A multi-component fusion identification method for safety fence defects

Through artificial intelligence technology, local fusion algorithms and fault classification networks are used to realize multi-component fusion identification of security fences, solving the security threat problems caused by the lack of monitoring in traditional technologies, improving the identification accuracy and positioning accuracy of security fences, and ensuring the protection of security fences Construction safety has been improved and intelligent management of infrastructure construction sites.

CN113902908BActive Publication Date: 2024-11-12GUANGZHOU POWER SUPPLY BUREAU GUANGDONG POWER GRID CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202111188154.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-10-12
Publication Date
2024-11-12
Estimated Expiration
2041-10-12

AI Technical Summary

Technical Problem

The lack of safety fence status monitoring technology in the traditional infrastructure industry, which leads to timely detection of equipment defects or failures, threatens the safety of construction personnel and causes economic losses.

Method used

A multi-component fusion recognition method for security fences based on artificial intelligence is designed, and a defect positioning prediction model and fault classification network are constructed through local fusion algorithms to achieve accurate identification and positioning of security fences.

Benefits of technology

The functional safety of safety fences has been improved, the accuracy of defect positioning and identification speed have been improved, construction safety has been ensured and the intelligent management level of infrastructure construction sites has been improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113902908B_ABST
    Figure CN113902908B_ABST
Patent Text Reader

Abstract

The present invention discloses a multi-component fusion identification method for safety fence defects. The method includes: collecting safety fence images of infrastructure construction sites, making corresponding sample label files for each image and forming a training set; constructing a safety fence defect positioning prediction model and a safety fence defect recognition model; inputting the training set into the safety fence defect positioning prediction model, and the outputs of the training set and the safety fence defect positioning prediction model are jointly input into the safety fence defect recognition model, and training is performed simultaneously to obtain a trained safety fence defect positioning prediction model and a safety fence defect recognition model; inputting the safety fence image to be detected into the trained safety fence defect positioning prediction model and the safety fence defect recognition model to obtain the defect recognition result of the safety fence. The present invention improves the positioning speed of safety fence defects, improves the detection accuracy of safety fence defects, ensures the construction safety of construction site personnel, and improves the intelligent management level of infrastructure construction sites.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a safety fence defect recognition method, and in particular to a multi-component fusion recognition method for safety fence defects. Background Art

[0002] Various types of equipment in the traditional infrastructure industry, such as safety fences, are closely related to safe construction. However, due to the lack of corresponding status monitoring technology, once these devices have defects or failures, it will lead to very serious safety accidents, threaten the lives of construction workers, and cause significant economic losses.

[0003] Smart infrastructure based on computer vision technology can realize intelligent identification of defect status of various safety equipment at the infrastructure site, detect and replace faulty equipment in advance, and ensure the safety of life and property. Summary of the invention

[0004] In order to solve the problem of locating and identifying safety fence defects, the present invention proposes a multi-component fusion identification method for full fence defects. For defects such as broken, damaged, and missing safety fences, a safety fence defect locating prediction model based on a local fusion algorithm is designed. The output result of the locating prediction model is input into a safety fence defect recognition model based on a fault classification network to complete the accurate identification of safety fence defects.

[0005] The present invention aims at the common safety fences at infrastructure construction sites and designs a safety fence authority positioning and identification method based on artificial intelligence technology. While ensuring the identification speed, it increases the defect identification accuracy, thereby improving the functional safety of the safety fence and creating a new smart infrastructure.

[0006] The technical solution of the present invention adopts the following steps:

[0007] 1) Collect safety fence images of infrastructure construction sites, create corresponding sample label files for each image and form a training set;

[0008] 2) Construct a safety fence defect location prediction model based on local fusion algorithm;

[0009] 3) Construct a safety fence defect recognition model based on fault classification network;

[0010] 4) Inputting the training set into the safety fence defect location prediction model, the output of the training set and the safety fence defect location prediction model are jointly input into the safety fence defect recognition model, the safety fence defect location prediction model and the safety fence defect recognition model are trained simultaneously, and the trained safety fence defect location prediction model and the safety fence defect recognition model are obtained;

[0011] 5) The safety fence image to be detected is input into the trained safety fence defect location prediction model and the safety fence defect recognition model, and the trained safety fence defect recognition model outputs the defect recognition result of the safety fence.

[0012] The safety fence image in step 1) contains multiple components, specifically the anti-fall bracket, fence chain lock, safety fence bottom roller, warning sign and metal fence.

[0013] The safety fence defect location prediction model in step 2) satisfies the following formula:

[0014]

[0015] Where L is the output of the safety fence defect location prediction model, N c is the number of cross components, satisfying N c =N·(N-1) / 2; L c is the component cross entropy, p i Score the component category, For p i The true value, N r is the number of smoothing components, satisfying N r =N,L r is the smoothing loss, q i is the border regression value, for q i The true value, S is the area of ​​the candidate box, G i is the i-th truth box of the safety fence image, A is the candidate box, and N is the number of parts in the safety fence image.

[0016] The safety fence defect recognition model K in step 3) satisfies the following formula:

[0017] K=sign(L)·ResNet(50,L,DCGAN(L,A))

[0018] Among them, ResNet() represents the ResNet classification network, L is the output of the safety fence defect positioning prediction model, A is the candidate box, sign(·) represents the symbolic function operation, and DCGAN() represents the deep convolutional adversarial generation network.

[0019] In the step 2), the COCO dataset is first used to pre-train the safety fence defect location prediction model to obtain the initial parameters of the model.

[0020] In the step 3), the ResNet classification network in the safety fence defect recognition model is first pre-trained on the ImageNet dataset to obtain the initial parameters of the ResNet classification network.

[0021] The beneficial effects of the present invention are:

[0022] 1. The present invention improves the accuracy of safety fence defect positioning by designing a safety fence defect positioning prediction model, enhances the robustness and anti-interference of safety fence defect positioning under various interferences, and plays an important role in timely discovering the defect location of the safety fence and organizing the maintenance of the safety fence.

[0023] 2. The present invention increases the recognition speed and accuracy of safety fence defects by designing a safety fence defect recognition model, which is of great value for timely identifying the types of safety fence defects and formulating maintenance and replacement plans, ensuring the construction safety of construction site personnel and improving the intelligent management level of infrastructure construction sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0025] The specific implementation modes of the present invention are described in detail below.

[0026] like Figure 1 As shown, the present invention comprises the following steps:

[0027] 1) Collect safety fence images of infrastructure construction sites, create corresponding sample label files for each image and form a training set; the label files meet the Pascal VOC format xml label file standard. The label file contains the image name, image path, image height and width, and the center point position and width and height of the real target box.

[0028] The safety fence image in step 1) includes multiple components, specifically anti-fall brackets, fence chain locks, warning signs and metal fences.

[0029] 2) Construct a safety fence defect location prediction model based on local fusion algorithm;

[0030] The safety fence defect location prediction model in step 2) satisfies the following formula:

[0031]

[0032] Where L is the output of the safety fence defect location prediction model, N c is the number of cross components, satisfying N c =N·(N-1) / 2; the cross component is to cross between components, L c is the component cross entropy, p i Score the component category, For p i The true value, N ris the number of smooth parts, mainly the rollers at the bottom of the safety fence, which meets N r =N,L r is the smoothing loss, q i is the border regression value, Q i The true value, S is the area of ​​the candidate box, G i is the i-th truth box of the safety fence image, A is the candidate box, and N is the number of parts in the safety fence image.

[0033] In step 2), the COCO dataset is first used to pre-train the safety fence defect location prediction model to obtain the initial parameters of the model.

[0034] 3) Construct a safety fence defect recognition model based on fault classification network;

[0035] The safety fence defect recognition model K in step 3) satisfies the following formula:

[0036] K=sign(L)·ResNet(50,L,DCGAN(L,A))

[0037] Among them, ResNet() represents the ResNet classification network (the principle of ResNet can be found in the paper “Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, Deep Residual Learning for Image Recognition, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016”), L is the output of the safety fence defect positioning prediction model, A is the candidate box, sign(·) represents the symbolic function operation, and DCGAN() represents the deep convolutional generative adversarial network (the principle of DCGAN can be found in the paper “Alec Radford, Luke Metz, and Soumith Chintala, Unsupervised representation learning with deep convolutional generative adversarial networks, In ICLR, 2016”).

[0038] In step 3), the ResNet classification network in the safety fence defect recognition model is pre-trained on the ImageNet dataset to obtain the initial parameters of the ResNet classification network.

[0039] 4) Inputting the training set into the safety fence defect location prediction model, the output of the training set and the safety fence defect location prediction model are jointly input into the safety fence defect recognition model, the safety fence defect location prediction model and the safety fence defect recognition model are trained simultaneously, and the trained safety fence defect location prediction model and the safety fence defect recognition model are obtained;

[0040] 5) The safety fence image to be detected is input into the trained safety fence defect location prediction model and the safety fence defect recognition model, and the trained safety fence defect recognition model outputs the defect recognition result of the safety fence.

[0041] In the specific implementation, when the safety fence defect location prediction model is trained using the safety fence image dataset, the model training optimizer uses stochastic gradient descent with an initial learning rate of 0.002. The initial learning rate is used in the first 20,000 iterations, and the learning rate is reduced to 0.08 times the current learning rate at the 20,000th and 50,000th iterations, respectively. The total number of model training iterations is 90,000.

[0042] The ResNet classification network in the safety fence defect recognition model is first pre-trained on the ImageNet dataset, and the pre-training results are used as the initial parameters of the model. Then, DCGAN is used to expand the training of the defective images of the safety fence. When using the safety fence image dataset for training, the model training optimizer uses stochastic gradient descent, with an initial learning rate of 0.002. It is stipulated that one round of training samples is one training cycle. The learning rate of the first five training cycles remains the initial learning rate, and the learning rate of the sixth and 16th rounds is reduced to 0.1 times the current learning rate, and a total of 30 rounds of training are carried out.

[0043] The above specific implementation modes are used to explain the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A multi-component fusion identification method for safety fence defects, characterized in that: The following steps are involved: 1) Collect safety fence images of infrastructure construction sites, create corresponding sample label files for each image and form a training set; 2) Construct a safety fence defect location prediction model based on local fusion algorithm; 3) Construct a safety fence defect recognition model based on fault classification network; 4) Inputting the training set into the safety fence defect location prediction model, the output of the training set and the safety fence defect location prediction model are jointly input into the safety fence defect recognition model, the safety fence defect location prediction model and the safety fence defect recognition model are trained simultaneously, and the trained safety fence defect location prediction model and the safety fence defect recognition model are obtained; 5) Inputting the safety fence image to be detected into the trained safety fence defect location prediction model and the safety fence defect recognition model, the trained safety fence defect recognition model outputs the defect recognition result of the safety fence; The safety fence defect location prediction model in step 2) satisfies the following formula: Where L is the output of the safety fence defect location prediction model, N c is the number of cross components, satisfying N c =N·(N-1) / 2; L c is the component cross entropy, p i Score the component category, For p i The true value, N r is the number of smoothing components, satisfying N r =N,L r is the smoothing loss, q i is the border regression value, Q i The true value, S is the area of ​​the candidate box, G i is the i-th truth box of the safety fence image, A is the candidate box, and N is the number of parts in the safety fence image.

2. The multi-component fusion identification method for safety fence defects according to claim 1 is characterized in that: The safety fence image in step 1) contains multiple components, specifically the anti-fall bracket, fence chain lock, safety fence bottom roller, warning sign and metal fence.

3. The multi-component fusion identification method for safety fence defects according to claim 1 is characterized in that: The safety fence defect recognition model K in step 3) satisfies the following formula: K=sign(L)·ResNet(50,L,DCGAN(L,A)) Among them, ResNet() represents the ResNet classification network, L is the output of the safety fence defect positioning prediction model, A is the candidate box, sign(·) represents the symbolic function operation, and DCGAN() represents the deep convolutional adversarial generation network.

4. The multi-component fusion identification method for safety fence defects according to claim 1 is characterized in that: In the step 2), the COCO dataset is first used to pre-train the safety fence defect location prediction model to obtain the initial parameters of the model.

5. The multi-component fusion identification method for safety fence defects according to claim 1 is characterized in that: In the step 3), the ResNet classification network in the safety fence defect recognition model is first pre-trained on the ImageNet dataset to obtain the initial parameters of the ResNet classification network.

Citation Information

Patent Citations

  • Method and device for automatically identifying defects of insulator of high voltage transmission tower

    CN108648169A

  • Part defect detection method and system

    CN109934821A

  • Power transmission line iron tower structure defect detection method and device

    CN110675387A

  • Tower defect identification method based on unmanned aerial vehicle inspection

    CN111311569A

  • Power transmission line fitting target detection and defect identification method and terminal

    CN111881937A