ConstructFaultDeect-based model for steel beam incremental launching construction quality detection and construction method and detection method of ConstructFaultDeect-based model for steel beam incremental launching construction quality detection
By applying a model based on ConstructFaultDetect in steel beam top push construction, image features are extracted and fused, defect monitoring problems during construction are solved, and high-accurate defect detection and construction quality monitoring are achieved.
Patent Information
- Application Number
- CN202510214630.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-26
- Publication Date
- 2025-06-13
AI Technical Summary
During the steel beam top push construction process, it is difficult for the existing technology to effectively monitor the risk factors and construction quality during the construction process, resulting in poor conditions of damage to the project quality.
A model based on ConstructFaultDetect is proposed, including a backbone network module, a neck module and a prediction head module, which is used to extract features from the input image and perform object detection to identify defects in steel beam over-push construction.
By extracting multi-level features and fusing different scale features, the model can accurately detect construction defects of steel beams with different sizes, improving the accuracy and efficiency of construction quality monitoring.
Smart Images
Figure CN120147862A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of intelligent construction safety management, and particularly to a model for quality inspection of steel beam jacking construction based on ConstructFaultDetect, and a construction method and an inspection method thereof. Background Art
[0002] In recent years, with the continuous expansion and improvement of the domestic transportation network, the intersections between existing railway lines and newly built highways, railways, and municipal roads have become increasingly frequent. Among them, the steel beam jacking construction technology is applicable to various scenarios in highway construction. Its main feature is that it can construct bridges without interrupting traffic or with less restrictions. For example, when building a bridge above an existing highway or railway, jacking construction can avoid long-term traffic interruption and reduce the impact on existing traffic; or in the city center or building-intensive areas, due to limited space, the jacking method can complete the construction of long-span bridges within a limited space. During the jacking construction process, it is easy to occur that the quality of the project is damaged due to the failure to timely monitor the risk factors during the construction process and the operating status of the operating line.
[0003] The advantages of machine vision are reflected in its automation level and information integration ability, in occasions with requirements. The automated and intelligent detection technology based on machine vision has been successfully applied to roads and tunnels, and has also been preliminarily applied to bridges, but mainly focuses on the acquisition technology of the apparent images of high-altitude concrete components with open views, and the automatic recognition of quality and safety issues still remains in the theoretical research stage. During the current steel beam jacking construction process, the means of safety and quality management are single, and there are easy visual blind spots in monitoring, making it difficult to effectively manage the safety and quality of bridges. Therefore, it is of great significance to propose a quality monitoring method for the steel beam jacking construction process based on machine vision recognition algorithm to ensure the quality and safety of bridge construction. Summary of the Invention
[0004] Aiming at the problems existing in the prior art, the present invention proposes a model for quality inspection of steel beam jacking construction based on ConstructFaultDetect. The model includes: a backbone network module (Backbone) for extracting features in the input image; a neck module (Neck) located after the backbone network for enhancing feature expression or fusing multi-scale features; and a prediction head module (Prediction) for receiving the feature map output from the neck module and then performing the target detection task.
[0005] Based on the above solution, the backbone network module includes a CBL module, a Res1 unit, a Res2 unit, a Res8 unit, a Res8 unit, and a Res4 unit connected in sequence. Among them, the image input into the backbone network module is marked as A, the output of the first Res8 unit is marked as B, the output of the second Res8 unit is marked as C, and the output of the Res4 unit is marked as D.
[0006] Based on the above solution, B, C, and D output by the backbone network module will all be used as inputs to the neck module. First, D is input into a series of CBL blocks connected by 5 CBLs, and its output is marked as E; E will be used as an output of the neck module;
[0007] E will also be input into a CBL module, followed by an upsampling module to obtain H; H and the output C of the backbone network module are subjected to a vector concatenation operation to obtain J; J then passes through 5 consecutive CBL modules to obtain L; L will be used as an output of the neck module;
[0008] L will also be input into a CBL module to obtain G, followed by an upsampling module to obtain I; I and the output B of the backbone network module are subjected to a vector concatenation operation to obtain K; K then passes through 5 consecutive CBL modules to obtain M; M will be used as an output of the neck module layer.
[0009] Based on the above solution, the prediction head module receives three outputs from the neck module, namely E, L, and M; among them, E passes through a CBL module and a convolutional module to obtain a feature map O for prediction with a size of 16x16x255;
[0010] L also passes through a CBL module and a convolutional (Conv) module to obtain a feature map P for prediction with a size of 32x32x255;
[0011] M also passes through a CBL module and a convolutional (Conv) module to obtain a feature map Q for prediction with a size of 64x64x255.
[0012] Based on the above solution, the CBL module is composed of a convolutional layer, a batch normalization layer, and a LeakyRelu activation function connected in sequence.
[0013] Based on the above solution, Res1, Res2, Res4, and Res8 all use ResUnit units (residual units), and the numbers 1, 2, 4, and 8 represent the number of residual units connected in series.
[0014] This application also provides a method for constructing the above model, including the following steps:
[0015] (1) Data collection:
[0016] Collect a large amount of image data during the jacking process of steel beams, including normal conditions and situations with various types of known defects, ensuring that the data covers different lighting conditions, angles, and environmental changes;
[0017] (2) Label the data:
[0018] For the collected images, use LabelMe for detailed annotation, and mark the positions and types of all existing construction defects with rectangular boxes;
[0019] (3) Model training
[0020] Initialize the model parameters in a randomized manner, and use the labeled dataset to train the selected model.
[0021] This application also provides a method for detecting the construction quality of steel beam jacking, using the above-mentioned model.
[0022] A server includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-mentioned method for detecting the construction quality of steel beam jacking.
[0023] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the steps of the method for detecting the construction quality of steel beam jacking as described above.
[0024] The backbone network module of the model of the present invention can extract multi-level features from the input image. These features contain different abstraction levels of the image, from low-level edges and textures to high-level semantic information. The backbone network can learn a certain degree of scale invariance, which is very important for detecting defect targets of steel beam jacking construction of different sizes. In addition, the neck module of the present invention can fuse different levels of features extracted by the backbone network module to make full use of the detailed information of low-level features and the semantic information of high-level features. The neck module can enhance the ability of feature representation, help the model better adapt to targets of different scales, and by adaptively adjusting the scale of the feature map, enable the model to make predictions at multiple scales. At the same time, the prediction head module of the present invention is responsible for classifying and positioning the features output by the neck module. The generated candidate boxes may contain multiple overlapping boxes. By non-maximum suppression (NMS), redundant boxes can be removed, and the most likely prediction results can be retained, making the detection of defect targets of steel beam jacking construction of different sizes very accurate.
[0025] In the model of the present invention, ResUnit units of Res1, Res2, Res4, and Res8 are adopted. ResUnit units with different scales can extract information on the defects of the steel beam quality in different dimensions, making the prediction results more accurate. The ResUnit is directly connected by two CBL modules and undergoes a residual addition, which enables the model to extract local features of the defects in the incremental launching construction of the steel beam. By using the convolutional kernel to learn information such as the texture and edges of the image, the data distribution becomes more stable, introducing non-linearity and enhancing the expression ability of the model, enabling the model to accurately detect multi-scale targets of the defects in the incremental launching construction of the steel beam. The DUS upsampling adopted in the model of the present invention uses local deformation attention feature upsampling, which enables the prediction model to not only aggregate the features of uniformly sampled neighborhood points, which may cause the model to overemphasize less important background regions, but to focus more on the feature extraction and recognition of the target itself. Especially in the pictures of the quality defects in the incremental launching construction of the beam in this example, it can highlight the extraction of subtle features and improve the accuracy. Description of the Drawings
[0026] Figure 1 It is the architecture diagram of the model of this application;
[0027] Figure 2 It is the structure diagram of the CBL module in the model of this application;
[0028] Figure 3 It is the structure diagram of the ResUnit in the model of this application;
[0029] Figure 4 It is the schematic diagram of the DUS module in the model of this application;
[0030] Figure 5 It is an example diagram of typical defects in the quality of incremental launching construction of the beam (scratches on the surface of the steel beam);
[0031] Figure 6 It is an example diagram of typical defects in the quality of incremental launching construction of the beam (welding slag inclusions);
[0032] Figure 7 It is an example diagram of typical defects in the quality of incremental launching construction of the beam (rail deformation);
[0033] Figure 8 It is an example diagram of typical defects in the quality of incremental launching construction of the beam (coating bubbles). Detailed Embodiment
[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the following will clearly and completely describe the technical solutions in the embodiments of this application with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, rather than all of them. Usually, the components of the embodiments of this application described and shown in the accompanying drawings here can be arranged and designed in various different configurations.
[0035] Embodiment 1
[0036] The construction process of pushing the steel beam includes:
[0037] 1. Install embedded parts on the capping beam: The embedded parts on the capping beam are installed during the construction of the capping beam to fix the construction support. The possible quality problem is that the embedded parts are not firmly fixed. If the embedded parts are not firmly fixed in the capping beam, they may shift during the pushing process, leading to construction accidents.
[0038] 2. Install the support and guide rail: The flatness problem of the support and guide rail. If the surfaces of the support and guide rail are not flat, the steel beam may shift during the pushing process.
[0039] 3. Transport the steel beam: Deformation: During transportation, the steel beam may deform due to vibration, collision, or improper fixing methods. Damage: During transportation, scratches, dents, or other damages may occur on the surface of the steel beam due to improper loading and unloading operations or poor road conditions.
[0040] 4. Welding: Welding defects: Welding defects such as porosity, slag inclusions, cracks, and lack of fusion may occur during the welding process, affecting the structural strength of the steel beam.
[0041] 5. Push the steel beam across the span: Steel beam deformation: During the pushing process, the steel beam may deform due to uneven stress, unstable support, etc. Welding joint problems: During the pushing process, cracks or other damages may occur in the welding joints due to stress concentration.
[0042] 6. Transverse shift the steel beam to the designed position, install the transverse shift track, and transverse shift the steel beam to the designed position: Improper design of the track structure, poor construction quality, poor synchronization control, etc. will all cause the track to deform or be distorted and damaged during the transverse shift process.
[0043] 7. Install the steel bridge diaphragm: If the welding process is improper or the welding quality is unqualified, weld defects such as cracks, porosity, and slag inclusions may occur, affecting the connection strength of the diaphragm. If the diaphragm material does not meet the design requirements, such as material corrosion or deformation, it will affect the structural safety.
[0044] 8. Paint the steel beam: If there are defects such as bubbles, sagging, orange peel, and pinholes in the coating, it will affect the anti-corrosion performance of the coating.
[0045] 9. The cantilever end of the bridge deck is constructed using suspended formwork: Formwork deformation: During the construction process, the suspended formwork may deform due to its own weight, construction loads, or external forces, affecting the geometric shape and dimensional accuracy of the bridge deck.
[0046] Concrete pouring problems: The construction using suspended formwork may lead to uneven concrete pouring, resulting in quality problems such as honeycombing, pockmarks, and cracks.
[0047] Figures 5 - 8 Four typical defects in the incremental launching construction quality of the beam are shown.
[0048] The present invention provides a model based on ConstructFaultDetect for the quality inspection of the incremental launching construction of steel beams, as well as its construction method and inspection method.
[0049] In the process of incremental launching of steel beams, early detection of construction quality defects using the model can be achieved through the following steps:
[0050] 1. Data collection:
[0051] Collect a large amount of image data during the incremental launching of steel beams, including normal conditions and cases with various known types of defects. Ensure that the data covers different lighting conditions, angles, and environmental changes.
[0052] 2. Data annotation:
[0053] For the collected images, use labelme for detailed annotation, and mark the positions and types of all existing construction defects with rectangular frames. The types are as follows:
[0054] 1) The embedded parts are not firmly fixed in the capping beam
[0055] 2) The surfaces of the brackets and guide rails are uneven
[0056] 3) The steel beam is deformed
[0057] 4) Scratches, depressions, or other damages appear on the surface of the steel beam
[0058] 5) Welding defects such as porosity, slag inclusions, cracks, and lack of fusion during the welding process
[0059] 6) The track is deformed or distorted and damaged during the transverse movement
[0060] 7) The material of the diaphragm is corroded or deformed
[0061] 8) Defects such as bubbles, runs, orange peel, and pinholes exist in the coating of the steel beam painting
[0062] 9) The formwork of the suspended formwork is deformed
[0063] 10) The concrete pouring is uneven, resulting in honeycombing, pockmarks, and cracks
[0064] In this example, a total of 2,000 photos were collected from a road and bridge construction site in Shandong. Using a ratio of 3:1:1, the data was divided into a training set (1,200 photos), a test set (400 photos), and a validation set (400 photos).
[0065] 3. Model architecture selection and initialization:
[0066] The object detection model selected in this invention is the ConstructFaultDetect model, and the model parameters are initialized in a randomized manner.
[0067] 4. Training the model:
[0068] The selected model is trained using the labeled dataset. Ensure there are sufficient computing resources and hyperparameters can be adjusted to optimize the model performance.
[0069] Define the loss function, which consists of two parts:
[0070] Classification loss (cross-entropy loss): Used for predicting the target class.
[0071] Regression loss (Smooth L1 loss): Used for the regression of bounding boxes.
[0072] The optimizer SGD is selected. In the forward propagation, the model receives the input image and outputs the prediction results (class and bounding box), then calculates the loss, computes the gradients through backpropagation, and updates the model parameters using the optimizer. After 500 epochs, stop the training and save the model with the best performance on the validation set.
[0073] 5. Model validation and testing:
[0074] Evaluate the performance of the model on the validation set to ensure it can accurately identify various defects.
[0075] 6. Deployment and integration:
[0076] Deploy the trained model to the monitoring system at the construction site. This may involve edge computing devices or cloud servers. In this invention, the model is deployed to a cloud server.
[0077] 7. Real-time monitoring and warning:
[0078] Once the model is successfully deployed, it will be able to process the video stream from the cameras at the construction site in real time, automatically detect and classify any emerging defects. When potential quality issues are detected, the system should immediately issue an alarm for timely corrective measures.
[0079] As a specific implementation, the overall architecture of the ConstructFaultDetect model is as follows Figure 1 shown, and the size of the input image of the model is 512x512. Specifically, the ConstructFaultDetect model consists of three parts: (1) Backbone network module: The main purpose of the backbone network is to extract features from the input image. It is usually composed of a series of convolutional layers, which are responsible for capturing the spatial hierarchical information of the image, such as edges, textures, shapes, etc., and gradually constructing high-level semantic features. (2) Neck module: The neck is located after the backbone network and before the prediction head, and is used to enhance feature expression or fuse multi-scale features. (3) Prediction head module: The prediction head receives the feature map from the neck, and then applies a series of convolutional layers and activation functions to perform the final object detection task. Its task is to generate bounding box coordinates, object confidence scores, and class probabilities. Specifically, each grid cell will predict multiple bounding boxes and their corresponding class distributions. The prediction head will make predictions at multiple scales to adapt to objects of different sizes.
[0080] Specifically, at the input end, the size of the input image is 512x512, and 3 represents the three RGB channels. The input image is denoted as A and enters the backbone network module (Backbone). The backbone network module (Backbone) is connected by CBL modules, Res1 units, Res2 units, Res8 units, Res8 units, and Res4 units. The output of the first Res8 unit is denoted as B, the output of the second Res8 unit is denoted as C, and the output of the Res4 unit is denoted as D.
[0081] As Figure 2 shown, where CBL represents the connection of a convolutional function (Conv) + batch normalization (BN) + activation function (LeakyRelu), that is: Conv represents the convolutional function, and BN is batch normalization (Batch Normalization). LeakyReLU (Leaky Rectified Linear Unit) is an improved version of the ReLU (Rectified Linear Unit) activation function used for non-linear transformation in neural networks.
[0082] As Figure 1 shown, Res1, Res2, Res4, and Res8 all use ResUnit units (residual units), and the numbers 1, 2, 4, and 8 represent the number of residual units connected in series.
[0083] As Figure 3As shown, the ResUnit is first directly connected by two CBL modules, and the input is directly added to the output vector of the second CBL module to obtain the output of the ResUnit.
[0084] The outputs B, C, and D of the backbone network module (Backbone) will all be used as inputs to the neck module (Neck). First, the D input passes through a CBL series block connected by 5 CBLs, and its output is denoted as E. E will be used as an output of the neck module (Neck).
[0085] E also inputs to a CBL module, followed by an upsampling module (DUS), to obtain H. H and the output C of the backbone network module (Backbone) are concatenated (Concat) to obtain J, and J then passes through 5 consecutive CBL modules to obtain L. L will be used as an output of the neck module (Neck).
[0086] L also inputs to a CBL module to obtain G, followed by an upsampling module (DUS), to obtain I. I and the output B of the backbone network module (Backbone) are concatenated (Concat) to obtain K, and K then passes through 5 consecutive CBL modules to obtain M. M will be used as an output of the neck module layer (Neck).
[0087] Thus, the neck module (Neck) has three outputs, namely E, L, and M. E passes through a CBL module and a convolution (Conv) module to obtain O, which is a 16x16x255 feature map for prediction; L also passes through a CBL module and a convolution (Conv) module to obtain P, which is a 32x32x255 feature map for prediction; M also passes through a CBL module and a convolution (Conv) module to obtain Q, which is a 64x64x255 feature map for prediction.
[0088] It adopts an architecture consisting of a Backbone network module, a Neck module, and a Prediction head module. The Backbone network module is responsible for extracting features from the input image. The Backbone network module can generate feature maps at different levels, and these feature maps contain information from low-level (details) to high-level (semantics). The Neck module is used to fuse the feature maps from different layers of the Backbone network module, so that the high resolution of the low-level features and the strong semantic information of the high-level features can be utilized simultaneously. By fusing features at different scales, the Neck module helps the model detect targets at different scales, which is very important for detecting objects of different sizes. The Backbone network module layer is responsible for object localization (bounding box prediction) and classification (class prediction) based on the fused feature maps. The design of the ConstructFaultDetect model allows the Backbone network module, the Neck module, and the Prediction head module layers to be trained end-to-end together, which enables the entire network to be jointly optimized and improve the detection performance. The design of the Backbone network module, the Neck module, and the Prediction head module in the ConstructFaultDetect model enables the model to accurately detect multi-scale targets in images while maintaining high speed and efficiency.
[0089] Specifically, the implementation of the Dual-Up-Sampling (DUS) module is as follows:
[0090] For the input feature map and the upsampling factor α ∈ [1, +∞], the output feature map can be obtained through feature upsampling At the beginning, the query (Q), key (K), and value (V) of the input feature map can be obtained through linear mapping as follows:
[0091] (Q, K, V) = (XW Q , XW K , XW V )
[0092] Assume that the kernel size of neighborhood sampling is k u = 3. As Figure 4 shown, let p = (x, y) be the coordinates of the point to be interpolated, where x ∈ [0, W - 1], y ∈ [0, H - 1]. Taking the grid arrangement format of aligned corner points as an example, the coordinates of p' = (x', y') can be obtained through the following formula:
[0093]
[0094] Where, W represents the width, H represents the height, α represents the upsampling factor, and p = (x, y) is the coordinate of the point to be interpolated.
[0095] Let r = {(x 1 , y 1 ), (x 2 , y 1 ), …, (x 3 , y 3 )} represent the absolute coordinates of the unified adjacent points of p'. Therefore, the upsampling result of point p based on local self-attention can be expressed as:
[0096]
[0097] Where Q, K, and V represent Query, Key, and Value respectively, p ′ represents Figure 4 the coordinate p' = (x', y') in, r represents the absolute coordinates of the unified adjacent points of p', and both s and t represent the points in the set r.
[0098] Figure 1 In, the prediction head module obtains three outputs O, P, and Q respectively. The features with different resolutions of the output will be segmented into a fixed number of grid cells. Usually, these grids are evenly distributed. Each grid cell is responsible for detecting objects in the image. In each grid cell, the model predicts multiple bounding boxes, and each bounding box is defined by a set of parameters, including the position of the bounding box (center coordinates and width and height), the confidence of the bounding box (representing the probability of containing an object and the accuracy of the bounding box), and a class probability distribution (representing the class of the object within the bounding box). First, filter out those bounding boxes that are unlikely to contain objects according to the confidence threshold. To handle the problem that the same object is detected by multiple grid cells, the model of the present application uses non-maximum suppression (NMS) to merge overlapping bounding boxes. The NMS algorithm retains the bounding box with the highest confidence and suppresses other bounding boxes that have a high overlap with it. After NMS processing, the remaining bounding boxes are considered as the final detection results, which contain the class and position information of the objects. These steps enable it to quickly and accurately detect multiple objects in the image without complex post-processing.
[0099] To verify the effectiveness of the model, by using the picture set of the construction quality of the beam jacking construction of Shandong Road and Bridge as the data set for training, excellent results are obtained. The present invention uses the following evaluation indicators for evaluation: Precision, Recall, and mAP.
[0100] The Intersection over Union (IoU) is an index used in object detection algorithms to evaluate the similarity between two rectangular boxes. IoU = the area of intersection of the two rectangular boxes / the area of union of the two rectangular boxes. TP, TN, FP, and FN are the abbreviations for true positive, true negative, false positive, and false negative respectively. Positive and negative refer to the results obtained from prediction. If the prediction is for the positive class, it is a positive example; if the prediction is for the negative class, it is a negative example. True and false indicate whether the predicted result is the same as the actual result. If they are the same, it is true; if they are different, it is false.
[0101] Precision - that is, the precision rate, the formula is as follows:
[0102] Recall - that is, the recall rate, the formula is as follows:
[0103] mAP is an indicator that can be used to measure whether the predicted box categories and positions of the model are accurate. Average Precision (AP) is the result of evaluating the detection quality of each class. Assuming that the IoU value is greater than a pre-set threshold (usually set to 0.5), it means that this predicted box is correct, and at this time this box is a TP; assuming that the IoU value is less than the pre-set threshold (usually set to 0.5), it means that this predicted box is wrong, and at this time this box is an FP. Taking the average of all APs gives mAP.
[0104] Through experiments, this model was compared with common object detection models, and the results are shown in Table 1:
[0105] Table 1 Recognition test results of different models
[0106]
[0107] As can be seen from Table 1, when using the ConstructFaultDetect model for training and prediction, it can particularly achieve very good results in defect recognition of the quality of beam launching construction, with a 1.91% improvement in precision, a 3.2% improvement in recall rate, and a 3.53% improvement in mAP_0.5:0.95.
[0108] It should be noted that, without conflict, the features in the embodiments of this application can be combined with each other.
[0109] The above are only the preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various modifications and variations can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.
Claims
1. A model for steel beam jacking construction quality detection based on ConstructFaultDetect, characterized in that: The model includes Backbone network module for extracting features from input images; Located after the backbone network, the neck module (Neck) is used to enhance feature expression or fuse multi-scale features: The prediction head module (Prediction) is used to receive the feature map output from the neck module and then perform the target detection task.
2. The model for steel beam jacking construction quality detection based on ConstructFaultDetect according to claim 1 is characterized in that: The backbone network module includes a CBL module, a Res1 unit, a Res2 unit, a Res8 unit, a Res8 unit and a Res4 unit connected in sequence, wherein the image input to the backbone network module is marked as A, the output of the first Res8 unit is marked as B, the output of the second Res8 unit is marked as C, and the output of the Res4 unit is marked as D.
3. The model for steel beam jacking construction quality detection based on ConstructFaultDetect according to claim 2 is characterized in that: The outputs of the backbone network module, B, C, and D, will be used as the inputs of the neck module. First, D is input into the CBL series block composed of 5 CBLs, and its output is recorded as E; E will be used as an output of the neck module; E is also input into a CBL module, followed by an upsampling module to get H; H and the output C of the backbone network module are concatenated to get J; J then passes through 5 consecutive CBL modules to get L; L is used as an output of the neck module; L is also input into a CBL module to obtain G, followed by an upsampling module to obtain I; I is concatenated with the output B of the backbone network module to obtain K; K then passes through 5 consecutive CBL modules to obtain M; M will be used as an output of the neck module layer.
4. The model for steel beam jacking construction quality detection based on ConstructFaultDetect according to claim 2 is characterized in that: The prediction head module receives three outputs from the neck module, namely E, L and M. Among them, E passes through a CBL module and a convolution module to obtain a 16x16x255 feature map O for prediction; L also passes through a CBL module and a convolution (Conv) module to obtain a 32x32x255 feature map P for prediction; M also passes through a CBL module and a convolution (Conv) module to obtain a 64x64x255 feature map Q for prediction.
5. The model for steel beam jacking construction quality detection based on ConstructFaultDetect according to claim 2 is characterized in that: The CBL module consists of a convolutional layer, a batch normalization layer and a LeakyRelu activation function connected in sequence.
6. The model for steel beam jacking construction quality detection based on ConstructFaultDetect according to claim 5 is characterized in that: Res1, Res2, Res4 and Res8 all use ResUnit units (residual units), and the numbers 1, 2, 4, and 8 represent the number of residual units connected in series.
7. A method for constructing the model according to any one of claims 1 to 6, characterized in that: The steps include: (1) Data collection: Collect a large amount of image data during the steel beam pushing process, including normal conditions and conditions where various types of defects are known to exist, ensuring that the data covers different lighting conditions, angles, and environmental changes; (2) Annotated data: The collected images are annotated in detail using LabelMe, and the locations and types of all existing construction defects are marked with rectangular boxes; (3) Model training Initialize the model parameters in a random manner and train the selected model using the labeled dataset.
8. A method for quality inspection of steel beam jacking construction, characterized in that: A model constructed using the model described in any one of claims 1 to 6 or the method of claim 7.
9. A server comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method for quality inspection of steel beam jacking construction as described in claim 8 are implemented.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, the steps of the method for quality inspection of steel beam jacking construction as claimed in claim 8 are implemented.