Electric power high-altitude operation safety belt detection method

Abstract

The invention discloses an electric power high-altitude operation safety belt detection method, and the method specifically comprises the following steps: 1, collecting a data set image, and dividing a data set into a training set and a test set; step 2, establishing a pyramid segmentation attention module PSA; step 3, constructing an EPSA network model based on a result obtained in the step 2; step 4, establishing an EPSA-YOLOV5 network model based on a result obtained in the step 3; step 5, the EPSA-YOLOV5 network model in the step 4 is improved; and 6, training the model improved in the step 5 by adopting the training set in the step 1. According to the invention, comprehensive detection of the wearing condition of the high-altitude safety belt is realized, and the accuracy is high.

Description

Technical Field

[0001] The invention belongs to the technical field of high-altitude safety belt detection, and relates to a method for detecting a safety belt for high-altitude electric power operations. Background Art

[0002] Power facilities have become an important pillar of my country's economy. The maintenance of power equipment is very important. With the increase in the number of high-altitude power workers and the complexity and changeability of high-altitude working scenes, there are many potential dangers. Data show that falling from height is the most common accident among all kinds of safety accidents in the power industry. The working environment at height is complex and diverse, and workers need to move frequently during the operation. In this case, correctly wearing safety belts plays a vital role in protecting the lives of workers. Although well-designed high-altitude working protective equipment is available, accidents still occur frequently. The reason is that it is difficult to ensure that the safety belts play their real protective role only by the conscious awareness of the workers. Therefore, it is very important to design an intelligent safety belt detection algorithm, which can not only improve the efficiency of supervision of the wearing of safety belts by workers at heights, but also realize the intelligent safety management of violators and ensure the safety of workers. Summary of the invention

[0003] The purpose of the present invention is to provide a method for detecting a safety belt for electric power high-altitude operations, which realizes a comprehensive detection of the wearing condition of the safety belt at high altitudes with a high accuracy rate.

[0004] The technical solution adopted by the present invention is a method for detecting a safety belt for electric power aerial work, which specifically comprises the following steps:

[0005] Step 1: collect data set images and divide the data set into training set and test set;

[0006] Step 2, establish the pyramid segmentation attention module PSA;

[0007] Step 3, constructing the EPSA network model based on the results obtained in step 2;

[0008] Step 4: Build the EPSA-YOLOV5 network model based on the results obtained in step 3;

[0009] Step 5, improve the EPSA-YOLOV5 network model in step 4;

[0010] Step 6: Use the training set in step 1 to train the improved model in step 5.

[0011] The present invention is also characterized in that:

[0012] The specific process of step 1 is:

[0013] Step 1.1, ground personnel use cameras to collect images of high-altitude workers in the power plant;

[0014] Step 1.2, divide the images collected in step 1.1 into 4 categories, namely: people included in the work site, guardian armbands worn at the work site, people in an off-ground state at the work site, and safety belts worn at the work site. These four categories are labeled as "person", "garders", "offgroundperson", and "safebelt-person" respectively. 90% of the labeled images are used as training sets and 10% as test sets.

[0015] The specific process of step 2 is:

[0016] Step 2.1: Use the squeeze connection module SPC to take the image collected in step 1 as input, divide the input image into S parts, and use the SIFT method to extract multi-scale features for the spatial information on the channel feature map after each segmentation to obtain a multi-scale feature map F i ;

[0017] Step 2.2: Based on the multi-scale feature map F extracted in step 2.1 i , use the SEWeight module to extract the channel attention of the multi-scale feature map and obtain the channel attention weight Z at each different scale i and attention vector Z;

[0018] Step 2.3, based on the channel attention weights Z of different scales obtained in step 2.2 i , use Softmax to recalibrate and get the new attention weight att i ;

[0019] Step 2.4, get the new attention weight att from step 2.3 i And the feature map F of the corresponding scale extracted in step 2.1 i Perform dot multiplication on each element and output the feature map Y after multi-scale feature information attention weighting i ;

[0020] Step 2.5: The feature map Y obtained in step 2.4 i Perform dimension splicing and output multi-scale information feature map Out.

[0021] The specific process of step 3 is: replace the bottleneck 3x3 convolution in the convolutional residual network ResNet with the PSA module built in step 2 to obtain the EPSANET network model.

[0022] The specific process of step 4 is: replace the backbone network in the YOLOV5 model with the EPSANET network model constructed in step 3 to obtain the EPSA-YOLOV5 network model.

[0023] The specific process of step 5 is to improve the non-maximum suppression NMS in the EPSA-YOLOV5 network model, that is, to improve the confidence function. The formula is as follows:

[0024]

[0025] Among them, Si is the confidence of the i-th prediction box, M is the current highest scoring box, bi is the prediction box to be processed, IOU is the intersection and union ratio of M and bi, and N is the set threshold.

[0026] The beneficial effects of the present invention are as follows:

[0027] 1. This paper introduces a new attention module EPSA, which can effectively integrate local and global attention and establish long-range channel dependencies.

[0028] 2. The present invention can more accurately merge contextual features of adjacent scales by extracting spatial information of different scales from each channel feature map.

[0029] 3. The present invention constructs cross-dimensional interaction by extracting channel attention weights of multi-scale feature maps, providing a new solution for the current industrial detection and recognition fields. The detection and recognition of scene seat belts also has strong practical significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of the electric power high-altitude working safety belt detection method of the present invention;

[0031] Figure 2 It is a block diagram of the EPSA module in the electric power aerial work safety belt detection method of the present invention;

[0032] Figure 3 It is a block diagram of the EPSA-YOLOV5 model in the electric power aerial work safety belt detection method of the present invention;

[0033] 4(a) to (c) are simulation result accuracy diagrams of the power aerial work safety belt detection method of the present invention. DETAILED DESCRIPTION

[0034] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] Example 1

[0036] The present invention provides a method for detecting a safety belt for high-altitude electric power operations. The process is as follows: Figure 1As shown, the specific steps include:

[0037] Step 1: Dataset collection.

[0038] Step 1.1: Ground personnel use cameras to collect pictures of aerial workers in a power plant. The pictures obtained are all simulated training pictures of aerial workers in power plants, and the data are labeled using labelimg annotation software;

[0039] Step 1.2, divide the collected images into 4 categories, namely: people at the work site (including supervisors and unrelated personnel), supervisory armbands at the work site (these people are supervisors), people off the ground at the work site (these people are high-altitude power maintenance personnel), and safety belts worn at the work site. These four categories are labeled as "person", "garders", "offgroundperson", and "safebelt-person" respectively. 90% of the labeled images are used as training sets and 10% as test sets.

[0040] Step 2: Establish the mathematical model of PSA (pyramid segmentation attention module):

[0041] Step 2.1, use SPC (squeeze connection module) to take the image collected in step 1 as input, and segment the channels of the two-dimensional image. First, the input image is segmented into S parts, and [X0, X1..., X S -1], the number of channels of each segmented part is C'=CS, where C is the number of channels before segmentation, and each is the feature map Xi∈C'×H×W, i=0,1,...,S-1 after segmentation, where H and W are the height and width of the feature map after segmentation, respectively. After segmentation, the SIFT (scale-invariant feature transformation) method is used to extract multi-scale features for the spatial information on the channel feature map after segmentation. The specific calculation method of the extraction process is as follows: Formula (1) is the multi-scale feature map F extracted using the SIFT algorithm i :

[0042] F i =Conv(K i ×K i , Gi)(X i ),i=0,1,2,....,S-1) (1);

[0043] Among them, K i =2×(i+1)+1, R is a real number.

[0044] Step 2.2: Based on the multi-scale feature map F extracted in step 2.1 i, use the SEWeight module to extract the channel attention of the multi-scale feature map and obtain the channel attention weight Z at each different scale i And the attention vector Z, the specific calculation process is as follows:

[0045] Z i =SEWeight(F i ),i=0,1,2,...,S-1 (2);

[0046] Z=Z0⊕Z1⊕.....⊕Z S-1 (3);

[0047] Step 2.3, based on the channel attention weights Z of different scales obtained in step 2.2 i , use Softmax to recalibrate it and get the new attention weight att i , the weight formula is shown in formula (4).

[0048]

[0049] Step 2.4, get the new attention weight att from step 2.3 i And the feature map F of the corresponding scale extracted in step 2.1 i Perform dot multiplication on each element and output the feature map Y after multi-scale feature information attention weighting i , the calculation process is as follows:

[0050] Y i =F i ⊙att i ,i=1,2,3,...S-1 (5);

[0051] Step 2.5: The feature map Y obtained in step 2.4 after multi-scale channel attention weighting i After dimension concatenation, a feature map with richer multi-scale information is output. The specific calculation is as follows:

[0052] Out=Cat([Y0,Y1...Y S-1 ]) (6).

[0053] Step 3, EPSA (Efficient Pyramid Attention Segmentation Module) network model design. Based on the PSA model established in step 2, the 3x3 convolution in the bottleneck part of ResNet is replaced with PSAmodule to obtain EPSANET to complete the design of the EPSA network model. The structure is as follows: Figure 2 As shown;

[0054] Step 4, EPSA-YOLOV5 network model design, such as Figure 3 As shown;

[0055] Based on the EPSANET obtained in step 3, replace the backbone network in the YOLOV5 model with EPSANET.

[0056] Step 5: Based on the EPSA-YOLOV5 obtained in step 4, the NMS (non-maximum suppression) in its model is improved, namely SOFT-NMS, to reduce the phenomenon of missed detection. The specific process is to improve the confidence function, and the formula is as follows:

[0057]

[0058] In formula (7), Si is the confidence of the i-th prediction box, M is the current box with the highest score, bi is the prediction box to be processed, IOU is the intersection-over-union ratio of M and bi, and N is the set threshold. From formula (7), we can see that the larger the intersection-over-union ratio of M and bi, the faster the score S of bi decreases, and it decays linearly.

[0059] Step 6: Use the training set in step 1 to train the EPSA-YOLOV5 model improved in step 5. The training loss function can be divided into three parts: confidence loss I obj , classification loss I cls And the target box positioning loss I box , the overall calculation formula is as follows:

[0060] LOSS=I obj +I cls +I box (8);

[0061] Step 7, for the network model trained in step 6, select H, mean average precision (mAP), and loss function to evaluate the model training effect. The H value is the weighted harmonic average of precision and recall, which is a comprehensive evaluation indicator proposed by combining the two, and can better avoid problems caused by large differences between the two. Precision is also called precision rate, which refers to the proportion of correct charts in all positive samples predicted by the algorithm;

[0062] The recall rate is also called the recall rate, which is the ratio of the correct results detected by the algorithm to all the targets in the chart. The specific calculation formula is shown in (9):

[0063]

[0064] Among them, AP (Average Precision) can reflect the detection performance of a single target category, mAP (meanAverage Precision) is defined as the mean of AP of all categories. With Recall value as the horizontal axis and Precision value as the vertical axis, we can get the PR curve. AP is the average precision. Simply put, it is the average of the Precision values ​​on the PR curve. The higher it is, the easier it is for the model to maintain high accuracy at a high recall rate. The calculation formula is as follows:

[0065]

[0066]

[0067] Example 2

[0068] On the basis of the embodiment, in the training process of step 6, the training set image marked in step 1 is sent to the EPSA-YOLOV5 detection network for training for 300 iterations (Epoch), and the weight parameters are saved once each Epoch is completed. When the number of iterations reaches about 200 times, the model learning efficiency gradually reaches saturation. After the parameter setting is completed, the network model is trained, and the whole training process is 200 epochs, wherein the first 100 epochs use the frozen part of the network to speed up the training speed, the learning rate is set to 0.001, the batchsize is set to 16, the learning rate of the latter 200 epochs is set to 0.0001, the batchsize is set to 8, and the optimizer uses Adam. The following is respectively based on the mean average precision, accuracy, and recall rate obtained by experiments under four different models of YOLOv5, EPSANet, EPSANet+soft-NMS, and EPSA-YOLOV5 as shown in Figures 4 (a) to (c). Figure 4(a) shows the average precision, Figure 4(b) shows the accuracy, and Figure 4(c) shows the recall.

[0069] Example 3

[0070] The four network models YOLOv5, EPSANet, EPSANet+soft-NMS, and EPSA-YOLOV5 used in the present invention are trained in the training set and validation set of the data set, and the performance of the models is compared and analyzed on the test set. The performance comparison of different network models on the test set is shown in Table 1.

[0071] Table 1 Comparison of detection results of different network models

[0072] Network Model mAP_0.5: 0.95 Precision recall YOLOV5 0.6926 0.8993 0.9133 EPSANET 0.7245 0.9050 0.9235 EPSANET-SOFTNMS 0.6898 0.9122 0.9031 EPSA-YOLOv5 0.7683 0.9227 0.9502

[0073] Experimental results show that EPSANET-YOLOV5 has a 7.57% higher MAP and a 3.69% higher recall rate than the original YOLOV5 network, and its detection effect is better than other common target detection networks.

Claims

1. A method for detecting safety belts for electric power aerial work, characterized in that: The specific steps include: Step 1: collect data set images and divide the data set into training set and test set; Step 2, establish the pyramid segmentation attention module PSA; Step 3, constructing the EPSA network model based on the results obtained in step 2; Step 4: Build the EPSA-YOLOV5 network model based on the results obtained in step 3; Step 5, improve the EPSA-YOLOV5 network model in step 4; Step 6: Use the training set in step 1 to train the improved model in step 5.

2. The method for detecting a safety belt for electric power aerial work according to claim 1, characterized in that: The specific process of step 1 is as follows: Step 1.1, ground personnel use cameras to collect images of high-altitude workers in the power plant; Step 1.2, divide the images collected in step 1.1 into 4 categories, namely: people included in the work site, guardian armbands worn at the work site, people in an off-ground state at the work site, and safety belts worn at the work site. These four categories are labeled as "person", "garders", "offgroundperson", and "safebelt-person" respectively. 90% of the labeled images are used as training sets and 10% as test sets.

3. The method for detecting a safety belt for electric power aerial work according to claim 1, characterized in that: The specific process of step 2 is: Step 2.1: Use the squeeze connection module SPC to take the image collected in step 1 as input, divide the input image into S parts, and use the SIFT method to extract multi-scale features for the spatial information on the channel feature map after each segmentation to obtain a multi-scale feature map F i ; Step 2.2: Based on the multi-scale feature map F extracted in step 2.1 i , use the SEWeight module to extract the channel attention of the multi-scale feature map and obtain the channel attention weight Z at each different scale i and attention vector Z; Step 2.3, based on the channel attention weights Z of different scales obtained in step 2.2 i , use Softmax to recalibrate and get new attention weights atti ; Step 2.4, get the new attention weight att from step 2.3 i And the feature map F of the corresponding scale extracted in step 2.1 i Perform dot multiplication on each element and output the feature map Y after multi-scale feature information attention weighting i ; Step 2.5: The feature map Y obtained in step 2.4 i Perform dimension splicing and output multi-scale information feature map Out.

4. The method for detecting a safety belt for electric power aerial work according to claim 3 is characterized in that: The specific process of step 3 is: replacing the bottleneck 3x3 convolution in the convolutional residual network ResNet with the PSA module built in step 2 to obtain the EPSANET network model.

5. The method for detecting a safety belt for electric power aerial work according to claim 4, characterized in that: The specific process of step 4 is: replacing the backbone network in the YOLOV5 model with the EPSANET network model constructed in step 3 to obtain the EPSA-YOLOV5 network model.

6. The method for detecting a safety belt for electric power aerial work according to claim 5, characterized in that: The specific process of step 5 is: improving the non-maximum suppression NMS in the EPSA-YOLOV5 network model, that is, improving the confidence function, the formula is as follows: Among them, Si is the confidence of the i-th prediction box, M is the current highest scoring box, bi is the prediction box to be processed, IOU is the intersection and union ratio of M and bi, and N is the set threshold.

Images