format.

9. The method for modeling medium and low voltage power distribution equipment according to claim 1, characterized in that: The specific steps of step G are as follows: in the result visualization and information management stage, the collected medium and low voltage power distribution equipment information is first combined with the geographic information system, and a three-dimensional point cloud model of the equipment is generated using a three-dimensional reconstruction algorithm. The model is embedded in the geographic environment to create an intuitive three-dimensional layout map, allowing users to freely browse and query the medium and low voltage power distribution equipment in the layout map on the web-based intelligent medium and low voltage layout management platform; in this way, transparent and integrated management of the layout data of medium and low voltage power distribution equipment is achieved.

Images