Reinforced concrete defect detection and risk early warning method based on lightweight YOLOv8 network

Abstract

The invention belongs to the technical field of computer vision, and relates to a reinforced concrete defect detection and risk early warning method based on a lightweight YOLOv8 network. According to the lightweight RC-YOLOv8 network model provided by the invention, through the innovative design of the adaptive convolution module, the dual convolution module and the convolution block attention module, the parameter quantity of the model is significantly reduced, and the detection precision and generalization ability of fine defects such as cracks and exposed bars are significantly improved, so that the model is more suitable for the actual deployment of resource-constrained equipment, and the method has good application prospects. And the application range in an actual engineering scene is further expanded. And meanwhile, the model shows higher detection precision and stronger robustness under a complex background condition, and the situations of false alarm and missing detection are effectively reduced. According to the risk scoring and early warning mechanism designed by the invention, by defining a scientific and reasonable risk scoring formula, the severity of the structural defect can be rapidly and accurately evaluated, and clear early warning management and response measures can be taken for different risk levels.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a reinforced concrete defect detection and risk warning method based on a lightweight YOLOv8 network. Background Art

[0002] With the rapid development of global infrastructure, the importance of project quality management in the construction sector has become increasingly prominent. Reinforced concrete, in particular, is widely used in various building structures due to its exceptional durability and structural strength. However, in actual use, reinforced concrete is often affected by complex environmental factors, resulting in defects such as cracks and exposed rebar. These defects not only affect structural stability but also increase the cost and difficulty of subsequent maintenance. Traditional manual inspection methods, due to their high labor and time costs, are increasingly unable to meet the needs of large-scale structural inspections.

[0003] In recent years, automated detection methods based on deep learning technology have been gradually applied to the field of reinforced concrete defect detection. For example, the YOLO series of models, due to its efficient single-step regression analysis method, can directly predict target boundaries and category probabilities from image pixels, thereby significantly improving detection efficiency. However, existing YOLO models are generally complex in structure and have a large number of model parameters, making it difficult to run efficiently on resource-constrained devices. In addition, these models perform poorly when detecting complex backgrounds and small defects, and have problems with accuracy and robustness. For example, the lightweight model based on YOLOv5s proposed by Hao et al. improved detection speed, but the detection accuracy in complex environments still needs to be improved; the YOLOv5-DE model proposed by Ma et al. showed good results for crack detection, but its generalization ability was insufficient.

[0004] At the same time, most current research on reinforced concrete defect detection remains at the detection and identification stage, with limited research on how to use detection results for risk assessment and early warning management. Research by Mucci et al. indicates that combining AI technology to achieve a closed loop from detection to risk management has important application value for improving engineering safety management. However, existing research has failed to effectively achieve this closed loop, lacks a systematic risk scoring and early warning mechanism, and is unable to meet the actual needs of engineering management. Therefore, there is an urgent need to develop a lightweight, high-precision reinforced concrete defect detection model and to construct a risk scoring and early warning system to achieve comprehensive automated management from defect identification to risk management. Summary of the Invention

[0005] The present invention aims to overcome the shortcomings of the existing technology and proposes an efficient, lightweight and accurate reinforced concrete defect detection and risk warning method. Specifically, it includes an improved YOLOv8 network model (RC-YOLOv8) and a corresponding risk scoring and warning mechanism to achieve intelligent closed-loop management of structural health monitoring.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A reinforced concrete defect detection and risk warning method based on a lightweight YOLOv8 network, the steps are as follows:

[0008] Step (1) Construction of lightweight reinforced concrete defect detection model RC-YOLOv8

[0009] (1.1) Introducing the Adaptive Convolution Module (AKConv): Traditional convolution layers use fixed-size convolution kernels, which are difficult to effectively capture defect features of different scales. This invention designs an AKConv module to adaptively adjust the size and shape of the convolution kernel based on the local information of the input feature map, thereby improving the model's sensitivity to defects such as cracks and exposed rebar. The adaptive adjustment of the convolution kernel is expressed as follows:

[0010]

[0011] Among them, Y i,j is the output after convolution, X i+m,j+n Represents the pixel value of the input feature map at the corresponding position, m and n represent the index position of the convolution kernel along the row and column directions on the input feature map, respectively, and W m,n is the weight of the convolution kernel, and K is the size of the convolution kernel.

[0012] (1.2) Design of the DualConv module: The DualConv module introduces the concept of multi-scale feature extraction, aiming to enhance the detection of defects at different scales. When dealing with concrete defects, the scale of the defects may vary significantly, and traditional convolutional networks have difficulty effectively handling these changes. The DualConv module uses multiple convolution kernels in parallel to extract feature information at different scales, and then fuses them into a unified output feature map, thus avoiding the limitations of single-scale feature extraction.

[0013] Assume that the input feature map is in, is a set of real numbers, H and W are the height and width of the feature map, and C is the number of channels. After convolution operations of different scales, multiple feature maps Z1, Z2, ..., Z n , where each Z iCorresponding to a convolution operation of different scales. The convolution operation of each scale can be expressed as:

[0014] Z i =X*W i (2)

[0015] Among them, * represents the convolution operation, W i is the convolution kernel of the i-th scale, i∈{1,2,…,n} represents the convolution kernel of different scales. After the multi-scale convolution operation, the DualConv module fuses the feature maps of these different scales to obtain the final output feature map Z final :

[0016]

[0017] Among them, α i is a learned fusion coefficient used to adjust the contribution of features at different scales. In this way, DualConv can effectively integrate information at different scales and improve the detection capability of defects of various sizes;

[0018] (1.3) Integrate Convolutional Block Attention Module (CBAM): Add CBAM modules in the early and middle stages of the network to enhance the model's ability to identify key defect areas through channel attention and spatial attention mechanisms; specifically, assume that the input feature map H and W are the height and width of the feature map, respectively, and C is the number of channels. The channel attention module first extracts two independent feature descriptions through global average pooling (GAP) and global maximum pooling (GMP). and

[0019]

[0020] Among them, F ij Represents the value of the input feature map at channel C, row i, column j, represents the global average value of the C-th channel feature, Indicates the maximum value of the C-th channel feature.

[0021] The two feature descriptions are then processed by a shared multi-layer perceptron (MLP), which consists of two fully connected layers with the middle layer dimension set to C / r (r is the reduction ratio, usually 16) to reduce the number of parameters and control the model complexity. The output of the channel attention is and The weighted sum of the features constitutes M c (F);

[0022]

[0023] Among them, σ represents the Sigmoid activation function, which is used to generate the attention weight of the channel.

[0024] (1.4) RC-YOLOv8 Overall Network Structure: To further enhance the network's ability to detect concrete defects of varying sizes, this paper proposes a lightweight, high-precision concrete defect detection network architecture based on the YOLOv8n network. While maintaining the existing efficient detection framework, the overall network strategically incorporates AKConv, CBAM, and DualConv modules to address the needs of different feature extraction stages, enabling precise extraction of fine-grained features and effective fusion of multi-scale information.

[0025] The network first inputs an image of concrete defects. This image is processed by the AKConv module, which extracts preliminary features, performs downsampling, and outputs a feature map. Subsequently, the AKConv module is applied again to the feature maps corresponding to Level 1, Level 3, Level 5, and Level 7 of YOLOv8 to achieve preliminary feature extraction.

[0026] After completing the initial feature extraction, the network enters the mid-layer feature enhancement stage. The first CBAM module is inserted at Level 16 to perform dual channel and spatial attention enhancement on the mid-layer features. The CBAM module extracts important area information by applying global average pooling (GAP) and global maximum pooling (GMP) operations to the feature map, generates a channel attention weight map after processing by a shared multi-layer perceptron (MLP), and performs weighted processing on the input feature map. Subsequently, the feature map undergoes further convolution and downsampling. The second CBAM module is introduced again at Level 20 to further improve the ability to focus on fine-grained information and output an enhanced feature map of the same size.

[0027] At the higher-level network layer, the DualConv module is introduced at Level 21 to enhance the ability to fuse multi-scale features. The DualConv module extracts multi-scale feature information by using multiple convolution kernels with different receptive field sizes in parallel, generating several sub-feature maps. These sub-feature maps are then weighted and fused to create a unified output feature map.

[0028] Finally, after feature upsampling, splicing and further convolution processing, the network extracts multi-scale fusion feature maps and outputs them in the Detect module, realizing the identification of defects such as concrete cracks and exposed rebar.

[0029] (1.5) Model training: A concrete defect dataset containing annotations of cracks and exposed rebar defects was used for model training. This dataset provides a wealth of annotated images and is suitable for model training, validation, and testing. In order to enhance the robustness of the model, field data collection was conducted to record images of concrete defects under different environmental conditions. These field images come from multiple construction projects and cover cracks and exposed rebar defects under different lighting and weather conditions, increasing the diversity of the dataset. All images were manually annotated by professionals, and the annotations included the specific location and type of cracks and exposed rebar defects, and a cross-validation method was used to ensure the consistency of the annotations. The LabelImg tool was used for annotation, and the annotation information included the coordinates and type of the defect area.

[0030] Step (2) Design of reinforced concrete defect risk scoring and early warning mechanism

[0031] (2.1) Risk score calculation method: By setting risk thresholds for different defect types such as cracks and exposed rebar, the risk level of each defect can be clarified. For example, the alarm level can be set based on the defect size, distribution density, and severity level (such as length, width, and depth). Based on the defect situation in each inspection image, a formula for calculating the risk score is defined. A weighted scoring method can be used to assign different weights to different types of defects to calculate the overall risk score. The specific calculation formula is as follows;

[0032]

[0033] Among them, L i is the weight of the defect, Defect i is the corresponding defect detection result. The higher the score, the greater the structural safety risk;

[0034] (2.2) Risk level and early warning trigger mechanism: Based on actual project requirements and expert recommendations, risk level thresholds are set. A score of over 80 is considered high risk, triggering an emergency warning immediately; a score of 50-80 is considered moderate risk, with regular inspections recommended; a score below 50 is considered low risk, with only record keeping and long-term monitoring required.

[0035] In the risk scoring and early warning triggering mechanism of the present invention, the scoring levels are determined based on discussions among multiple building safety experts to ensure the scientific and practical nature of the scoring. Specifically, when the overall risk score exceeds 80 points, the score reflects that the structure may have major safety hazards and is therefore classified as high risk. The system will immediately generate an emergency warning, requiring the engineering management team to respond quickly, immediately dispatch personnel to the defect location for on-site assessment, and take reinforcement or sealing measures to curb further expansion of the risk. For defects with risk scores between 50 and 80 points, which are defined as medium risk, the system will generate early warning recommendations for regular inspections and convey the information to the engineering management team, prompting them to arrange routine inspections to closely monitor risk changes in the area. If the score is less than 50 points, it means that the defect poses a small direct threat to the structure. Such defects are recorded and left for subsequent evaluation.

[0036] Beneficial effects of the present invention:

[0037] 1. The lightweight RC-YOLOv8 network model proposed in this paper significantly reduces the model's parameter count by approximately 300,000 through the innovative design of the Adaptive Convolution Module (AKConv), the Dual Convolution Module (DualConv), and the Convolutional Block Attention Module (CBAM). This significantly improves the detection accuracy and generalization capability of subtle defects such as cracks and exposed rebar, making the model more suitable for deployment in resource-constrained equipment and further broadening its applicability in real-world engineering scenarios. Furthermore, the model demonstrates higher detection accuracy and greater robustness under complex background conditions, effectively reducing false positives and missed detections.

[0038] 2. The risk scoring and early warning mechanism designed in this invention, by defining a scientific and rational risk scoring formula, can quickly and accurately assess the severity of structural defects and implement clear early warning management and response measures for different risk levels. When the risk score exceeds the set threshold, the system automatically generates an emergency warning and transmits the defect location and detailed characteristics to the project management team in real time, helping managers to formulate a timely response plan. This greatly improves the response efficiency and scientific decision-making of project safety management, significantly reduces the risk of potential structural safety accidents, and comprehensively enhances the intelligent level of project management. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic diagram of the overall architecture of the RC-YOLOv8 network of the present invention;

[0040] Figure 2 Schematic diagram of the AKConv adaptive convolution module structure;

[0041] Figure 3 Schematic diagram of the DualConv dual convolution module structure;

[0042] Figure 4 Schematic diagram of the CBAM convolutional block attention mechanism module;

[0043] Figure 5 The performance comparison of different networks in detecting cracks and exposed rebar is shown in the figure. DETAILED DESCRIPTION

[0044] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0045] In this embodiment, the present invention first constructs a concrete defect dataset containing 1,600 images. The image data covers cracks and exposed reinforcement defects under different lighting and environmental conditions, and uses data enhancement technology to improve data diversity.

[0046] Secondly, the PyTorch framework is used to build a lightweight RC-YOLOv8 network model (such as Figure 1 As shown). The AKConv module is introduced in the initial stage ( Figure 2 ), automatically adjust the convolution kernel size to improve the defect feature extraction capability; introduce the DualConv module in the late stage of the network ( Figure 3 ) to achieve multi-scale feature fusion; the CBAM module is introduced in the front end of the network ( Figure 4 ), enhance the expression of key defect characteristics. The details are as follows:

[0047] The network first inputs a 640×640×3 concrete defect image. This image is processed by the AKConv module, which extracts preliminary features and performs downsampling, outputting a feature map of size 320×320×64. Subsequently, the AKConv module is applied again to the 320×320×64, 160×160×128, 80×80×256, and 40×40×512 feature maps at Levels 1, 3, 5, and 7, respectively, to progressively extract higher-level and richer feature information. Spatial downsampling is performed each time, ensuring that the feature map size is halved and the number of channels is doubled. The AKConv module introduces an adaptive convolution mechanism, allowing the feature extraction process at each layer to dynamically adjust the shape and size of the convolution kernel based on the specific characteristics of the input local area, significantly improving the ability to capture subtle defect features such as cracks and exposed rebar.

[0048] After completing the initial feature extraction, the network enters the mid-level feature enhancement stage. To further strengthen the model's focus on key defect areas, the first CBAM module is inserted at Level 16 (i.e., the 80×80×256 feature map) to perform both channel-wise and spatial attention enhancement on mid-level features. The CBAM module extracts important region information by applying global average pooling (GAP) and global maximum pooling (GMP) operations to the feature map. After processing the feature map using a shared multi-layer perceptron (MLP), a channel-wise attention weight map is generated and weighted on the input feature map. The enhanced output feature map remains at 80×80×256. Subsequently, the feature map undergoes further convolution and downsampling, reducing its size to 40×40×512. A second CBAM module is introduced at Level 20 (40×40×512 feature map) to further enhance the focus on fine-grained information and output an enhanced feature map of the same size.

[0049] At the higher-level network level, the DualConv module is introduced at Level 21 (40×40×512 feature map) to enhance the ability to fuse multi-scale features. The DualConv module uses multiple convolution kernels with different receptive field sizes in parallel to extract multi-scale feature information, generating several sub-feature maps. These sub-feature maps are then weighted and fused to produce a unified output feature map. The output feature map processed by the DualConv module remains 40×40×512, but incorporates a rich set of features at different scales, improving the model's robustness in detecting concrete defects of varying sizes (particularly variations in crack length and exposed rebar area).

[0050] Ultimately, after feature upsampling, concatenation, and further convolution, the network extracts multi-scale fused feature maps at three scales (80×80, 40×40, and 20×20). These are then output by the Detect module, achieving high-precision identification of defects such as concrete cracks and exposed rebar. By rationally incorporating AKConv, CBAM, and DualConv modules, the proposed detection network significantly improves detection accuracy and robustness while maintaining its lightweight nature, making it particularly suitable for fine-grained defect detection in complex backgrounds.

[0051] In this example, the dataset contains 1,600 images, which are divided into 70% for training, 15% for validation, and 15% for testing. All images are normalized to 640x640 pixels and resized before being fed into the network to ensure data consistency and accuracy.

[0052] The model was trained using the SGD optimization algorithm, with an initial learning rate of 0.01, a momentum of 0.937, a weight decay of 0.0005, a batch size of 4, and 300 iterations. After training, the model parameters were reduced by approximately 300,000, and compared to the traditional model, the mean average prediction accuracy (mAP) increased to 78.4% and the accuracy reached 88.1%.

[0053] To comprehensively evaluate the RC-YOLOv8 object detection model, we used several key metrics, including precision, recall, F1 score, mean average precision (mAP), intersection over union (IOU), and frames per second (FPS). The following is a detailed description of these metrics and their calculation formulas, illustrating the model's performance in different areas:

[0054] Precision measures the proportion of samples predicted by the model as positive that are actually positive, and is an important indicator for evaluating model accuracy. Its calculation formula is:

[0055]

[0056] Among them, TP (True Positives) represents true positive examples, that is, the number of positive examples correctly predicted by the model, and FP (False Positives) represents false positive examples, that is, the number of positive examples incorrectly predicted by the model.

[0057] Recall shows the proportion of the model's correct identification among all actual positive classes, focusing on the model's coverage ability. Its calculation formula is:

[0058]

[0059] Among them, FN (False Negatives) represents false negative examples, that is, the number of actual positive classes not detected by the model.

[0060] The F1 score is the harmonic mean of precision and recall, and is used to evaluate the balance between these two metrics. Its calculation formula is:

[0061]

[0062] mAP provides a measure of the average precision of the model under different categories and different thresholds. This paper focuses on the two indicators mAP@0.5 and mAP@0.5:0.95. Among them, AP refers to the average accuracy of the prediction results calculated by the model for each category under different categories. mAP refers to the average AP calculated over all categories. It provides a comprehensive performance evaluation of the model across all categories. The calculation formulas for AP and mAP are:

[0063]

[0064] FPS reflects the speed at which the model processes images and is a very important indicator for evaluating model performance, especially in real-time applications. Its calculation formula is:

[0065]

[0066] Among them, t1 is the image preprocessing time; t2 is the inference time, which refers to the time from the preprocessed image input to the model to the result; t3 is the conversion time of the model output.

[0067] When evaluating model performance, direct comparison with other classic object detection models can more clearly demonstrate the model's advantages. We systematically compare the OURS model with several common object detection models (YOLOv8n, YOLOv7, and YOLOv5).

[0068] In the crack detection task, the OURS model demonstrates superior performance compared to YOLOv8n. As shown in Table 1, in crack classification detection, the OURS model achieves a precision of 81.7%, higher than the 76.0% of the YOLOv8n model; a recall of 74.7%, higher than the 68.8% of the YOLOv8n model; and a mAP@0.5 score of 72.2%, nearly 4 percentage points higher than the YOLOv8n model; and a mAP@0.5:0.95 score of 44.7%, better than the 43.8% of the YOLOv8n model. This demonstrates that the OURS model has higher accuracy and robustness in crack detection, and is able to more effectively capture the characteristics of small cracks.

[0069] The OURS model also achieved a clear advantage in the exposed rebar detection task. Specifically, the OURS model achieved a precision of 94.6% in the exposed rebar category, significantly higher than the YOLOv8n model's 81.6%; the recall rate was 79.1%, higher than the YOLOv8n model's 74.0%; the mAP@0.5 score was 84.5%, an increase of over 10 percentage points from the YOLOv8n model's 74.1%; and the mAP@0.5:0.95 score reached 59.5%, significantly higher than the YOLOv8n model's 49.0%. This demonstrates that the OURS model demonstrates superior detection accuracy and stability when identifying complex defects such as exposed rebar.

[0070] In terms of risk scoring, as shown in Table 2, for exposed rebar defects, the two parameters of exposed rebar area and rebar depth are examined. In the defect case No. 1, the exposed area is 85cm 2 , the depth is 8cm, and the final weighted score is 46.5 points; in the defect case No. 2, the area is 65cm 2, the depth is 6cm, and the final score is 35.5 points; in the defect case No. 3, the area is 45cm 2 , with a depth of 4cm and a final score of 24.5 points. Through weighted calculation, the risk level of defects can be accurately quantified, providing data support for subsequent risk scoring and early warning mechanisms.

[0071] In summary, by comparing with the existing classic detection model system, the OURS model proposed in this invention has obvious advantages in key indicators such as precision, recall rate, and mAP, especially in complex environments and subtle defect detection, showing stronger robustness and accuracy, which fully verifies the effectiveness and engineering application value of the technical solution of this invention.

[0072] Table 1 Concrete crack and exposed reinforcement detection results

[0073]

[0074] Table 2 Calculation results of three-dimensional quantitative indicators of concrete spalling and exposed reinforcement

[0075]

Claims

1. A reinforced concrete defect detection and risk warning method based on a lightweight YOLOv8 network, characterized in that: Here are the steps: Step (1) Construction of lightweight reinforced concrete defect detection model RC-YOLOv8 (1.1) Introducing the adaptive convolution module AKConv: The AKConv module adaptively adjusts the size and shape of the convolution kernel based on the local information of the input feature map. The adaptive adjustment of the convolution kernel is expressed as follows: Among them, Y i,j is the output after convolution, X i+m,j+n Represents the pixel value of the input feature map at the corresponding position, m and n represent the index position of the convolution kernel along the row and column directions on the input feature map, respectively, and W m,n is the weight of the convolution kernel, K is the size of the convolution kernel; (1.2) Design of the DualConv module: The DualConv module uses multiple convolution kernels in parallel to extract feature information at different scales and then fuses them into a unified output feature map. The details are as follows: Assume that the input feature map is in, is a set of real numbers, H and W are the height and width of the feature map respectively, and C is the number of channels; after convolution operations of different scales, multiple feature maps Z1, Z2, ..., Z n , where each Z i Corresponding to a convolution operation of different scales; the convolution operation of each scale is expressed as: Z i =X*W i (2) Among them, * represents the convolution operation, W i is the convolution kernel of the i-th scale, i∈{1,2,…,n} represents the convolution kernel of different scales; after the multi-scale convolution operation, the DualConv module fuses the feature maps of these different scales to obtain the final output feature map Z final : Among them, α i is a learned fusion coefficient; (1.3) Integrate the convolutional block attention module CBAM: The CBAM module enhances the model's ability to identify key defect areas through channel attention and spatial attention mechanisms; specifically, let the input feature map H and W are the height and width of the feature map respectively, C is the number of channels, and the channel attention module first extracts two independent feature descriptions through global average pooling GAP and global maximum pooling GMP and Among them, F ij Represents the value of the input feature map at channel C, row i, column j, represents the global average value of the C-th channel feature, Represents the maximum value of the C-th channel feature; The two feature descriptions are then processed by a shared multi-layer perceptron (MLP), which consists of two fully connected layers with the middle layer dimension set to C / r, where r is the reduction ratio; the output of the channel attention is given by and The weighted sum of the features constitutes M c (F); Among them, σ represents the Sigmoid activation function, which is used to generate the attention weight of the channel; (1.4) The overall network structure of RC-YOLOv8: First, the network input is a concrete defect image, which is processed by the AKConv module to extract preliminary features and perform downsampling, outputting a feature map. Then, the AKConv module is applied again to the feature maps corresponding to Level 1, Level 3, Level 5, and Level 7 of YOLOv8 to achieve preliminary feature extraction. After completing the initial feature extraction, the network enters the mid-layer feature enhancement stage. The first CBAM module is inserted at Level 16 to perform channel-wise and spatial dual attention enhancement on the mid-layer features. The CBAM module extracts important region information by applying global average pooling and global maximum pooling operations to the feature map. After processing it with a shared multi-layer perceptron, a channel attention weight map is generated and the input feature map is weighted. Subsequently, the feature map undergoes further convolution and downsampling. The second CBAM module is introduced again at Level 20 to further improve the ability to focus on fine-grained information and output an enhanced feature map of the same size. At the high-level network stage, the DualConv module is introduced at Level 21 to enhance the multi-scale feature fusion capability. The DualConv module extracts multi-scale feature information by using multiple convolution kernels with different receptive field sizes in parallel, obtaining several sub-feature maps of different scales. These sub-feature maps are then weighted and fused to generate a unified output feature map. Finally, after feature upsampling, splicing, and further convolution processing, the network extracts a multi-scale fusion feature map, which is then output in the Detect module to identify concrete defects. (1.5) Model training: The model is trained using a concrete defect dataset containing annotations of cracks and exposed rebar defects. Step (2) Design of reinforced concrete defect risk scoring and early warning mechanism (2.1) Risk score calculation method: By setting risk thresholds for different defect types, such as cracks and exposed rebar, the risk level of each defect is clarified. A weighted scoring method is used to assign different weights to different defect types, thereby calculating the overall risk score. The specific calculation formula is as follows; Among them, L i is the weight of the defect, Defect i is the corresponding defect detection result; (2.2) Risk level and warning trigger mechanism: Set risk level thresholds. A score of more than 80 points is considered high risk, and an emergency warning is triggered immediately; a score of 50-80 points is considered medium risk, and regular inspections are recommended; a score below 50 points is considered low risk, and only records are kept and monitored over the long term.

Images