Electric power operation safety detection method based on target detection network

Abstract

The invention discloses an electric power operation safety detection method based on a target detection network, and the method comprises the following steps: firstly collecting pictures shot in an electric power operation scene as a data set, carrying out the corresponding preprocessing of the pictures of the data set, and dividing the pictures into a training set and a test set; then designing a structure of a target detection network suitable for power operation safety detection; inputting the data set picture into a target detection network for training to obtain a target detection model; and finally, performing safety detection by using the trained target detection model to obtain a safety detection result in the electric power operation scene. According to the invention, the improved channel attention mechanism is integrated into the target detection network with a lightweight structure, and a more effective loss function is designed, so that the interference of picture background information is more effectively inhibited, better target positioning is realized, and the detection of key targets is enhanced. Therefore, the accuracy, the detection speed and the reliability of the overall safety detection of the model are improved, and safer and more efficient operation support is provided for power workers.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and more specifically, to a power operation safety detection method based on a target detection network. Background Art

[0002] With the rapid development of science and technology, power operation safety detection has become a key link to ensure the stable operation of the power system and the safety of personnel. Traditional power operation safety detection methods often rely on manual inspections and simple sensor equipment, which is not only inefficient but also easy to miss potential safety hazards. Therefore, how to use advanced technical means to improve the accuracy and efficiency of power operation safety detection is an urgent problem to be solved.

[0003] In the field of computer vision, target detection is one of the most popular research areas, with wide applications in intelligent monitoring, autonomous driving and other fields. In the safety detection of power operations, the application of target detection technology is also particularly important. Through real-time monitoring of the operation site, abnormal situations can be discovered in time, such as illegal operations by personnel, equipment failures, etc., so that timely and effective measures can be taken to avoid accidents.

[0004] In recent years, target detection models based on deep neural networks have made significant progress in the field of target detection. These models have the advantages of high recognition accuracy and fast speed, and have become the mainstream in target detection algorithms. However, for the specific application scenario of power operation safety detection, the existing target detection models still have some shortcomings. First, the on-site environment of power operations is complex and changeable, and there are a large number of interference factors, which requires the model to have stronger robustness and adaptability. Secondly, power operation safety detection has high real-time requirements, so it is necessary to increase the detection speed as much as possible while ensuring accuracy.

[0005] The Chinese patent with publication number CN114155487A discloses "a method for detecting electric power workers based on multi-group convolution fusion", which enhances the feature expression ability of the target by adding N groups of parallel convolution fusion modules to the feature extraction network, achieves high-precision real-time processing effects, and improves the performance of safety detection of electric power workers. However, this method lacks processing of image background information and has a large computational overhead, which is far from enough for electric power operation sites that require models to have stronger robustness and detection speed. Summary of the invention

[0006] In view of the above defects or improvement needs of the prior art, the present invention provides a power operation safety detection method based on a target detection network, which aims to integrate the improved channel attention mechanism into a target detection network with a lightweight structure and suppress the interference of background information in the detection image by designing a more effective loss function, so as to achieve better target positioning and strengthen the detection of key targets, thereby improving the robustness of the overall safety detection of the model, the detection speed and accuracy, and providing safer and more efficient operation support for power workers.

[0007] To achieve the above purpose, according to one aspect of the present invention, a method for detecting safety of electric power operations based on a target detection network is provided. The method for detecting safety of electric power operations comprises the following steps:

[0008] The output obtained by the power operation safety detection model detection includes the location information and category probability of each detected object. The pictures taken in the power operation scene are collected as a data set, and the data set pictures are preprocessed accordingly and divided into a training set and a test set;

[0009] S2. Design the structure of the target detection network suitable for power operation safety detection;

[0010] S3, input the data set images into the target detection network for training to obtain the power operation safety detection model;

[0011] S4. Use the trained power operation safety detection model to perform safety detection and obtain safety detection results in the power operation scenario.

[0012] Optionally, the S1 includes the following sub-processes:

[0013] S11. Organize the photos collected at the power operation site into a data set, and then format the data set images, remove the operation sound and other pre-processing;

[0014] S12. Use Labelimg annotation tool to label the preprocessed dataset images. The specific labeling types are safety helmet, safety belt, work clothes jacket, work clothes pants, insulating gloves, and insulating boots.

[0015] S13. The labeled data set is divided into a training set and a test set in a ratio of 8:2.

[0016] Optionally, the structure of the target detection network consists of 6 depth-separable convolutional layers with different convolution kernel sizes, 6 identical channel attention modules and a security detection prediction module.

[0017] Optionally, when the dataset images are input into the target detection network for training, the total loss function used is a combination of the classification loss and the positioning loss, as follows:

[0018] First, the classification loss Loss1 is expressed as follows:

[0019]

[0020] Among them, N is the number of input samples, M is the number of object categories detected, and w j represents the weight of category j, y ij represents the true label of category j in sample i, p ij represents the predicted probability of category j in sample i, and γ is a focusing parameter used to control the degree of weight adjustment of difficult and easy samples;

[0021] Then there is the positioning loss Loss2, which is expressed as follows:

[0022]

[0023] Where N is the number of input samples, M is the number of object categories detected, and each category in each sample predicts 4 bounding box coordinates, tx ijk represents the horizontal coordinate of the kth bounding box coordinate of category j in sample i, is the corresponding real horizontal coordinate, ty ijk represents the ordinate of the kth bounding box coordinate of category j in sample i, is the corresponding real ordinate;

[0024] The total loss of the training model is Loss_all, which is expressed as:

[0025] Loss_all=β·Loss1+(1-β)·Loss2

[0026] Among them, β is a hyperparameter used to balance the weight between classification loss Loss1 and localization loss Loss2.

[0027] Optionally, use the trained power operation safety detection model to perform safety detection, including:

[0028] S41, the input image is first processed by data enhancement, and then input into the first convolution layer for feature extraction. The convolution kernel size used in the convolution layer is 3×3, and the step size is 1. After the convolution operation, batch normalization and ReLU activation function are also used. Each subsequent convolution layer will also perform the same processing after the convolution operation;

[0029] S42, input the feature map processed by the first convolution into the second convolution layer and the first channel attention module at the same time, the convolution kernel size used in the second convolution layer is 5×5, and the step size is 1;

[0030] S43, the feature map obtained after the second convolution layer is processed continues to be input into the third convolution layer and the second channel attention module, wherein the convolution kernel size used in the third convolution layer is 1×1 and the step size is 1;

[0031] S44, adding the output of the first channel attention module and the output of the second channel attention module, and inputting the output of the third convolutional layer into the fourth convolutional layer and the third channel attention module, wherein the convolution kernel size used in the fourth convolutional layer is 3×3 and the step size is 1;

[0032] S45. Similar to the previous step, the output of the channel attention module is summed, and the output of the convolution layer continues to be input into the next convolution layer and the next channel attention module. The next convolution layer is the fifth convolution layer, and its convolution kernel size is 5×5 and the step size is 1.

[0033] S46, repeat the above operation, but this time the convolution kernel size in the convolution layer is 7×7 and the step size is 1;

[0034] S47. Add the output of the sixth convolutional layer and the output of the sixth channel attention module, and then send the fused feature map to the safety detection prediction module. The safety detection prediction module will obtain the location information and category probability of each detected object, and output the power operation safety detection result based on the obtained information, where the location information includes 4 bounding box coordinates.

[0035] Optionally, the channel attention module uses an improved channel attention, and the processing of the feature map in it includes the following specific steps:

[0036] A1. Perform global average pooling along the width W and the height H on each channel of the input feature map of size C×H×W, and obtain a C×H×1 feature map and a C×1×W feature map, capturing information in the height direction and the width direction, where C is the number of channels, H is the height, and W is the width;

[0037] A2. Concatenate the two feature maps obtained in the previous step in the channel dimension to obtain a feature map of C×1×(W+H);

[0038] A3. Input the concatenated feature map into a 1×1 convolutional layer to fuse and transform features.

[0039] A4. Input the feature map of size C / m×1×(W+H) after convolution into the post-processing layer. The post-processing layer will batch normalize the feature map after convolution and increase the expression ability of the model through nonlinear activation function, where m is the scaling factor.

[0040] A5. Divide the feature map of size C / m×1×(W+H) processed by the post-processing layer into two parts and input them into a 1×1 convolutional layer respectively;

[0041] A6. Two 1×1 convolutional layers output a feature map of size C×H×1 and a feature map of size C×1×W. The two feature maps are respectively input into a weight activation layer to obtain two different weight maps.

[0042] A7. Multiply the input feature map by the two weight maps channel by channel to perform feature weighting, and obtain the final weighted feature map of size C×H×W.

[0043] Optionally, in step S4, obtaining safety detection results in the power operation scenario specifically includes:

[0044] The output obtained by the power operation safety detection model detection includes the location information and category probability of each detected object;

[0045] According to the category probability and location information in the model output, it is compared with the predefined threshold. If the category probability is higher than the threshold and the location information is correct, the existence of the category object can be determined and the security factor is met.

[0046] A comprehensive judgment is made based on the judgment results of each category object. If all necessary safety factors exist and meet the requirements, the safety of the power operation can be judged and the output power operation safety test is qualified. Otherwise, the output power operation safety test is unqualified.

[0047] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0048] S1. The target detection network is designed with a lightweight structure, which helps the model to have a faster detection speed when performing power operation safety detection.

[0049] S2. By integrating the improved channel attention mechanism into the designed target detection network, the accuracy and robustness of the model in power operation safety detection are improved.

[0050] S3. Use innovative loss functions to make the target positioning more accurate during model training, while also achieving faster convergence and better training results, making it more suitable for power operation safety detection scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A technical solution flow chart of a power operation safety detection method based on a target detection network is provided in an implementation case of the present invention.

[0052] Figure 2A schematic diagram of the target detection network structure of a power operation safety detection method based on a target detection network is provided in an implementation example of the present invention.

[0053] Figure 3 A schematic diagram of a channel attention module of a power operation safety detection method based on a target detection network is provided in an implementation example of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0055] The present invention provides an implementation scheme of a power operation safety detection method based on an improved channel attention mechanism, see the attached Figure 1 , Figure 1 This is a technical solution flow chart of a power operation safety detection method based on an improved channel attention mechanism provided by the implementation use case, which specifically includes the following steps:

[0056] S1. Collect pictures taken in power operation scenes as a data set, pre-process the pictures in the data set accordingly, and divide them into a training set and a test set;

[0057] Wherein, the S1 includes the following sub-processes:

[0058] S11. Organize the photos collected at the power operation site into a data set, and then format the data set images, remove the operation sound and other pre-processing;

[0059] S12. Use Labelimg annotation tool to label the preprocessed dataset images. The specific labeling types are safety helmet, safety belt, work clothes jacket, work clothes pants, insulating gloves, and insulating boots.

[0060] S13. The labeled data set is divided into a training set and a test set in a ratio of 8:2.

[0061] S2. Design the structure of the target detection network suitable for power operation safety detection;

[0062] The designed target detection network structure consists of 6 depth-separable convolutional layers with different convolution kernel sizes, 6 identical channel attention modules, and a security detection prediction module. Figure 2 , Figure 2This is a schematic diagram of the target detection network structure of a power operation safety detection method based on an improved channel attention mechanism provided by the implementation case.

[0063] S3, input the data set images into the target detection network for training to obtain the power operation safety detection model;

[0064] Among them, the total loss function used when training the target detection network is composed of classification loss and positioning loss. The details are as follows:

[0065] First, the classification loss Loss1 is expressed as follows:

[0066]

[0067] Among them, N is the number of input samples, M is the number of object categories detected, and w j represents the weight of category j, y ij represents the true label of category j in sample i, p ij represents the predicted probability of category j in sample i, and γ is a focusing parameter used to control the degree of weight adjustment of difficult and easy samples;

[0068] Then there is the positioning loss Loss2, which is expressed as follows:

[0069]

[0070] Where N is the number of input samples, M is the number of object categories detected, and each category in each sample predicts 4 bounding box coordinates, tx ijk represents the horizontal coordinate of the kth bounding box coordinate of category j in sample i, is the corresponding real horizontal coordinate, ty ijk represents the ordinate of the kth bounding box coordinate of category j in sample i, is the corresponding real ordinate;

[0071] The total loss of the training model is Loss_all, which is expressed as:

[0072] Loss_all=β·Loss1+(1-β)·Loss2

[0073] Among them, β is a hyperparameter used to balance the weight between classification loss Loss1 and localization loss Loss2.

[0074] S4. Use the trained power operation safety detection model to perform safety detection and obtain safety detection results in the power operation scenario.

[0075] The trained power operation safety detection model is used for safety detection, including:

[0076] S41, the input image is first processed by data enhancement, and then input into the first convolution layer for feature extraction. The convolution kernel size used in the convolution layer is 3×3, and the step size is 1. After the convolution operation, batch normalization and ReLU activation function are also used. Each subsequent convolution layer will also perform the same processing after the convolution operation;

[0077] S42, input the feature map processed by the first convolution into the second convolution layer and the first channel attention module at the same time, the convolution kernel size used in the second convolution layer is 5×5, and the step size is 1;

[0078] S43, the feature map obtained after the second convolution layer is processed continues to be input into the third convolution layer and the second channel attention module, wherein the convolution kernel size used in the third convolution layer is 1×1 and the step size is 1;

[0079] S44. Add the output of the first channel attention module and the output of the second channel attention module. At the same time, the output of the third convolutional layer is input into the fourth convolutional layer and the third channel attention module, wherein the convolution kernel size used in the fourth convolutional layer is 3×3, and the step size is 1.

[0080] S45. Similar to the previous step, the output of the channel attention module is summed, and the output of the convolution layer continues to be input into the next convolution layer and the next channel attention module. The next convolution layer is the fifth convolution layer, and its convolution kernel size is 5×5 and the step size is 1.

[0081] S46, repeat the above operation, but this time the convolution kernel size in the convolution layer is 7×7 and the step size is 1;

[0082] S47. Add the output of the sixth convolutional layer and the output of the sixth channel attention module, and then send the fused feature map to the safety detection prediction module. The safety detection prediction module will obtain the location information and category probability of each detected object, and output the power operation safety detection result based on the obtained information, where the location information includes 4 bounding box coordinates.

[0083] Among them, the channel attention module uses the improved channel attention, see Figure 3 , Figure 3 A schematic diagram of a channel attention module of a power operation safety detection method based on a target detection network provided in an implementation example of the present invention, wherein the processing of a feature graph therein includes the following specific steps:

[0084] A1. Perform global average pooling along the width W and the height H on each channel of the input feature map of size C×H×W, and obtain a C×H×1 feature map and a C×1×W feature map, capturing information in the height direction and the width direction, where C is the number of channels, H is the height, and W is the width;

[0085] A2. Concatenate the two feature maps obtained in the previous step in the channel dimension to obtain a feature map of C×1×(W+H);

[0086] A3. Input the concatenated feature map into a 1×1 convolutional layer to fuse and transform features.

[0087] A4. Input the convolved feature map (size is C / m×1×(W+H), r is a scaling factor) into the post-processing layer. The post-processing layer will batch normalize the convolved feature map and increase the model's expressiveness through a nonlinear activation function, where m is the scaling factor.

[0088] A5. Divide the feature map of size C / m×1×(W+H) processed by the post-processing layer into two parts and input them into a 1×1 convolutional layer respectively;

[0089] A6. Two 1×1 convolutional layers output a feature map of size C×H×1 and a feature map of size C×1×W. The two feature maps are respectively input into a weight activation layer to obtain two different weight maps.

[0090] A7. Multiply the input feature map by the two weight maps channel by channel to perform feature weighting, and obtain the final weighted feature map of size C×H×W.

[0091] Furthermore, the safety detection results in the power operation scenario are obtained, which specifically includes the following sub-processes:

[0092] The output obtained by the power operation safety detection model detection includes the location information and category probability of each detected object;

[0093] According to the category probability and location information in the model output, it is compared with the predefined threshold, where the threshold is 0.5. If the category probability is higher than the threshold and the location information is correct, the existence of the category object can be determined and the security factor is satisfied;

[0094] A comprehensive judgment is made based on the judgment results of each category object. If all necessary safety factors exist and meet the requirements (for example, safety helmets, safety belts, insulating gloves, etc.), the safety of the power operation can be judged to be qualified, otherwise, the safety test of the output power operation is unqualified;

[0095] In summary, this application loads a pre-trained target detection model with a lightweight structure, performs image preprocessing and model inference on the power operation site pictures to be analyzed, and obtains the location information and category probability of each detection object. Then, it is determined whether the location information is correct, whether the category probability is greater than the predefined threshold (set to 0.5), whether the category object exists, and whether the safety factor is satisfied. Finally, according to the judgment results of each category object, a comprehensive judgment of power operation safety detection is made.

[0096] Those skilled in the art may make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for detecting power operation safety based on a target detection network, characterized in that: The power operation safety detection method comprises the following steps: The output obtained by the power operation safety detection model detection includes the location information and category probability of each detected object. The pictures taken in the power operation scene are collected as a data set, and the data set pictures are preprocessed accordingly and divided into a training set and a test set; S2. Design the structure of the target detection network suitable for power operation safety detection; S3, input the data set images into the target detection network for training to obtain the power operation safety detection model; S4. Use the trained power operation safety detection model to perform safety detection and obtain safety detection results in the power operation scenario.

2. The method for detecting safety of electric power operations based on a target detection network according to claim 1, characterized in that: The S1 includes the following sub-processes: S11. Organize the photos collected at the power operation site into a data set, and then format the data set images, remove the operation sound and other pre-processing; S12. Use Labelimg annotation tool to label the preprocessed dataset images. The specific labeling types are safety helmet, safety belt, work clothes jacket, work clothes pants, insulating gloves, and insulating boots. S13. The labeled data set is divided into a training set and a test set in a ratio of 8:

2.

3. The method for detecting safety of electric power operations based on a target detection network according to claim 2, characterized in that: The structure of the target detection network consists of 6 depth-wise separable convolutional layers with different convolution kernel sizes, 6 identical channel attention modules and a security detection prediction module.

4. The method for detecting safety of electric power operations based on a target detection network as claimed in claim 3, characterized in that: When the dataset images are input into the target detection network for training, the total loss function used is a combination of classification loss and positioning loss, as follows: First, the classification loss Loss1 is expressed as follows: Among them, N is the number of input samples, M is the number of object categories detected, and w j represents the weight of category j, y ij represents the true label of category j in sample i, p ij represents the predicted probability of category j in sample i, and γ is a focusing parameter used to control the degree of weight adjustment of difficult and easy samples; Then there is the positioning loss Loss2, which is expressed as follows: Where N is the number of input samples, M is the number of object categories detected, and each category in each sample predicts 4 bounding box coordinates, tx ijk represents the horizontal coordinate of the kth bounding box coordinate of category j in sample i, is the corresponding real horizontal coordinate, ty ijk represents the ordinate of the kth bounding box coordinate of category j in sample i, is the corresponding real ordinate; The total loss of the training model is Loss_all, which is expressed as: Loss_all=β·Loss1+(1-β)·Loss2 Among them, β is a hyperparameter used to balance the weight between classification loss Loss1 and localization loss Loss2.

5. The method for detecting safety of electric power operations based on a target detection network as claimed in claim 4, characterized in that: Use the trained power operation safety detection model to perform safety detection, including: S41, the input image is first processed by data enhancement, and then input into the first convolution layer for feature extraction. The convolution kernel size used in the convolution layer is 3×3, and the step size is 1. After the convolution operation, batch normalization and ReLU activation function are also used. Each subsequent convolution layer will also perform the same processing after the convolution operation; S42, input the feature map processed by the first convolution into the second convolution layer and the first channel attention module at the same time, the convolution kernel size used in the second convolution layer is 5×5, and the step size is 1; S43, the feature map obtained after the second convolution layer is processed continues to be input into the third convolution layer and the second channel attention module, wherein the convolution kernel size used in the third convolution layer is 1×1 and the step size is 1; S44, adding the output of the first channel attention module and the output of the second channel attention module, and inputting the output of the third convolutional layer into the fourth convolutional layer and the third channel attention module, wherein the convolution kernel size used in the fourth convolutional layer is 3×3 and the step size is 1; S45. Similar to the previous step, the output of the channel attention module is summed, and the output of the convolution layer continues to be input into the next convolution layer and the next channel attention module. The next convolution layer is the fifth convolution layer, and its convolution kernel size is 5×5 and the step size is 1. S46, repeat the above operation, but this time the convolution kernel size in the convolution layer is 7×7 and the step size is 1; S47. Add the output of the sixth convolutional layer and the output of the sixth channel attention module, and then send the fused feature map to the safety detection prediction module. The safety detection prediction module will obtain the location information and category probability of each detected object, and output the power operation safety detection result based on the obtained information, where the location information includes 4 bounding box coordinates.

6. A method for detecting safety of electric power operations based on a target detection network as claimed in claim 5, characterized in that: The channel attention module uses improved channel attention, and the processing of feature maps in it includes the following specific steps: A1. Perform global average pooling along the width W and the height H on each channel of the input feature map of size C×H×W, and obtain a C×H×1 feature map and a C×1×W feature map, capturing information in the height direction and the width direction, where C is the number of channels, H is the height, and W is the width; A2. Concatenate the two feature maps obtained in the previous step in the channel dimension to obtain a feature map of C×1×(W+H); A3. Input the concatenated feature map into a 1×1 convolutional layer to fuse and transform features. A4. Input the feature map of size C / m×1×(W+H) after convolution into the post-processing layer. The post-processing layer will batch normalize the feature map after convolution and increase the expression ability of the model through nonlinear activation function, where m is the scaling factor. A5. Divide the feature map of size C / m×1×(W+H) processed by the post-processing layer into two parts and input them into a 1×1 convolutional layer respectively; A6. Two 1×1 convolutional layers output a feature map of size C×H×1 and a feature map of size C×1×W. The two feature maps are respectively input into a weight activation layer to obtain two different weight maps. A7. Multiply the input feature map by the two weight maps channel by channel to perform feature weighting, and obtain the final weighted feature map of size C×H×W.

7. A method for detecting safety of electric power operations based on a target detection network as claimed in claim 6, characterized in that: In step S4, the safety detection results in the power operation scenario are obtained, which specifically include: The output obtained by the power operation safety detection model detection includes the location information and category probability of each detected object; According to the category probability and location information in the model output, it is compared with the predefined threshold. If the category probability is higher than the threshold and the location information is correct, the existence of the category object can be determined and the security factor is satisfied; A comprehensive judgment is made based on the judgment results of each category object. If all necessary safety factors exist and meet the requirements, the safety of the power operation can be judged and the output power operation safety test is qualified. Otherwise, the output power operation safety test is unqualified.

Images