A safety monitoring method for complex substation environment based on high dynamic target enhancement

Abstract

The application discloses a kind of complex substation environment safety monitoring methods based on high dynamic target enhancement. Including obtaining the video and image data of substation under various working conditions as standard sample data set;Standard sample data set is processed to basic data, and standard sample data set is divided into training set, test set and verification set;The target detection model is constructed, the data of training set are input into the target detection model for training, and the parameters of the target detection model are optimized through verification set and test set;Real-time image in substation is analyzed using the target detection model trained, and the detection result is output.The application can monitor the state of substation under various working conditions, with the advantages of high recognition rate and strong accuracy.

Description

Technical Field

[0001] This invention relates to the field of substation environmental safety monitoring technology, specifically a method for monitoring the environmental safety of complex substations based on high dynamic target enhancement. Background Technology

[0002] With the expansion of the power grid and the growth of electricity demand, ensuring the safe, reliable, and stable operation of the power grid has become particularly important. As a key node connecting the main grid and the distribution network, the safety status of substations directly affects the reliability of the power grid. Common safety hazards in substations include fires, personnel not wearing safety helmets, and abnormal instrumentation. If these problems are not addressed promptly, they can lead to serious economic losses and casualties. Traditional fire alarm systems typically rely on sensors that detect specific thresholds, such as temperature or smoke sensors. However, these systems suffer from delayed responses and are unable to detect early signs of fires. With the rapid development of deep learning technology, video-based fire monitoring technology is gradually becoming feasible.

[0003] Deep learning models such as convolutional neural networks (CNNs) have significantly improved the accuracy of flame detection and safety helmet recognition. These models can automatically learn feature representations in complex scenes, accurately distinguishing flames from flame-like objects and reducing false positives and false negatives. Furthermore, deep learning can also perform human pose estimation and feature extraction, accurately determining whether a worker is wearing a safety helmet even without a frontal face or complete human body information, thus overcoming the limitations of traditional methods.

[0004] However, the performance of deep learning models largely depends on the diversity and quality of the training data. Data augmentation techniques have proven to be an effective means of increasing the diversity of training data and improving the generalization ability of deep learning models. By performing a series of transformations on the training data, such as rotation, scaling, cropping, and color adjustment, data augmentation can provide more data variations during model training. This is particularly important for applications such as substations, where monitoring images are often affected by factors such as changes in lighting, seasonal variations, and equipment aging. Although existing deep learning models perform well in many scenarios, they face specific challenges in the special environment of substations, such as the need for large amounts of labeled data, complex background interference, and the problem of accurately detecting flames and safety helmets.

[0005] Therefore, with limited data, how to effectively utilize data augmentation techniques to improve the detection accuracy of deep learning models in complex substation environments has become an important problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a method for environmental safety monitoring of complex substations based on high dynamic target enhancement. This invention can monitor the status of substations under various operating conditions and has the advantages of high recognition rate and high accuracy.

[0007] The technical solution of this invention: A method for environmental safety monitoring of complex substations based on high dynamic target enhancement, comprising the following steps:

[0008] Step S1: Data Acquisition: Acquire video and image data of the substation under various operating conditions as a standard sample dataset;

[0009] Step S2: Data preprocessing: Perform basic data processing on the standard sample dataset, dividing the standard sample dataset into training set, test set and validation set;

[0010] Step S3: Model Construction: Construct an object detection model, input the training set data into the object detection model for training, and optimize the object detection model parameters through validation and test sets; among them, implement an online dynamic image enhancement method to adjust the attributes of training images in the training set; introduce a coordinate-guided multi-granularity enhanced feature fusion module in the neck structure of the object detection model to enhance the model's ability to handle objects in complex environments; and adopt a dynamic shape similarity boundary intersection-union loss function that introduces shape similarity measurement and dynamic weight adjustment mechanism to improve the performance of the object detection model.

[0011] Step S4: Safety Monitoring: Analyze real-time images within the substation using the trained target detection model and output the detection results.

[0012] In the above-mentioned method for environmental safety monitoring of complex substations based on high dynamic target enhancement, step S1 involves using the photography equipment in the substation and existing databases to collect video and image data of the substation under various operating conditions, which are then used as a standard sample dataset.

[0013] In the aforementioned method for environmental safety monitoring of complex substations based on high dynamic target enhancement, step S1 includes image data of flames, image data of personnel wearing safety helmets, and image data of instrument status.

[0014] In the aforementioned method for environmental safety monitoring of complex substations based on high dynamic target enhancement, step S2 includes basic data processing, which includes improving image clarity, removing image noise, and standardizing image size.

[0015] In the aforementioned method for environmental safety monitoring of complex substations based on high dynamic target enhancement, step S2, which involves dividing the standard sample dataset into training, testing, and validation sets, includes automatically screening the standard sample dataset, removing overly blurry images, manually correcting them using labelimg software, and labeling them with tags. The standard sample dataset is divided into training, validation, and testing sets in a ratio of 8:1:1 to ensure the balance and representativeness of the image data distribution.

[0016] In the aforementioned method for environmental safety monitoring of complex substations based on highly dynamic target enhancement, step S3 involves constructing a target detection model, which includes implementing online dynamic image enhancement technology to dynamically adjust the attributes of training set images; introducing a coordinate-guided multi-granularity enhanced feature fusion module into the target detection model to enhance its ability to handle targets in complex environments; and employing a dynamic shape similarity boundary intersection-union ratio loss function that incorporates shape similarity measurement and dynamic weight adjustment mechanism to improve the performance of the target detection model.

[0017] The aforementioned method for environmental safety monitoring of complex substations based on high dynamic target enhancement includes the following steps in step S3: First, loading the labeled standard sample dataset images, extracting flame images and their label files, and refining the flame morphology; then, superimposing the flame images onto the target background image, performing image fusion processing, and updating the corresponding label files according to the new position and size of the flame in the fused image; finally, generating a controllable number of random points for each flame image, calculating the movement of a random vector direction for each random point based on the maximum displacement parameter of the random points, obtaining the target position of the random point, and deforming the fused image through local affine transformation to generate a flame image with high variability.

[0018] The aforementioned method for environmental safety monitoring of complex substations based on high-dynamic target enhancement, in step S3, involves introducing a coordinate-guided multi-granularity enhanced feature fusion module into the neck structure of the target detection model to enhance the model's ability to handle complex environmental targets. This includes the following steps: In the neck structure of the target detection model, the input feature matrix is ​​divided into two parts; one part is subjected to a CoordConv operation; x and y coordinate channels are generated and convolved with the input feature map to enhance the spatial location information of the feature map; then, a linear transformation is performed on the feature map after the CoordConv operation to generate a query and key matrix; a standard convolution operation is performed on the other part of the input feature matrix; and the feature map is extracted through the convolution operation. The original features are linearly transformed to generate a value matrix. After generating the query, key, and value matrices, convolution operations are performed on the query, key, and value matrices respectively to further enhance feature extraction. A second dimensionality reduction and convolution operation is performed on the high-dimensional features to obtain mid-dimensional features. Finally, a third dimensionality reduction and convolution operation is performed on the mid-dimensional features to obtain low-dimensional features. The three-dimensional features of the convolved query, key, and value matrices are input into a multi-head self-attention mechanism. By splitting them into 8 heads, attention weights and weighted values ​​are calculated separately to capture long-distance dependencies and global context information in the feature map. The three-dimensional features are then subjected to a unified-dimensional convolution operation to improve the effectiveness of feature fusion. Finally, maximum value fusion is performed to obtain the fused feature representation.

[0019] In the aforementioned method for environmental safety monitoring of complex substations based on highly dynamic target enhancement, the specific calculation formula for the dynamic shape similarity boundary intersection-union ratio, which incorporates shape similarity measurement and dynamic weight adjustment mechanism, in step S3 is as follows:

[0020] H(A,B)=max(h(A,B),h(B,A));

[0021]

[0022] ShapeSim(A,B)=exp(-δ*H(A,B));

[0023] Where H(A, B) is the Hausdorff distance between set A and set B; h(A, B) is the maximum distance from each point in set A to the nearest point in set B; h(B, A) is the maximum distance from each point in set B to the nearest point in set A; δ is the adjustment parameter;

[0024] The weight parameters in the loss function are adjusted in real time based on the aspect ratio and area of ​​the target; the formula for the dynamic weight adjustment mechanism is as follows:

[0025]

[0026] Where Wpred is the width of the predicted bounding box; hpred is the height of the predicted bounding box; wgt is the width of the ground truth bounding box; hgt is the height of the ground truth bounding box; Apred is the area of ​​the predicted bounding box; A gt η is the area of ​​the true bounding box; η and ζ are adjustment parameters; r is the aspect ratio of the target.

[0027] The traditional loss function is calculated using the following formula:

[0028]

[0029] Where p is the distance between the center points of the predicted bounding box and the ground truth bounding box; c is the diagonal length of the smallest closed region enclosing the two bounding boxes; C w C is the length of the closed region of width. h b is the length of the closed region with height; b is the coordinate of the center point of the predicted bounding box; b gt is the coordinates of the center point of the ground truth bounding box; w is the width of the predicted bounding box; w gt h is the width of the ground truth bounding box; h is the height of the predicted bounding box; gt The height of the actual bounding box;

[0030] The expression for the dynamic shape similarity boundary intersection-union ratio loss function, which incorporates shape similarity measurement and dynamic weight adjustment mechanism, is as follows:

[0031] Loss=α*L EIoU +β*ShapeSim(A, B)+γ*DynWeight(r, A);

[0032] Where α is the weight coefficient of the traditional loss function; β is the weight coefficient of the shape similarity measure; and γ is the weight coefficient of the dynamic weight adjustment.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This invention acquires video and image data of substations under various operating conditions as a standard sample dataset; performs basic data processing on the standard sample dataset; constructs a target detection model, trains the target detection model, and optimizes the target detection model parameters; uses the trained target detection model to analyze real-time images within the substation and outputs detection results; this invention can monitor the status of substations under various operating conditions, thereby improving substation safety; by using the trained target detection model to perform real-time analysis and detection of images in the substation, it can quickly identify potential safety hazards, issue timely warnings, and reduce the risk of safety accidents.

[0035] 2. This invention employs Online Dynamic Image Enhancement (ODIA) technology, which effectively increases data diversity and improves the generalization ability of the object detection model under limited data conditions. The ODIA strategy not only includes DMDA, such as flame fusion techniques and dynamic changes, but also integrates conventional enhancement techniques, thereby increasing the diversity of training data and enabling the object detection model to better adapt to various visual changes.

[0036] 3. The coordinate-guided multi-granularity enhanced feature fusion module (CMGE-FFM) of this invention significantly enhances the object detection model's ability to capture spatial location information and multi-scale features. Through multi-head self-attention mechanism and multiple dimensionality reduction feature extractions, the object detection model can better handle complex scenes and targets of different scales, improving detection accuracy and robustness. Compared with traditional object detection models, this invention achieves more refined feature extraction and more effective feature fusion, significantly improving the detection performance of the object detection model in complex scenes.

[0037] 4. The DSS-IoU loss function proposed in this invention combines the latest shape similarity measurement and dynamic weight adjustment mechanism. It not only has significant advantages in geometric attribute optimization, but also makes the target detection model perform better in complex environments for high aspect ratio or small targets by comprehensively capturing the shape and size features of the target. Attached Figure Description

[0038] Figure 1 This is a flowchart of the present invention;

[0039] Figure 2 This is a structural diagram of the object detection model;

[0040] Figure 3 This is a structural diagram of ODIA;

[0041] Figure 4 This is a flowchart of a coordinate-guided multi-granularity enhanced feature fusion module;

[0042] Figure 5 This is the PR curve of the target detection model proposed in this invention. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0044] Example: A method for environmental safety monitoring of complex substations based on high dynamic target enhancement, such as... Figure 1 As shown, it includes the following steps:

[0045] Step S1: Data Acquisition: Acquire video and image data of the substation under various operating conditions as a standard sample dataset.

[0046] In step S1, video and image data of the substation under various operating conditions are collected using the camera equipment in the substation and the existing database, which serve as a standard sample dataset.

[0047] In step S1, the standard sample dataset includes image data of flames, image data of personnel wearing safety helmets, and image data of instrument status.

[0048] Step S2: Data preprocessing: Perform basic data processing on the standard sample dataset, dividing the standard sample dataset into training set, test set and validation set.

[0049] In step S2, the basic data processing includes improving image sharpness, removing image noise, and standardizing image size;

[0050] In step S2, dividing the standard sample dataset into training, testing, and validation sets includes automatically filtering the standard sample dataset, removing overly blurry images, manually correcting them using labelimg software, and labeling them with tags; dividing the standard sample dataset into training, validation, and testing sets in a ratio of 8:1:1 to ensure the balance and representativeness of the image data distribution.

[0051] Step S3: Model Building: Construct the object detection model, such as... Figure 2 As shown, the data from the training set is input into the object detection model for training, and the parameters of the object detection model are optimized using the validation set and the test set.

[0052] In step S3, constructing the target detection model includes implementing Online Dynamic Image Augmentation (ODIA) technology, such as... Figure 3 As shown, the attributes of the training set images are dynamically adjusted; the ODIA strategy includes Dynamic Morphological Transformation Enhancement (DMDA) technique.

[0053] A coordinate-guided multi-granularity enhanced feature fusion module (CMGE-FFM) is introduced into the target detection model to enhance its ability to handle targets in complex environments.

[0054] A Dynamic Shape Similarity Intersection over Union (DSS-IoU) loss function, which incorporates shape similarity measurement and dynamic weight adjustment mechanism, is adopted to improve the performance of the object detection model.

[0055] In step S3, the online dynamic image enhancement technology includes Dynamic Morphological Transformation Enhancement (DMDA). DMDA is used to simulate the natural changes of highly dynamic targets such as flames. DMDA plays a key role in this invention, aiming to enhance the model's ability to detect highly dynamic targets such as flames. The Dynamic Morphological Transformation Enhancement technology includes the following steps:

[0056] 1. Feature extraction:

[0057] Image Import: First, load the images from the labeled standard sample dataset to prepare for extracting the flame image and its label file. Parse the label file to extract the flame's bounding box information, ensuring subsequent processing focuses on the flame itself.

[0058] Color space conversion: The HSV color space can better separate color and brightness information, highlighting the color characteristics of flames. Therefore, it is necessary to convert the flame area from the BGR color space to the HSV color space.

[0059] Image noise processing: further refine the flame shape and remove background noise, smooth the flame boundary, fill in the internal voids of the flame, and remove noise around the flame.

[0060] 2. Image fusion:

[0061] Image edge processing: Through erosion operations, the natural edges of flames are simulated, removing black outlines caused by small noises and brightness differences in the flame image. Multiple erosions are performed on the flame image using dynamics and sizes related to the flame size to simulate the irregularity of the flame edges. This not only removes noise but also makes the flame outline clearer, facilitating subsequent identification and analysis.

[0062] Image fusion processing: After resizing the flame image, it is overlaid onto the target background image. The target background image is converted to grayscale, black outlines, noise, and background are removed, and a threshold is applied to create a mask. The target area where the flame appears in the target background is calculated based on the actual boundaries of the flame, accurately simulating the appearance of the flame in the target background. Based on the color and brightness characteristics of the flame, the flame image is optimized through morphological operations to make the fusion of the flame and the target background more natural, improving the realism of flame recognition and simulation.

[0063] Label file update: To maintain data consistency and facilitate subsequent model training, after the flame is superimposed onto the target scene, the corresponding label file is updated based on the new position and size of the flame in the image. If the updated label file overlaps with the original label file, the two label files are automatically merged to ensure that the information in the label file accurately reflects the actual situation of the flame in the target scene, providing accurate training and validation data for the model.

[0064] 3. Image feature morphological changes:

[0065] Random point generation: To cope with the high dynamics of flames, a controllable number of random points are generated for each flame image.

[0066] Random vector change: Based on the maximum displacement parameter of the random point, calculate the movement of a random vector direction for each random point to obtain the target position of the random point, and simulate the natural dynamic shape change of the flame caused by wind direction change or other external factors.

[0067] Highly variable flame image generation: By utilizing generated random points and their corresponding target locations, the entire image is deformed using local affine transformation techniques. This method allows for flexible deformation of the image within local regions, thereby creating new flame shapes. In this way, not only can the natural dynamics of flames be simulated, but also highly variable flame images can be generated, which is particularly important for training deep learning models to recognize various flame scenes.

[0068] In step S3, the introduction of a coordinate-guided multi-granularity enhanced feature fusion module into the target detection model, such as... Figure 4 As shown, the process includes the following steps: First, in the neck structure of the object detection model, the input feature matrix is ​​divided into two parts: one part undergoes a CoordConv operation, and the other part undergoes a standard convolution operation. For the part undergoing the CoordConv operation, x and y coordinate channels are generated and convolved with the input feature map to enhance the spatial location information of the feature map. Then, a linear transformation is performed on the feature map after the CoordConv operation to generate query (Q) and key (K) matrices. For the part undergoing the standard convolution operation, the original features are extracted through the convolution operation and then linearly transformed to generate a value (V) matrix. After generating the Q, K, and V matrices, convolution operations are performed on them to further enhance feature extraction. A second dimensionality reduction and convolution operation is performed on the high-dimensional features to obtain mid-dimensional features. Finally, a third dimensionality reduction and convolution operation is performed on the mid-dimensional features to obtain low-dimensional features. Then, the three-dimensional features of the convolutional Q, K, and V matrices are input into a multi-head self-attention mechanism. By splitting the features into eight heads, attention weights and weighted values ​​are calculated separately to capture long-distance dependencies and global contextual information in the feature maps. A unified-dimensional convolution operation is then performed on these three-dimensional features to improve the effectiveness of feature fusion. Finally, maximum value fusion is performed to obtain the fused feature representation.

[0069] In step S3, the specific calculation formula for the dynamic shape similarity boundary intersection-union ratio, which introduces shape similarity measurement and dynamic weight adjustment mechanism, is as follows:

[0070] H(A,B)=max(h(A,B),h(B,A));

[0071]

[0072] ShapeSim(A,B)=exp(-δ*H(A,B));

[0073] Where H(A, B) is the Hausdorff distance between set A and set B; h(A, B) is the maximum distance from each point in set A to the nearest point in set B; h(B, A) is the maximum distance from each point in set B to the nearest point in set A; δ is the adjustment parameter;

[0074] By employing the DSS-IoU (Dynamic Shape Similarity Intersection over Union) loss function and introducing a shape similarity metric and dynamic weight adjustment mechanism, the accuracy and robustness of object detection are significantly improved. Building upon traditional loss functions, DSS-IoU introduces a shape similarity metric, measuring the maximum distance between two sets to effectively capture the differences between the bounding box and the target shape. This allows the object detection model to perform better when dealing with complex-shaped targets. Through this approach, the object detection model can better handle shape differences between the bounding box and the target, particularly significantly improving the detection performance of irregularly shaped targets.

[0075] Secondly, this invention designs a dynamic weight adjustment mechanism that adjusts the weight parameters in the loss function in real time based on the aspect ratio and area of ​​the target. The dynamic weight adjustment function, based on the target's aspect ratio and area, calculates dynamic weights in real time, enabling the target detection model to more flexibly handle targets of different shapes and sizes, thus improving its ability to detect targets in various complex scenes. The formula for the dynamic weight adjustment mechanism is as follows:

[0076]

[0077] Among them, w pred To predict the width of the bounding box; h pred To predict the height of the bounding box; w gt h is the width of the actual bounding box. gt A is the height of the actual bounding box. pred To predict the area of ​​the bounding box; A gt η is the area of ​​the true bounding box; η and ζ are adjustment parameters; r is the aspect ratio of the target.

[0078] The traditional loss function is calculated using the following formula:

[0079]

[0080] Where p is the distance between the center points of the predicted bounding box and the ground truth bounding box; c is the diagonal length of the smallest closed region enclosing the two bounding boxes; C w C is the length of the closed region of width. h b is the length of the closed region with height; b is the coordinate of the center point of the predicted bounding box; b gt is the coordinates of the center point of the ground truth bounding box; w is the width of the predicted bounding box; w gt h is the width of the ground truth bounding box; h is the height of the predicted bounding box; gt The height of the actual bounding box;

[0081] The expression for the dynamic shape similarity boundary intersection-union ratio loss function, which incorporates shape similarity measurement and dynamic weight adjustment mechanism, is as follows:

[0082] Loss=α*L EIoU +β*ShapeSim(A, B)+γ*DynWeight(r, A);

[0083] Where α is the weight coefficient of the traditional loss function; β is the weight coefficient of the shape similarity measure; and γ is the weight coefficient of the dynamic weight adjustment.

[0084] Building upon IoU, DSS-IoU not only considers the distance between bounding box center points and aspect ratio differences, but also enhances its ability to handle irregular shapes and targets of varying sizes compared to traditional loss functions. Furthermore, through a dynamic weight adjustment mechanism, it makes the object detection model more flexible and robust when facing diverse targets. Moreover, by introducing shape similarity measurement and dynamic weight adjustment, the DSS-IoU loss function can more accurately assess the difference between the target bounding box and the true target, thereby improving the detection accuracy and reliability of the object detection model.

[0085] Step S4: Safety Monitoring: Analyze real-time images within the substation using the trained target detection model and output the detection results.

[0086] The trained target detection model is used to perform real-time analysis and detection of images in the substation, identifying flames, personnel wearing safety helmets, and instrument status, and obtaining detection results.

[0087] Furthermore, the method of the present invention was tested in comparison with several other methods, and the results are shown in Table 1.

[0088] FasterR-CNN 0.81 YOLOv5s 0.847 0.889 0.614 CBAM-YOLO 0.856 0.884 0.599 YOLOv5-ECA 0.845 0.89 0.613 SE-YOLOv5s 0.854 0.89 0.61 YOLOv8s 0.853 0.885 0.614 This invention 0.868 0.901 0.619

[0089] Table 1

[0090] Extensive experiments were conducted to compare the performance of the object detection model proposed in this invention with several popular object detection models. The performance PR curve of the object detection model proposed in this invention is shown below. Figure 5 As shown ( Figure 5 In the example, instrument 2 has a PR value of 0.995 (coinciding with the bounding box). The results show that the target detection model proposed in this invention performs excellently across all key performance indicators, particularly achieving the highest scores of 0.901 and 0.619 in mAP and mAP 0.5–0.95, respectively. This performance improvement is attributed to the introduction of the ODIA data augmentation strategy and coordinate convolution, which significantly enhance the accuracy and generalization ability of the target detection model in identifying key targets in complex substation environments. Compared to other target detection models, the target detection model proposed in this invention exhibits a higher recall rate (0.868), indicating that it can detect more real targets more effectively and reduce missed detections. Furthermore, its performance in the mAP 0.5–0.95 range demonstrates its ability to maintain stable performance across various IoU thresholds, which is crucial for reliability in practical applications.

[0091] In summary, this invention acquires video and image data of substations under various operating conditions as a standard sample dataset; performs basic data processing on the standard sample dataset; constructs a target detection model, trains the target detection model, and optimizes the target detection model parameters; uses the trained target detection model to analyze real-time images within the substation and outputs detection results; this invention can monitor flames, personnel safety helmet wearing status, and instrument status in the substation environment in real time, thereby improving substation safety; by using the trained target detection model to perform real-time analysis and detection of images in the substation, it can quickly identify potential safety hazards, issue timely warnings, and reduce the risk of safety accidents.

Claims

1. A method for environmental safety monitoring of complex substations based on high dynamic target enhancement, characterized in that: Includes the following steps: Step S1: Data Acquisition: Acquire video and image data of the substation under various operating conditions as a standard sample dataset; Step S2: Data preprocessing: Perform basic data processing on the standard sample dataset, dividing the standard sample dataset into training set, test set and validation set; Step S3: Model Construction: Construct an object detection model, input the training set data into the object detection model for training, and optimize the object detection model parameters through validation and test sets; among them, implement an online dynamic image enhancement method to adjust the attributes of training images in the training set; introduce a coordinate-guided multi-granularity enhanced feature fusion module in the neck structure of the object detection model to enhance the model's ability to handle objects in complex environments; and adopt a dynamic shape similarity boundary intersection-union loss function that introduces shape similarity measurement and dynamic weight adjustment mechanism to improve the performance of the object detection model. Step S4: Safety Monitoring: Analyze real-time images within the substation using the trained target detection model and output the detection results; In step S3, the implementation of the online dynamic image enhancement method to adjust the attributes of training images in the training set includes the following steps: First, load the images of the labeled standard sample dataset, extract the flame images and their label files, and refine the flame morphology; then, superimpose the flame images onto the target background image, perform image fusion processing, and update the corresponding label files according to the new position and size of the flames in the fused image; finally, generate a controllable number of random points for each flame image, calculate the movement of a random vector direction for each random point based on the maximum displacement parameter of the random points, obtain the target position of the random point, and deform the fused image through local affine transformation to generate a flame image with high variability; the introduction of a coordinate-guided multi-granularity enhanced feature fusion module in the neck structure of the target detection model to enhance the model's ability to handle targets in complex environments includes the following steps: In the neck structure of the target detection model, divide the input feature matrix into two parts, perform a CoordConv operation on one part; generate x and y coordinate channels, The feature map is then convolved with the input feature map to enhance its spatial location information. A linear transformation is then applied to the CoordConv-operated feature map to generate query and key matrices. A standard convolution operation is performed on another part of the input feature matrix. The original features are extracted through convolution and linearly transformed to generate value matrices. After generating the query, key, and value matrices, convolution operations are performed on them to further enhance feature extraction. A second dimensionality reduction and convolution operation is performed on the high-dimensional features to obtain mid-dimensional features. Finally, a third dimensionality reduction and convolution operation is performed on the mid-dimensional features to obtain low-dimensional features. The three dimensions of the convolved query, key, and value matrices are input into a multi-head self-attention mechanism. Attention weights and weighted values ​​are calculated by dividing the matrix into eight heads to capture long-distance dependencies and global context information in the feature map. A unified-dimensional convolution operation is then performed on the three dimensions to improve the effectiveness of feature fusion. Finally, maximum value fusion is performed to obtain the fused feature representation.

2. The method for environmental safety monitoring of complex substations based on high dynamic target enhancement according to claim 1, characterized in that: In step S1, video and image data of the substation under various operating conditions are collected using the camera equipment in the substation and the existing database, which serve as a standard sample dataset.

3. The method for environmental safety monitoring of complex substations based on high dynamic target enhancement according to claim 1, characterized in that: In step S1, the standard sample dataset includes image data of flames, image data of personnel wearing safety helmets, and image data of instrument status.

4. The method for environmental safety monitoring of complex substations based on high dynamic target enhancement according to claim 1, characterized in that: In step S2, the basic data processing includes improving image sharpness, removing image noise, and standardizing image size.

5. The method for environmental safety monitoring of complex substations based on high dynamic target enhancement according to claim 1, characterized in that: In step S2, dividing the standard sample dataset into training, testing, and validation sets includes automatically filtering the standard sample dataset, removing overly blurry images, manually correcting them using labelimg software, and labeling them with tags; dividing the standard sample dataset into training, validation, and testing sets in a ratio of 8:1:1 to ensure the balance and representativeness of the image data distribution.

6. The method for environmental safety monitoring of complex substations based on high dynamic target enhancement according to claim 1, characterized in that: In step S3, the specific calculation formula for the dynamic shape similarity boundary intersection-union ratio, which introduces shape similarity measurement and dynamic weight adjustment mechanism, is as follows: in, The Hausdorff distance between set A and set B; Let A be the maximum distance from each point in set A to the nearest point in set B. Let B be the maximum distance from each point in set B to the nearest point in set A. To adjust the parameters; The weight parameters in the loss function are adjusted in real time based on the aspect ratio and area of ​​the target; the formula for the dynamic weight adjustment mechanism is as follows: in, To predict the width of the bounding box; To predict the height of the bounding box; The width of the actual bounding box; The height of the actual bounding box; To predict the area of ​​the bounding box; This represents the area of ​​the actual bounding box; and These are the adjustment parameters; The aspect ratio of the target; The traditional loss function is calculated using the following formula: in, To predict the distance between the center points of the bounding box and the ground truth bounding box; The length of the diagonal of the smallest closed region that encloses the two bounding boxes; The length of the closed region is its width; The length of the closed region is its height. To predict the coordinates of the center point of the bounding box; These are the coordinates of the center point of the actual bounding box; To predict the width of the bounding box; The width of the actual bounding box; To predict the height of the bounding box; The height of the true bounding box; the expression for the dynamic shape similarity boundary intersection-union ratio loss function, which introduces shape similarity measurement and dynamic weight adjustment mechanism in IOU, is as follows: in, These are the weighting coefficients of the traditional loss function; Weighting coefficients for shape similarity measurement; The weighting coefficients are dynamically adjusted.

Images