Method and system for identifying safe power transformation operation scene based on lightweight Yolov7

By improving the lightweight Yolov7 model, combined with optical flow changes and attention mechanisms, the problems of high computing resources and insufficient real-time performance in substation safety supervision are solved, efficient and accurate safety operation scene recognition and early warning are achieved, the scene adaptability is enhanced, and detailed safety management data support is provided.

CN120833579APending Publication Date: 2025-10-24YONGCHUAN POWER SUPPLY BRANCH STATE GRID CHONGQING ELECTRIC POWER COMPANY
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202510907732.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-02
Publication Date
2025-10-24

AI Technical Summary

Technical Problem

Existing technologies for substation safety supervision have problems such as high computing resource requirements, insufficient real-time performance, and poor scenario adaptability. Traditional deep learning algorithms such as Faster R-CNN have complex model structures and large computational workloads, which cannot meet the needs of real-time safety supervision. Single-frame image algorithms cannot recognize complex behaviors, and video analysis algorithms cannot adapt to complex and changeable substation operation scenarios.

Method used

A lightweight Yolov7 model is used to extract key frames by analyzing optical flow changes, construct a labeled sample set, and improve the model through random weighted features and residual connections. Combined with the void convolution and CBAM channel spatial attention mechanism, the model's detection ability for multi-scale targets is enhanced, a safety and supervision recognition model is constructed, and the safety status of operators is judged using logical decision functions.

Benefits of technology

The algorithm is lightweight and efficient, accurately identifies key safety indicators, and has strong scenario adaptability. It can detect the location information of workers and safety indicators in real time, generate safety warning information, provide objective and reliable data support, and improve the efficiency and accuracy of safety supervision.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120833579A_ABST
    Figure CN120833579A_ABST
Patent Text Reader

Abstract

The invention discloses an identification method and system for a safe power transformation operation scene based on lightweight Yolov7, and relates to the technical field of image target detection, and the method comprises the steps: based on power transformation operation scene data, extracting a key frame through the analysis of the optical flow change, constructing a labeling sample set, training an improved Yolov7 model through the labeling sample set, and obtaining a Yolov7 model; a safety supervision identification model is obtained; carrying out target detection on the real-time power transformation operation scene by using a safety supervision identification model, obtaining position information of an operator and a safety index, analyzing the position information through a preset logic judgment function, and judging a safe operation state of the operator; and generating safety early warning information according to the judgment result of the safety operation state, and displaying the safety early warning information through a human-computer interaction interface. Aiming at the defect of target identification of a Yolov7 basic model in a power transformation operation scene, the detection capability of the model on targets with different sizes can be improved, and the adaptability to a complex operation scene is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image target detection, in particular to a safety substation operation scene recognition method and system based on a light-weight Yolov7. BACKGROUND

[0002] As a key component of the power system, the substation involves complex electrical equipment operation, high-voltage environment, and numerous personnel activities. Once a safety accident occurs, it will not only threaten the life safety of the operating personnel, but also cause large-scale power outages, affect the normal production and living order of society, and bring huge economic losses to power companies. Traditional substation operation safety supervision mainly relies on manual supervision. Supervisors discover illegal operations and safety hazards by on-site patrol, monitoring video viewing, and other methods. However, manual supervision has many drawbacks: it cannot achieve full-time, full-range monitoring, and is prone to supervision loopholes; long-term supervision work is easy to cause the supervisor to be tired, reducing the supervision efficiency and accuracy; manual judgment has subjectivity, and different supervisors may have different understanding and execution scales of safety standards. Therefore, substation operation safety supervision needs technological innovation to improve supervision efficiency and accuracy.

[0003] With the development of technology, in recent years, the comprehensive exploration and development of artificial intelligence in different business scenarios in the power field have formed a large number of new achievements represented by artificial intelligence and power Internet of Things and have been widely applied. In the field of safety supervision and identification, systems based on deep learning target detection algorithms have gradually been applied, such as personnel behavior detection and equipment state identification using traditional target detection algorithms such as Faster R-CNN. At present, domestic power companies have carried out video monitoring, robot inspection, and other businesses, and substation visualization devices have been basically popularized and equipped in large quantities. Management units arrange personnel with construction safety supervision professional knowledge to track construction details and carry out on-site operation inspection in the background by manually operating on-site cameras to judge whether there are safety hazards in the operation. Current operation site monitoring and analysis algorithms are mainly divided into two categories based on single-frame images and video sequences, which to some extent assist the safety management of substations.

[0004] However, the prior art has obvious deficiencies in practical application. First, the traditional deep learning target detection algorithm such as the Faster R-CNN model structure is complex, and the calculation amount is large, and when the substation is deployed, high-performance hardware equipment support is required, such as high bandwidth, high GPU, etc., which increases the construction cost and operation cost of the system. Secondly, with the deepening of the frequency and application field of monitoring applications, the massive data set transmission problem has been faced, in order to ensure the quality of the shooting video, high-definition cameras are mostly used, which brings the contradiction between high-resolution video transmission and limited bandwidth, and this problem has not been effectively solved at present. Thirdly, the processing speed of the complex deep learning algorithm is slow, especially when processing high-resolution videos and multi-target scenes, delays are prone to occur, which cannot meet the real-time safety supervision requirements of complex and variable substation operation scenes, and the adaptability is poor. Fourthly, the existing single-frame image algorithm cannot identify complex human behaviors and dynamically understand construction scenes, and can only solve simple problems such as wearing safety helmets and work clothes; and the video analysis algorithm needs to consume a large amount of calculation and storage resources, and cannot adapt to the complex environment of the infrastructure management and control site and the edge computing device (such as the edge agent device), which restricts its actual application. In addition, the one-key control scene is affected by objective factors such as dark light and low resolution of the shooting device, resulting in poor recognition effect. With the increase of the number of monitoring sites, it is difficult to comprehensively monitor the power site, and there are hidden dangers that the site safety cannot be found and rectified in time.

[0005] At present, no effective solution has been proposed for the problems in the related art. SUMMARY

[0006] In view of the problems in the related art, the present application proposes a safety substation operation scene recognition method and system based on a lightweight Yolov7, which has the advantages of lightweight and efficient algorithm, accurate recognition of key safety indicators, and strong scene adaptability, thereby solving the problems of high calculation resource demand, insufficient real-time performance and poor scene adaptability in the prior art.

[0007] To this end, the specific technical solutions adopted by the present application are as follows:

[0008] According to one aspect of the present application, a safety substation operation scene recognition method based on a lightweight Yolov7 is provided, which comprises:

[0009] S1, based on the substation operation scene data, the key frames are extracted by analyzing the light flow change to construct the labeled sample set, and the improved Yolov7 model is trained through the labeled sample set to obtain an safety supervision recognition model;

[0010] S2, the safety supervision recognition model is used for target detection on the real-time substation operation scene, position information of the operation personnel and safety indexes is acquired, and the position information is analyzed through a preset logical judgment function to judge the safety operation state of the operation personnel;

[0011] S3, according to the judgment result of the safety operation state, safety warning information is generated, and is displayed through a man-machine interaction interface.

[0012] Further, based on the substation operation scene data, key frames are extracted by analyzing the optical flow change to construct a labeled sample set, and the improved Yolov7 model is trained through the labeled sample set to obtain the safety supervision recognition model, including:

[0013] S11, based on the personnel behavior video data of the substation operation scene, key frames are extracted according to the analysis of the optical flow change to construct a key frame data set;

[0014] S12, the key frame data set is semantically labeled and hierarchically divided according to the safety semantic category to form a labeled sample set;

[0015] S13, the labeled sample set is used to train a lightweight Yolov7 model improved by random weighted features and residual connection to obtain a safety supervision recognition model for the substation operation scene;

[0016] The lightweight Yolov7 model includes an input module, a backbone network, a detection head and an output module, and the detection head includes a spatial pyramid pooling module.

[0017] Further, based on the personnel behavior video data of the substation operation scene, key frames are extracted according to the analysis of the optical flow change to construct a key frame data set, including:

[0018] S111, using a video acquisition device covering the substation operation area, personnel behavior videos in different time periods and different operation stages are collected and numbered to obtain an original video sequence;

[0019] S112, based on the original video sequence, the optical flow change between adjacent frames is compared through motion vector analysis, the current frame exceeding the threshold is extracted as a key frame, and an initial key frame set is generated;

[0020] S113, the initial key frame set is screened, and the image parameters are optimized to obtain a key frame data set.

[0021] Further, the extracting the current frame exceeding the threshold as the key frame by analyzing and comparing the optical flow changes between adjacent frames through the motion vector includes: calculating motion components of all pixel points in the video frame in horizontal and vertical directions to obtain pixel displacement values; performing square sum and square root operations on the displacement values of all pixel points to obtain pixel motion amplitudes; calculating an average value of all pixel motion amplitudes and comparing the average value with a preset threshold; and determining the current frame as the key frame when the average motion amplitude is greater than the preset threshold; wherein a first preset threshold is set for a video of a static scene, and a second preset threshold is set for a video of a dynamic scene, and the second preset threshold is greater than the first preset threshold.

[0022] Further, the training of the improved lightweight Yolov7 model using random weighted features and residual connection using the labeled sample set to obtain the safety supervision recognition model for the substation operation scene includes:

[0023] S131, in the feature extraction stage of the lightweight Yolov7 model, based on the spatial pyramid pooling module, a plurality of parallel atrous convolution layers are used to replace the maximum pooling layer, the input feature map is convolved through atrous convolution operations with different atrous rates, and the obtained multi-scale feature maps are spliced in the channel dimension to obtain multi-scale fusion features of the labeled sample set;

[0024] S132, in the feature enhancement stage of the lightweight Yolov7 model, the channel spatial attention mechanism is applied to the multi-scale fusion features of the labeled sample set, and the attention weighted features are obtained by using the channel attention module and the spatial attention module;

[0025] S133, in the feature fusion stage of the lightweight Yolov7 model, the attention weighted features are connected with the input feature map through residual connection to generate output feature maps, so as to complete the model training and obtain the safety supervision recognition model for the substation operation scene.

[0026] Further, the replacing the maximum pooling layer with a plurality of parallel atrous convolution layers and convolving the input feature map through atrous convolution operations with different atrous rates includes: setting a plurality of parallel atrous convolution layers, each atrous convolution layer uses a convolution kernel with the same size and the same number of output channels, and is configured with different atrous rates; the input feature map is convolved by each atrous convolution layer, respectively, under the premise of keeping the consistent receptive field, small target detail features, medium scale features and large scene outline features are captured respectively; and the output feature maps of each atrous convolution layer are spliced in the channel dimension to form a multi-scale fusion feature map.

[0027] Further, the channel attention module is used to identify key feature channels related to safety equipment and assign weights, and the spatial attention module is used to locate the work risk area and strengthen the spatial feature weight; the channel spatial attention mechanism is applied to the multi-scale fusion features of the labeled sample set, and the attention weighted features are obtained by using the channel attention module and the spatial attention module: based on the channel attention module, global average pooling and global maximum pooling are performed on the multi-scale fusion feature map, the pooling results are input into the shared multi-layer perception to fuse the features, and the channel attention weight is obtained after the activation function processing, and the channel attention weighted feature map is obtained by multiplying the multi-scale fusion feature map; according to the spatial attention module, the channel attention weighted feature map is subjected to average pooling and maximum pooling in the channel dimension, the pooling results are spliced, and then subjected to convolution dimension reduction processing and activation function activation to obtain the spatial attention weight matrix, and the spatial attention weighted feature map is obtained by multiplying the channel attention weighted feature map.

[0028] Further, the safety supervision recognition model is used to detect the target of the real-time substation work scene, obtain the position information of the work personnel and safety indicators, and analyze the position information by using a preset logical judgment function to determine the safety work state of the work personnel, including:

[0029] S21, according to the identified substation work scene type, determine the corresponding key safety indicator set, wherein the key safety indicator set includes safety elements determined according to different work scenes;

[0030] S22, using the safety supervision recognition model to detect the target of the real-time work scene, obtaining the position information of the work personnel bounding box and the key safety indicator bounding box;

[0031] S23, according to the position relationship between the work personnel bounding box and the key safety indicator bounding box, judging the relevance of the key safety indicator and the work personnel, and using the logical judgment function to determine whether it meets the safety work requirements.

[0032] Further, the correlation between the key safety indicator and the worker is determined according to the positional relationship between the worker annotation box and the key safety indicator annotation box, including: setting the lower left corner of the image as the coordinate origin to establish a two-dimensional coordinate system, and obtaining the upper left corner coordinates and the lower right corner coordinates of the worker annotation box to determine the position range of the worker annotation box; obtaining the upper left corner coordinates and the lower right corner coordinates of the key safety indicator annotation box to determine the position range of the key safety indicator annotation box; determining the positional relationship between the key safety indicator annotation box and the worker annotation box, if the key safety indicator annotation box is completely located inside the worker annotation box, it is determined that the key safety indicator is related to the worker; if the key safety indicator annotation box is completely located outside the worker annotation box without overlapping, it is determined that the key safety indicator is irrelevant to the worker; if the key safety indicator annotation box and the worker annotation box partially overlap, the correlation between the key safety indicator and the worker is determined based on the comparison result of the minimum distance between the edges of the two and a preset threshold, wherein the preset threshold is determined by statistics of the positional relationship of the standard wearing sample.

[0033] According to another aspect of the present application, a safety substation operation scene recognition system based on light-weight Yolov7 is also provided, which comprises:

[0034] A model construction module is configured to construct a labeled sample set by analyzing optical flow changes to extract key frames based on substation operation scene data, and train an improved Yolov7 model through the labeled sample set to obtain an safety supervision recognition model;

[0035] A scene recognition module is configured to use the safety supervision recognition model to perform target detection on a real-time substation operation scene to obtain position information of workers and safety indicators, and analyze the position information through a preset logical judgment function to determine the safety operation state of the workers;

[0036] A warning display module is configured to generate safety warning information according to the determination result of the safety operation state, and display the information through a human-computer interaction interface.

[0037] The present application has the following advantages:

[0038] (1) The present application installs specific cameras at appropriate positions in the substation to comprehensively cover the operation area, and extracts key frames based on optical flow change analysis to effectively reduce data redundancy while retaining key information, and reasonably divides high-quality training sample sets by combining a hierarchical division method according to semantic categories, thereby providing a rich and balanced data basis for model training.

[0039] (2) The present application aims at the defects of the basic Yolov7 in the substation operation scene, such as small target missing detection and inaccurate identification of occluded targets, and innovatively uses a hollow convolution to replace a maximum pooling layer to expand the receptive field, combines a CBAM channel spatial attention mechanism to dynamically weight and enhance key features, and retains original feature information through residual connection, thereby significantly improving the detection capability of the model for multi-scale targets in the substation scene and enhancing the adaptability of the model to complex operation scenes.

[0040] (3) The present application constructs a logical judgment function based on a safety index set, and combines the position of the labeled frame to accurately distinguish the relevance of key safety indicators and operation personnel, judge whether the personnel behavior in the operation scene conforms to the safety operation specification, distinguish the key indicators related and irrelevant, and discriminate and record the operation situation of the scene, thereby providing detailed operation situation data for the management personnel and providing objective and reliable data support for safety management.

[0041] (4) The present application displays the monitoring results of the operation site in real time through an intuitive man-machine interface, uses different color labeled frames to clearly distinguish compliant and non-compliant behaviors, integrates statistical charts to display safety data distribution, and is deeply integrated with a safety management system to realize data sharing, facilitate the query and analysis of the behavior mode and safety risk distribution of the operation personnel, and provide a complete solution for safety management decision and model continuous optimization. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0043] Figure 1 is a flowchart of a safety substation operation scene recognition method based on a lightweight Yolov7 according to an embodiment of the present application;

[0044] Figure 2 is a safety supervision recognition logic flowchart in a safety substation operation scene recognition method based on a lightweight Yolov7 according to an embodiment of the present application;

[0045] Figure 3 is a safety supervision recognition model training flowchart in a safety substation operation scene recognition method based on a lightweight Yolov7 according to an embodiment of the present application;

[0046] Figure 4It is a schematic diagram of an improved random weighted feature module in a safety substation operation scene recognition method based on a lightweight Yolov7 according to an embodiment of the present application.

[0047] Figure 5 It is a principle block diagram of a safety substation operation scene recognition system based on a lightweight Yolov7 according to an embodiment of the present application. DETAILED DESCRIPTION

[0048] To further illustrate the embodiments, the present application provides drawings which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can explain the operating principle of the embodiments in conjunction with the related description of the specification. Those skilled in the art should understand other possible implementations and advantages of the present application by referring to these contents. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0049] According to an embodiment of the present application, a safety substation operation scene recognition method and system based on a lightweight Yolov7 are provided.

[0050] The present application will be further described in conjunction with the drawings and specific embodiments. Figure 1 As shown, according to an embodiment of the present application, a safety substation operation scene recognition method based on a lightweight Yolov7 is provided, which includes:

[0051] S1, based on substation operation scene data, key frames are extracted by analyzing optical flow changes to construct a labeled sample set, and an improved Yolov7 model is trained through the labeled sample set to obtain an safety supervision recognition model;

[0052] S2, using the safety supervision recognition model to detect targets in real-time substation operation scene, obtaining the position information of the operation personnel and safety indicators, and analyzing the position information through a preset logical judgment function to determine the safety operation state of the operation personnel;

[0053] S3, generating safety warning information according to the determination result of the safety operation state, and displaying through a man-machine interaction interface.

[0054] Specifically, the present application includes the following steps:

[0055] Step 1: Collect video data related to personnel behavior in substation operation scene, extract key frames therein as sample set of detection model, and label and divide the sample set in combination with artificial labeling method;

[0056] Step 2: Based on the Yolov7 model, the working scene is identified and judged, wherein a new design of random weighted features and residual connection is used to improve the basic Yolov7 network structure;

[0057] Step 3: According to the position information of the working personnel and the related safety indicators, a logical decision function is used to judge the safety of the personnel in the model identified scene;

[0058] Step 4: According to the model judgment result, a warning is given to realize the semi-automatic application of man-machine interaction and working scene recognition.

[0059] In one embodiment, based on the substation working scene data, the key frames are extracted by analyzing the optical flow changes to construct the labeled sample set, and the improved Yolov7 model is trained through the labeled sample set to obtain the safety supervision recognition model, which includes:

[0060] S11, based on the personnel behavior video data of the substation working scene, the key frames are extracted according to the analysis of the optical flow changes to construct the key frame data set;

[0061] S12, according to the safety semantic category, the key frame data set is semantically labeled and hierarchically divided to form a labeled sample set;

[0062] S13, using the labeled sample set to train the lightweight Yolov7 model improved by random weighted features and residual connection, to obtain the safety supervision recognition model for the substation working scene;

[0063] Among them, the lightweight Yolov7 model includes an input module, a backbone network, a detection head and an output module, and the detection head includes a spatial pyramid pooling module.

[0064] In one embodiment, based on the personnel behavior video data of the substation working scene, the key frames are extracted according to the analysis of the optical flow changes to construct the key frame data set, which includes:

[0065] S111, using the video acquisition device covering the substation working area, personnel behavior videos in different time periods and different working stages are collected and numbered to obtain the original video sequence;

[0066] Specifically, in step 1, the camera with wide-angle lens and good low-light performance is installed in the appropriate position of the substation to comprehensively cover the working area. In different time periods and different working stages, according to the characteristics of substation work and the changes of working environment, data collection and numbering are carried out respectively to diversely capture the personnel behavior in the substation working scene, and at the same time, it is convenient for subsequent retrieval and processing.

[0067] S112, based on the original video sequence, comparing the optical flow changes between adjacent frames by motion vector analysis, extracting the current frame exceeding the threshold as the key frame, and generating an initial key frame set;

[0068] Specifically, the key frame is extracted by comparing the optical flow changes between adjacent frames by motion vector analysis, including:

[0069] Calculate the motion components of all pixel points in the video frame in the horizontal and vertical directions to obtain the pixel displacement value;

[0070] Square and square root operations are performed on the displacement values of all pixel points to obtain the pixel motion amplitude;

[0071] Calculate the average value of the motion amplitude of all pixels and compare it with the preset threshold;

[0072] When the average motion amplitude is greater than the preset threshold, the current frame is determined as a key frame;

[0073] Among them, the first preset threshold is set for the video of the static scene, and the second preset threshold is set for the video of the dynamic scene, and the second preset threshold is greater than the first preset threshold.

[0074] Specifically, the core goal of key frame extraction is to filter out frames that can represent content changes from continuous video streams, reduce data redundancy while retaining key information, and adopt pixel-based motion vector analysis to judge the amplitude of scene changes by calculating the displacement (optical flow field) of pixels in adjacent frames:

[0075]

[0076] In the formula, KeyF is the key frame, N is the total number of pixels, u and v are the motion components of the pixel in the x and y directions, T h is the threshold value, which is adjusted according to the video content (small value for static scene, large value for dynamic scene).

[0077] S113, filtering the initial key frame set and optimizing the image parameters to obtain a key frame data set.

[0078] Specifically, when constructing the video frame data, the key frame is extracted according to the degree of change of the video content, the image similarity algorithm is used to compare the differences between adjacent frames, and when the difference exceeds a certain threshold, the current frame is defined as a key frame (selected according to the optical flow change). The extracted key frame is filtered and optimized, including: removing blurred, repeated or irrelevant frames; adjusting brightness, contrast, color balance, etc. to improve the quality and readability of the image.

[0079] In one embodiment, the key frame data set is semantically labeled and hierarchically divided according to the safety semantic category to form a labeled sample set.

[0080] The samples consisting of key frames are layered according to categories (layered according to business scenarios, and the specific number of layers is determined according to the sample data categories), such as different sample categories of wearing protective equipment such as safety helmets and checking the scene of personnel operation specifications.

[0081] According to business scenarios (such as: safety helmet wearing, insulating glove use, high-altitude operation specification, etc.), the key frame samples are divided according to semantic categories. When making the training set, 80% of the sample set is taken as the training set, including 80% of the safety helmet wearing category, 80% of the insulating glove operation, etc. The above method is used to divide the sample categories when making the verification set and the test set.

[0082] The proportion of each category in each subset (training / verification / test) ≈ the proportion of the category in the total sample set, for example: if the safety helmet category accounts for 30% in the total sample, then the safety helmet category in the training / verification / test set ≈ 30%.

[0083] In one embodiment, the labeled sample set is trained using a lightweight Yolov7 model improved by random weighted features and residual connection to obtain an occupational safety and health identification model for substation operation scenarios, which includes:

[0084] S131, in the feature extraction stage of the lightweight Yolov7 model, based on the spatial pyramid pooling module, a plurality of parallel hole convolution layers are used to replace the maximum pooling layer, the input feature map is convolved by hole convolution operation with different hole rates, and the obtained multi-scale feature map is spliced in the channel dimension to obtain the multi-scale fusion features of the labeled sample set;

[0085] Specifically, the plurality of parallel hole convolution layers are used to replace the maximum pooling layer, and the input feature map is convolved by hole convolution operation with different hole rates, which includes:

[0086] A plurality of parallel hole convolution layers are set, each hole convolution layer uses a convolution kernel of the same size and the same number of output channels, and is configured with different hole rates;

[0087] The input feature map is convolved by each hole convolution layer, respectively capturing small target detail features, medium scale features, and large scene outline features while keeping the receptive field consistent;

[0088] The output feature maps of each hole convolution layer are spliced in the channel dimension to form a multi-scale fusion feature map.

[0089] In the lightweight Yolov7 model feature enhancement stage, a channel and space attention mechanism is applied to the multi-scale fusion features of the labeled sample set, and attention weighted features are obtained by using the channel attention module and the space attention module.

[0090] Specifically, the channel attention module is used to identify key feature channels related to safety equipment and assign weights, and the space attention module is used to locate the work risk area and strengthen the spatial feature weight.

[0091] Specifically, the channel and space attention mechanism is applied to the multi-scale fusion features of the labeled sample set, and attention weighted features are obtained by using the channel attention module and the space attention module, including:

[0092] Based on the channel attention module, global average pooling and global maximum pooling are performed on the multi-scale fusion feature map, the pooling results are input into a shared multi-layer perceptron for feature fusion, and the channel attention weight is obtained after activation function processing, and then multiplied by the multi-scale fusion feature map to obtain the channel attention weighted feature map.

[0093] According to the space attention module, the channel attention weighted feature map is subjected to average pooling and maximum pooling in the channel dimension, the pooled results are spliced, and then subjected to convolution dimension reduction processing and activation function activation to obtain a space attention weight matrix, and then multiplied by the channel attention weighted feature map to obtain a space attention weighted feature map.

[0094] In the lightweight Yolov7 model feature fusion stage, the attention weighted features are connected in residual connection with the input feature map to generate an output feature map, so as to complete the model training and obtain an occupational safety and health identification model for the substation work scene.

[0095] Specifically, in step 2, since the existing Yolov7 model has problems such as small target missing detection and inaccurate detection of occluded targets in target identification in the substation work scene, the present application adopts a random weighted feature and residual connection method to improve the above defects.

[0096] Specifically, the lightweight Yolov7 model mainly consists of an input module, a backbone network, a detection head, and an output module. The training sample set planned in step ① is used as the input of the model. The backbone network performs feature extraction on the image through 4 layers of basic convolution layers. Then, the ELAN module composed of multiple CBS is used to further extract and fuse feature information. Finally, the MP layer changes the number of channels to generate feature maps of different scales to capture different feature information in the image. For the 32 times down-sampling feature map output by the backbone network, after processing by the SPPCSPC module (spatial pyramid pooling module) of the detection head, the feature map is up-sampled to increase its scale, and the corresponding feature map of the backbone network is convolved and combined. By fusing feature information of different scales, the detection ability of the model for objects of different sizes is improved. Finally, the three ELAN-A modules in the detection head (Head) output feature expressions of different scales of 80x80, 40x40, and 20x20, respectively. After two convolutions, the feature dimension is converted to 255.

[0097] Specifically, the root cause of the low detection ability of small targets mainly lies in the small target feature extraction ability and the loss of small target feature information in deep sampling. Therefore, the network structure is improved by using random weighted features and residual connections in the model backbone network. First, the maximum pooling layer in SPPCSPC is replaced by a dilated convolution, and the CBAM attention mechanism (channel and spatial attention mechanism) is used to randomly assign weights to the channels after feature fusion. Finally, the residual structure is used to compensate for the information loss caused by convolution operations.

[0098] Specifically, as shown in Figure 4 , a dilated convolution with a dilation rate of 2, 4, and 6, a convolution kernel of 3, and an output channel number of 3 is used to convolve the input feature map in SPPCSPC, ensuring that the receptive field before and after convolution is consistent, improving the feature extraction ability while reducing the parameter amount.

[0099] Specifically, for the substation operation scene and its safety operation indicators of different sizes, the importance of feature extraction in different channels also varies, but in the Yolov7 model, only the feature results of different branches are simply spliced, and the importance of special information is not distinguished. Therefore, the CBAM attention mechanism (channel and spatial attention mechanism) is adopted in the present application to randomly assign weights to the feature maps after the hole convolution, so as to highlight the spatial region features considered more important by the model and suppress relatively unimportant feature information. In the channel dimension, the channel weight is learned based on the average pooling and the fully connected layer, and the important feature information is focused on. In the spatial dimension, the spatial position weight is learned based on the channel splicing and the convolution operation, and the position of the feature information is focused on. Finally, by multiplying the feature value of each spatial position of the original feature map by the corresponding spatial attention weight, the original feature map in the spatial dimension is reweighted according to the spatial attention weight:

[0100] F spatial-weighted (i,j)=(F in (i,j,c)×F channel-atten (c))×F spatial-atten ;

[0101] In the formula, F spatial-weighted is the feature map after spatial dimension attention weighting, F in is the input feature map, i, j and c are the indexes of height, width and channel three dimensions, F channel-atten is the channel attention weight vector, F spatial-atten is the spatial attention weight (here, random does not mean random number, but means the weight learned through data-driven method, that is, attention weight; these weights are generated adaptively according to the content of the input feature map, so they are random, but in fact they are adaptive). The channel and spatial attention mechanism in the present application learns the importance of channels and spaces through two modules of channel attention and spatial attention, and then weights the feature map. Among them, the channel attention module is used to identify the key feature channels related to safety equipment and give higher weight to solve the problem of safety equipment features being submerged in the substation scene; the spatial attention module is used to locate the operation risk area and strengthen the spatial feature weight to solve the problem of complex background interference in the substation.

[0102] It should be noted that the present invention uses the channel-space dual attention mechanism of CBAM to achieve dynamic weighted enhancement of "key feature channels" (such as protective equipment, semantic features of operating actions) and "key spatial areas" (such as operators, equipment operating parts) in substation operation scenarios, solving the shortcomings of feature fusion of the original Yolov7 model, making the model more adaptable to the complex scenes of multi-scale and multi-feature differences in substation operations, and ultimately improving the accuracy and efficiency of safe operation identification. The core of this mechanism is to "let the model autonomously learn which features / areas are more important for safe operations and enhance them in a targeted manner", rather than simply splicing features. It is a deep optimization of traditional target detection models in specific scenarios.

[0103] Specifically, first, three parallel dilated convolutional layers (with a convolution kernel size of 3×3, dilation rates of 2, 4, and 6, and an output channel number of 3) are used to perform convolution operations on the input feature map. Without increasing the amount of computation, the receptive field is expanded to capture small target details (such as reflective strips on helmets), medium-scale features (such as personnel postures), and large scene outlines (such as equipment layout) in the substation scene, and three feature maps F1, F2, and F3 (all with dimensions of H×W×3) are obtained. These three feature maps are connected in the channel dimension to obtain the multi-scale fusion feature F cat (Dimensions are H×W×9).

[0104] F i =DilatedConv(F in ,r i ),r i ∈{2,4,6};

[0105]

[0106] Where, F in is the input feature map, DilatedConv represents the hole convolution operation, r i is the void ratio, F cat It is a multi-scale fusion feature, and Concat is a feature splicing operation.

[0107] Specifically, the present invention applies the channel attention mechanism to the connected multi-scale fusion features, dynamically learns and enhances channels that are strongly related to safety detection (such as the color channel of the helmet and the edge channel of the equipment operation), and suppresses irrelevant background channels (such as the sky and wall texture). The channel attention module first performs global average pooling and global maximum pooling on the input feature map to extract channel-level statistical information; then the pooling result is input into the shared MLP (shared multi-layer perceptron) for feature fusion MLP (avg) + MLP (max); after Sigmoid activation, the channel weight (1×1×C, where C is the number of channels) is obtained, and the channel attention weight M is used as the input feature map. channel-atten With the feature map Fcat Multiply (i.e., F cat Each channel of is multiplied by the corresponding weight value), and the channel attention weighted feature map F is obtained. channel (size is H×W×9), the channel attention weighted feature map is input to the spatial attention module:

[0108] M channel-atten =σ(MLP(AvgPool(F cat ))+MLP(MaxPool(F cat )));

[0109]

[0110] In the formula, σ is the Sigmoid function (activation function), MLP is the multi-layer perceptron, AvgPool is the average pooling operation, MaxPool is the maximum pooling operation, is channel-by-channel multiplication, M channel-atten is the attention weight, F channel It is the feature map after channel attention processing.

[0111] Specifically, the input of the spatial attention module is F channel (H×W×9), first, for F channel Perform average pooling and maximum pooling in the channel dimension to extract spatial statistical information. Then, concatenate the pooling results, perform 7×7 separation convolution and Sigmoid activation, and obtain the spatial weight matrix M. spatial-atten (H×W×1), the spatial attention weight matrix M spatial-atten With the input feature map F channel Multiply [i.e., F channel Each spatial position (h, w) is multiplied by the corresponding weight value M spatial-atten (h,w)], and obtain the spatial attention weighted feature map F spatial-weighted (H×W×9):

[0112] M spatial-atten =σ(f 7×7 (Concat[AvgPool c (F channel ); MaxPool c (F channel )]));

[0113]

[0114] Where σ is the sigmoid function, f 7×7 is a 7×7 convolution dimensionality reduction process, Concat is a splicing operation, AvgPool cMaxPool is the average pooling operation in the channel dimension, c MaxPool is the maximum pooling operation in the channel dimension, M is the channel-wise multiplication, spatial-atten F is the attention weight, spatial-weighted is the feature map after spatial attention processing.

[0115] Specifically, finally, the attention weighted feature (spatial attention weighted feature map) after assignment is connected with the original input feature map in residual, which promotes feature reuse and information transmission, and further improves the adaptability of the model to different data:

[0116] F out =F in +F spatial-weighted ;

[0117] In the formula, F out is the feature map output after residual processing; + is the element-wise addition of the matched feature map, which directly adds the feature values of each channel and each spatial position to fuse the original and enhanced features.

[0118] Specifically, according to the improved network structure, the training parameters are adjusted, and in the training process, the changes of the training loss, accuracy, recall rate and other indicators of the model are paid attention to. The validation set is used for evaluation to understand the performance of the improved Yolov7 model. The test set is used to comprehensively evaluate the improved model, including accuracy, recall rate, F1 value, detection speed, etc.

[0119] In one embodiment, the safety supervision recognition model is used to detect the target in the real-time substation operation scene, obtain the position information of the operation personnel and the safety indicators, and analyze the position information through a preset logical judgment function to determine the safety operation state of the operation personnel, including:

[0120] S21, according to the recognized substation operation scene type, determine the corresponding key safety indicator set, wherein the key safety indicator set includes safety elements determined according to different operation scenes;

[0121] S22, using the safety supervision recognition model to detect the target in the real-time operation scene, obtaining the position information of the operation personnel bounding box and the key safety indicator bounding box;

[0122] S23, according to the position relationship between the operation personnel bounding box and the key safety indicator bounding box, judging the relevance of the key safety indicator and the operation personnel, and using the logical judgment function to determine whether it meets the safety operation requirements.

[0123] In one embodiment, according to the position relationship between the operation personnel bounding box and the key safety indicator bounding box, the relevance of the key safety indicator and the operation personnel includes:

[0124] Setting the lower left corner of the image as the coordinate origin, a two-dimensional coordinate system is established, and the left upper corner coordinate and the right lower corner coordinate of the worker marking box are obtained to determine the position range of the worker marking box;

[0125] The left upper corner coordinate and the right lower corner coordinate of the key safety indicator marking box are obtained to determine the position range of the key safety indicator marking box;

[0126] The positional relationship between the key safety indicator marking box and the worker marking box is judged, if the key safety indicator marking box is completely located inside the worker marking box, it is determined that the key safety indicator is related to the worker;

[0127] If the key safety indicator marking box is completely located outside the worker marking box without overlapping, it is determined that the key safety indicator is not related to the worker;

[0128] If the key safety indicator marking box and the worker marking box exist partial overlap, the relevance of the key safety indicator and the worker is determined based on the comparison result of the minimum distance between the edges of the two and the preset threshold value, wherein the preset threshold value is determined by the positional relationship of the standard wearing sample.

[0129] Specifically, in step ③, since it is crucial to ensure that personnel behavior meets the safety operation specification in various work scenes, the present application constructs a logical judgment function to determine whether the personnel behavior in the work scene meets the safety operation specification after the model identifies the current work scene, including but not limited to the personnel's action, position, tools and equipment used, etc.

[0130] Specifically, first, the key safety indicators need to be determined according to different work scenes. For example, in the work scene of power equipment maintenance, the key indicators are whether to wear insulating gloves, safety hats, etc.; in the work scene of high-altitude detection, the key indicators include whether to fasten the safety belt. According to the safety operation requirements of different scenes, the corresponding key safety indicators are listed.

[0131] Specifically, let the key safety indicator set be S i ={S i1 ,S i2 ,...,S in}, where n is the number of safety indicators in the ith work scene. The value range of each key safety indicator is {0, 1}, where 1 represents compliance and 0 represents non-compliance. When there is a non-compliant indicator in the work scene, the work scene detected by the model is determined to be not in compliance with the safety operation requirement, and is marked as non-compliant. The logical judgment function f(S)f(S) is represented as:

[0132]

[0133] Specifically, when the model detects the work scene, the work personnel and the corresponding key indicators are labeled using the labeled boxes. Considering the standardization and accuracy of personnel behavior, the positional relationship between the personnel work scene marking box and the key indicator marking box is integrated to distinguish irrelevant key indicators in the scene, such as safety helmets placed on the workbench. The position judgment logic according to the marking box is as follows:

[0134] First, the center position of the work personnel marking box is determined as the reference, and then the center position of the key target marking box is judged in three cases, namely completely inside, no overlap and partial overlap. The logical judgment relationship is designed to distinguish the three relationships, and for the case of partial overlap, an error threshold is set to make the judgment result more accurate and consistent with the actual situation.

[0135] Let the lower left corner of the image be the origin (x n ,y n ), the horizontal right direction be the positive direction of the x-axis, and the vertical downward direction be the positive direction of the y-axis. The range of the work personnel marking box can be represented by the left upper corner coordinates and the right lower corner coordinates of the marking box Correspondingly, the position range of the key target marking box is represented as Then, by comparing the position of the key indicator marking box with the position of the work personnel marking box, the corresponding relationship between the key indicator position and the work personnel is judged:

[0136] If the key indicator marking box is completely inside the work personnel marking box, then the key indicator is related to the work personnel, such as the safety helmet worn by the work personnel, etc.:

[0137]

[0138] If the key indicator marking box is not completely inside the work personnel marking box, then the key indicator is not related to the work personnel, such as the safety helmet placed on the workbench, etc. which has no direct relationship with the work personnel:

[0139]

[0140] If the key indicator marking box is neither completely contained nor completely non-overlapping with the work personnel marking box, then further judgment is needed according to the specific situation. The present application sets a coordinate distance threshold d (adopting samples to obtain a value) to determine the correlation between the two, when the edge distance between the key indicator marking box and the work personnel marking box is within the threshold range, it is considered to be related, otherwise, it is not. For example, the marking box of the safety helmet and the marking box of the work personnel have partial overlap, at this time, the edge distance between the two is considered to judge whether it is related to the work personnel:

[0141]

[0142] Collect 100 groups of standard wearing samples, measure the minimum effective distance, record the maximum reasonable distance, and obtain the value of d according to the statistical data.

[0143] Wherein:

[0144] Completely contained: the safety helmet is completely in the personnel frame → standard wearing;

[0145] No overlap: the safety helmet is placed independently → environmental articles;

[0146] Partial overlap + distance ≤ d: the safety helmet is offset but still within the associated range → non-standard wearing;

[0147] Partial overlap + distance > d: adjacent but unrelated safety helmets → environmental articles;

[0148] In one embodiment, according to the judgment result of the safety operation state, safety warning information is generated and displayed through a man-machine interaction interface.

[0149] Specifically, in step IV, the man-machine interaction page displays a real-time video picture of the work site, and superimposes the model detection result on the picture. Different colored annotation boxes are used to mark the work site personnel and key indicators, wherein a preset green annotation box represents a compliant work site, and a red annotation box represents a non-compliant work site. Statistical information such as the number of current work sites, the number and type distribution of compliant and non-compliant behaviors, and warning information is displayed in the form of a chart on the left side of the interface, wherein the warning information expansion needs to include detailed non-compliant information such as work site, personnel information, and non-compliant behavior description, so as to facilitate the management personnel to quickly understand the overall situation and take corresponding measures.

[0150] Specifically, the work site recognition and warning system is integrated with the safety management system to realize the sharing and unified management of safety data. The management personnel can query the basic information, behavior records and historical warning information of a specific personnel by inputting the number of the work personnel, the time range and other selection conditions, so as to better evaluate the safety behavior of the personnel and targeted training; the safety situation of a certain work site in a certain time period can also be queried for post-analysis and evaluation. For some complex situations or model misjudgment situations, the management personnel can manually mark on the man-machine interaction interface and provide corresponding explanations, and these feedback information will be used for subsequent optimization and improvement of the model.

[0151] In order to facilitate the understanding of the above technical solutions of the present application, the following will be specifically described as follows by taking a substation maintenance operation as an example:

[0152] When equipment maintenance work is carried out in the 220kV substation, the safety substation operation scene recognition method based on the light-weight Yolov7 is adopted, and intelligent supervision of the safety behavior of the operation personnel is realized.

[0153] As shown in Figure 3 , the training process of the safety supervision recognition model first collects video data of different operation scenes through high-definition cameras installed in multiple areas of the substation, extracts key frames through optical flow analysis, and constructs an initial sample set. Then, these key frames are labeled and hierarchically divided according to safety semantic categories (such as wearing a safety helmet, using an insulating glove, and following a safe operation specification), forming a structured training sample set. Then, the improved light-weight Yolov7 model is trained using these sample sets, which uses a hollow convolution to replace the maximum pooling layer, a CBAM attention mechanism to enhance feature extraction capability, and a residual connection to retain original feature information. During the training process, the model parameters are continuously optimized by dynamically adjusting the learning rate and data enhancement strategy, and finally the safety supervision recognition model suitable for the substation operation scene is obtained.

[0154] As shown in Figure 2 , the recognition of the substation operation scene first acquires real-time operation video streams through a video acquisition device, and after preprocessing, inputs them into the safety supervision recognition model for target detection. The present application can identify the operation personnel and their key safety indicators (such as safety helmets, insulating gloves, safety belts, etc.), and mark the detection results on the video screen. Then, according to the type of the operation scene, the corresponding safety rule library is called, and the positional relationship between the operation personnel and the safety indicators is analyzed through a pre-set logical judgment function to determine whether the operation personnel comply with the safety operation specification. Finally, safety warning information is generated according to the judgment result, and is displayed in real time through a human-computer interaction interface, so that the supervisory personnel can timely discover and handle safety hazards.

[0155] In actual application, when it is detected that an operation personnel is operating in a high-voltage equipment area, the system automatically identifies that the scene type is "high-voltage equipment maintenance", and determines the corresponding set of key safety indicators (safety helmet, insulating glove, insulating shoe, etc.). By analyzing the positional relationship between the operation personnel and the safety indicators, the system determines that the operation personnel has worn a safety helmet but has not worn an insulating glove correctly, immediately marks this violation behavior with a red marked box on the monitoring interface, and generates the warning information "high-voltage equipment area operation without wearing insulating gloves" to push to the safety management personnel, effectively preventing the occurrence of safety accidents.

[0156] As shown in Figure 5 , according to another embodiment of the present application, a safety substation operation scene recognition system based on light-weight Yolov7 is also provided, which comprises:

[0157] A model construction module 1 is configured to construct a labeled sample set by analyzing optical flow changes to extract key frames based on substation operation scene data, and train an improved Yolov7 model through the labeled sample set to obtain an safety supervision recognition model;

[0158] A scene recognition module 2 is configured to use the safety supervision recognition model to perform target detection on a real-time substation operation scene, obtain position information of an operator and a safety index, and analyze the position information through a preset logical judgment function to determine a safe operation state of the operator;

[0159] A warning display module 3 is configured to generate safety warning information according to the determination result of the safe operation state, and display the safety warning information through a human-computer interaction interface.

[0160] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying a safe power transformation operation scene based on lightweight Yolov7, characterized in that, The method comprises the following steps: S1, based on the substation operation scene data, the key frames are extracted by analyzing the optical flow change to construct the labeled sample set, and the improved Yolov7 model is trained through the labeled sample set to obtain the safety supervision recognition model; S2, using the safety supervision recognition model to detect the target of the real-time substation operation scene, obtaining the position information of the operation personnel and the safety index, and analyzing the position information through a preset logical judgment function to determine the safety operation state of the operation personnel; S3, generating safety warning information according to the determination result of the safety operation state, and displaying through the man-machine interface.

2. The method according to claim 1, wherein The safety supervision recognition model is obtained by analyzing the optical flow change to extract the key frames based on the substation operation scene data, and training the improved Yolov7 model through the labeled sample set, which comprises the following steps: S11, based on the personnel behavior video data of the substation operation scene, the key frames are extracted according to the analysis of the optical flow change to construct the key frame data set; S12, the key frame data set is semantically labeled and hierarchically divided according to the safety semantic category to form a labeled sample set; S13, using the labeled sample set to train the lightweight Yolov7 model improved by random weighted features and residual connection to obtain the safety supervision recognition model for the substation operation scene; The lightweight Yolov7 model comprises an input module, a backbone network, a detection head and an output module, and the detection head comprises a spatial pyramid pooling module.

3. The method according to claim 2, wherein, The key frame data set is constructed by extracting the key frames according to the analysis of the optical flow change based on the personnel behavior video data of the substation operation scene, which comprises the following steps: S111, using the video acquisition device covering the substation operation area to acquire personnel behavior videos at different time periods and different operation stages and number them to obtain an original video sequence; S112, based on the original video sequence, the optical flow change between adjacent frames is compared by motion vector analysis, and the current frame exceeding the threshold value is extracted as a key frame to generate an initial key frame set; S113, screening the initial key frame set and optimizing the image parameters to obtain the key frame data set.

4. The method according to claim 3, wherein, The key frame is extracted by comparing the optical flow change between adjacent frames through motion vector analysis, which comprises the following steps: Calculate the motion components of all pixel points in the horizontal and vertical directions of the video frame to obtain the pixel displacement value; Square and square root operations are performed on the displacement values of all pixel points to obtain the pixel motion amplitude; Calculate the average value of the total pixel motion amplitude and compare it with the preset threshold value; When the average motion amplitude is greater than the preset threshold value, the current frame is determined as a key frame; Wherein, the first preset threshold value is set for the video of the static scene, and the second preset threshold value is set for the video of the dynamic scene, and the second preset threshold value is greater than the first preset threshold value.

5. The method according to claim 2, wherein The safety supervision recognition model for the substation operation scene is obtained by training the lightweight Yolov7 model improved by random weighted features and residual connection using the labeled sample set, which comprises the following steps: S131, in the lightweight Yolov7 model feature extraction stage, based on the spatial pyramid pooling module, a plurality of parallel hole convolution layers are used to replace the maximum pooling layer, the input feature map is convolved by different hole convolution operations with different hole rates, and the obtained multi-scale feature maps are spliced in the channel dimension to obtain the multi-scale fusion features of the labeled sample set; S132, in the lightweight Yolov7 model feature enhancement stage, the multi-scale fusion features of the labeled sample set are applied to the channel spatial attention mechanism, and the attention weighted features are obtained by using the channel attention module and the spatial attention module; S133, in the lightweight Yolov7 model feature fusion stage, the attention weighted features are connected with the input feature map in residual connection to generate the output feature map, so as to complete the model training and obtain the safety supervision identification model for the substation operation scene.

6. The method according to claim 5, wherein, The plurality of parallel hole convolution layers are used to replace the maximum pooling layer, and the input feature map is convolved by different hole convolution operations with different hole rates, including: A plurality of parallel hole convolution layers are set, each hole convolution layer uses a convolution kernel with the same size and the same number of output channels, and is configured with different hole rates; Each hole convolution layer is used to convolve the input feature map respectively, and small target detail features, medium scale features and large scene outline features are captured respectively under the premise of keeping the receptive field consistent; The output feature maps of each hole convolution layer are spliced in the channel dimension to form a multi-scale fusion feature map.

7. The method according to claim 5, wherein, The channel attention module is used to identify the key feature channels related to safety equipment and assign weights, and the spatial attention module is used to locate the risk area of the operation and enhance the spatial feature weight; The multi-scale fusion features of the labeled sample set are applied to the channel spatial attention mechanism, and the attention weighted features are obtained by using the channel attention module and the spatial attention module, including: Based on the channel attention module, the multi-scale fusion feature map is subjected to global average pooling and global maximum pooling, the pooling results are input into a shared multi-layer perception machine for feature fusion, and the channel attention weight is obtained after the activation function processing, and the channel attention weighted feature map is obtained by multiplying the multi-scale fusion feature map; According to the spatial attention module, the channel attention weighted feature map is subjected to average pooling and maximum pooling in the channel dimension, the pooling results are spliced, subjected to convolution dimension reduction processing and activated by an activation function, a spatial attention weight matrix is obtained, and a spatial attention weighted feature map is obtained by multiplying the channel attention weighted feature map.

8. The method according to claim 1, wherein, The safety supervision identification model is used for target detection on the real-time substation operation scene, the position information of the operation personnel and safety indicators is obtained, and the safety operation state of the operation personnel is judged by analyzing the position information through a preset logical judgment function, including: S21, according to the identified substation operation scene type, determine the corresponding key safety indicator set, wherein the key safety indicator set includes safety elements determined according to different operation scenes; S22, the safety supervision identification model is used for target detection on the real-time operation scene, and the position information of the operation personnel label frame and the key safety indicator label frame is obtained; S23, judging the relevance of the key safety indicators and the workers according to the position relationship between the worker annotation box and the key safety indicator annotation box, and determining whether it meets the safety operation requirements by using a logical judgment function.

9. The method according to claim 8, wherein, The judgment of the relevance of the key safety indicators and the workers according to the position relationship between the worker annotation box and the key safety indicator annotation box comprises: The lower left corner of the image is set as the coordinate origin to establish a two-dimensional coordinate system, and the upper left corner coordinates and the lower right corner coordinates of the worker annotation box are obtained to determine the position range of the worker annotation box; The upper left corner coordinates and the lower right corner coordinates of the key safety indicator annotation box are obtained to determine the position range of the key safety indicator annotation box; The position relationship between the key safety indicator annotation box and the worker annotation box is judged, and if the key safety indicator annotation box is completely located inside the worker annotation box, it is determined that the key safety indicator is related to the worker; If the key safety indicator annotation box is completely located outside the worker annotation box without overlapping, it is determined that the key safety indicator is irrelevant to the worker; If the key safety indicator annotation box and the worker annotation box have partial overlap, the relevance of the key safety indicator and the worker is determined based on the comparison result of the minimum distance between the edges of the two and a preset threshold value, wherein the preset threshold value is determined by statistical standard wearing samples.

10. A safety substation operation scene identification system based on lightweight Yolov7, used to implement the safety substation operation scene identification method based on lightweight Yolov7 in any one of claims 1-9, characterized in that, The system comprises: A model construction module is configured to construct a labeled sample set by analyzing optical flow changes to extract key frames based on substation operation scene data, and train an improved Yolov7 model through the labeled sample set to obtain an occupational safety and health identification model. A scene recognition module is configured to use the occupational safety and health identification model to perform target detection on a real-time substation operation scene, obtain position information of workers and safety indicators, and analyze the position information through a preset logical judgment function to determine the safety operation state of the workers. An early warning display module is configured to generate safety warning information according to the determination result of the safety operation state, and display the safety warning information through a human-computer interaction interface.

Citation Information

Cited By

  • Crane full-stroke operation safety monitoring system and method

    CN121292288A