Steel box girder welding transverse linear precision detection method and system based on deep convolution

Abstract

The present application relates to the technical field of structure detection, and particularly relates to a steel box girder welding transverse linear precision detection method and system based on deep convolution, comprising the following steps: obtaining a post-welding section image of a target steel box girder assembly welding section, and inputting the convolution deep network to identify a curvature release stage corresponding to the post-welding section image; outputting a stage curvature representation unit corresponding to the post-welding section image; based on the stage curvature representation unit, performing stage correction on a transition deformation profile in the post-welding section image; and obtaining a stable curvature profile of the target steel box girder assembly welding section. According to the present application, the stage curvature representation unit is used to analyze the transverse displacement vector field and the curvature distribution information, the reverse displacement correction is performed on the transition profile, the spline fitting is combined with the curvature constraint, the stable curvature profile is reconstructed, the conversion from the instantaneous observation curvature to the real stable curvature is realized, and the transition state misjudgment is avoided.

Description

Technical Field

[0001] This invention relates to the field of structural inspection technology, and in particular to a method and system for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution. Background Technology

[0002] Steel box girders are widely used in bridge construction due to their high load-bearing capacity, high overall stiffness, and suitability for long-span bridges. During the manufacturing and cantilever assembly of steel box girders, the lateral alignment accuracy after assembly and welding directly affects the joint quality of beam segments, the overall stress state, and the subsequent installation accuracy. In existing projects, to ensure the lateral alignment of steel box girders, mechanical adjustment devices are usually used for remedial correction during assembly or welding. For example, external jacking pressure is applied to the beam segments to be erected or the welded components using bidirectional jacks, vertical jacks, pads, jacking brackets, or other special tools to adjust their lateral deformation or joint misalignment, thereby improving the alignment matching accuracy.

[0003] However, the aforementioned mechanical adjustment measures are usually required to be implemented when the component is not yet fully stable and still has an adjustable margin. If the transverse alignment deviation is found to be out of tolerance only after the weld has completely cooled, the residual stress has been basically released, and the component shape has been basically solidified, the difficulty of correction by jacking, using jacks or tooling increases significantly, and the best window for remediation may even be missed. Therefore, the transverse alignment inspection of steel box girders after welding should not only determine whether the deviation is out of tolerance, but also predict the trend of stable alignment as early as possible during the post-weld cooling and shape release process, so as to reserve time for subsequent mechanical adjustment or process remediation.

[0004] Existing detection methods mostly use visual measurement, laser scanning, or point cloud fitting to obtain the post-weld cross-sectional profile and directly calculate curvature or linear deviation based on the current observed profile. However, after the steel box girder is assembled and welded, the cross-sectional shape is not immediately stable. Instead, it is simultaneously affected by welding thermal shrinkage, residual stress release, delayed springback after clamp removal, and local constraint differences. The current observed profile is often in a transitional deformation state. If this transitional state is directly used as the final linear shape for evaluation, it is easy to misjudge the transient deviation as a true deviation, or fail to identify the subsequent stable deviation trend in a timely manner.

[0005] Therefore, there is an urgent need for a method to detect the transverse alignment accuracy of steel box girder assembly and welding. This method should be able to identify the curvature release stage of the cross section while the welded component is still within a detectable and remedial temperature window, and reconstruct a stable curvature profile based on the current transition deformation profile. This would allow the curvature distribution to characterize the transverse alignment accuracy, determine in advance whether the stable alignment meets the design requirements, and provide a reliable basis for subsequent remedial correction by mechanical adjustment devices. Summary of the Invention

[0006] This invention provides a method and system for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution.

[0007] The first aspect is a method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution, which includes the following steps: The post-weld cross-sectional image of the welded section of the target steel box girder assembly is obtained and input into a convolutional depth network to identify the curvature release stage corresponding to the post-weld cross-sectional image; Output the stage curvature characterization unit corresponding to the post-weld cross-sectional image; Based on the stage curvature characterization unit, the transition deformation profile in the post-weld cross-sectional image is corrected in stages. Obtain the stable curvature profile of the welded section of the target steel box girder assembly; The transverse curvature detection value of the target steel box girder assembly and welding section is extracted based on the stable curvature profile, and the transverse curvature detection value is compared with the preset design curvature to output the transverse curvature accuracy result.

[0008] Secondly, the steel box girder welding transverse alignment accuracy detection system based on depth convolution includes the following modules: The image acquisition module is used to acquire post-weld cross-sectional images of the welded sections of the target steel box girder assembly. The stage recognition module is used to input the post-weld cross-sectional image into the convolutional depth network and identify the corresponding curvature release stage. The characterization output module is used to output the stage curvature characterization unit corresponding to the post-weld cross-sectional image. The stage correction module is used to perform stage correction on the transition deformation profile in the post-weld cross-sectional image based on the stage curvature characterization unit. The curvature reconstruction module is used to obtain the stable curvature profile of the welded section of the target steel box girder assembly. The accuracy detection module is used to extract the lateral curvature detection value based on the stable curvature profile, compare the lateral curvature detection value with the preset design curvature, and output the lateral curvature accuracy result.

[0009] The beneficial effects of this invention are: This invention identifies the curvature release stage of the post-weld cross-section through a convolutional depth network and constructs a stage curvature characterization unit. It further analyzes and obtains the lateral displacement vector field and curvature distribution information, performs reverse compensation and stable curvature reconstruction on the transition deformation profile, and realizes the transformation from the instantaneous post-weld observation shape to the final stable line shape. It can effectively distinguish the deviation between the transition deformation and the true line shape, avoid misjudgment caused by transient deformation, and improve the accuracy and reliability of the lateral line shape detection results.

[0010] This invention eliminates the need to wait for the steel box girder section to fully stabilize before conducting inspections. Instead, it predicts and reconstructs a stable curvature profile based on the current observed profile, improving inspection efficiency while ensuring accuracy. This better aligns with the production cycle requirements of steel box girder manufacturing sites. Furthermore, by automatically establishing a mapping relationship between the post-weld morphology and the curvature release state through a convolutional depth network, it reduces reliance on manual experience and judgment, enhancing the automation, adaptability, and consistency between inspection results and the actual service condition of the structure. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the method flow of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the system modules in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the architecture of the convolutional depth network in Embodiment 1 of the present invention. Detailed Implementation

[0013] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0014] Example 1 Linearity is a general description of the geometric shape of the cross section or outer contour of a steel box girder. It reflects whether the contour is smooth, whether it deviates, and whether it conforms to the design shape. Curvature is a quantitative result of the degree of bending of the linearity at a local location. It is used to describe how much the linearity bends and how fast it changes. Therefore, linearity is the overall geometric expression, while curvature is a specific parameter that characterizes the local bending characteristics of the overall linearity. The two are essentially related as an overall form and a local quantitative index.

[0015] like Figure 1 and Figure 3 As shown, the method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution includes the following steps: S1: Obtain the post-weld cross-sectional image of the welded section of the target steel box girder assembly, input it into the convolutional depth network, identify the curvature release stage corresponding to the post-weld cross-sectional image, and output the stage curvature representation unit corresponding to the post-weld cross-sectional image.

[0016] S11, the target detection cross section is predefined as a cross section located at the corresponding position of the weld to be detected and perpendicular to the longitudinal reference axis of the steel box girder. The visible contour to be detected includes at least the intersection line of the cross section with the weld reinforcement surface and the surface to be detected of the flange plates on both sides. Multiple high-resolution industrial cameras are arranged laterally on a rigid support or gantry above the steel box girder assembly and welding station. Each camera adopts a different lateral position and tilt angle towards the same target detection cross section, so that the fields of view of adjacent cameras both overlap and form an effective parallax. The camera array covers the visible contour to be detected, avoiding the use of texture edges in a single two-dimensional image directly above to characterize the deformation of the flange plate surface.

[0017] Before formal data acquisition, calibration plates or control points with known physical coordinates are arranged near the target detection cross-section. The intrinsic and extrinsic parameters and lens distortion parameters of each industrial camera are calculated, and distortion correction is performed on the acquired images. For visible contour points falling within the target detection cross-section plane, the intersection of the pixel back-projection ray and the target cross-section plane is taken as the physical coordinates of the cross-section corresponding to the pixel, based on the camera's intrinsic and extrinsic parameters and the target cross-section plane equation. Verification control points that were not involved in parameter calculation are selected, and the root mean square deviation between the mapped physical coordinates and their known coordinates is taken as the calibration error. The calibration result is used only if the calibration error is not greater than the preset allowable measurement error; otherwise, recalibration is performed.

[0018] For flange plate height or other non-coplanar contour points, matching is performed using the same features in the overlapping fields of view of adjacent cameras. Based on the calibrated camera extrinsic parameters, multi-view triangulation is used to recover their three-dimensional physical coordinates. Then, the three-dimensional points are transformed into a unified steel box girder coordinate system. The tolerance-filtered point set is selected according to the longitudinal position of the target detection cross section and the preset cross section, and projected onto the local physical coordinate system of the cross section to obtain the cross section contour with actual length scale.

[0019] Then, synchronous imaging acquisition is performed on the welded section of the target steel box girder assembly. Synchronous imaging acquisition is achieved through unified trigger control. All industrial cameras are connected to the same timing control module. When the temperature monitoring module determines that the section temperature meets the preset threshold condition, a trigger signal is sent to each camera simultaneously, so that each camera completes exposure acquisition under the same time reference, thereby obtaining post-weld section images at the same time and from different perspectives. After acquisition, the image is first corrected using the distortion parameters of each camera, and then the multi-view data is unified to the same steel box girder coordinate system based on the external parameters and the physical coordinate mapping of the section. For contour points in the plane of the target section, the physical coordinates are obtained by using the intersection of the pixel back projection ray and the plane of the target section. For contour points with out-of-plane height, the three-dimensional coordinates are recovered by multi-view triangulation using the corresponding contour points in the overlapping field of view. After selecting the point set according to the target section position and the preset section tolerance, the points are projected onto the local coordinate system of the section to form a section imaging and contour result with physical scale, rather than just spatially stitching the top view texture edge.

[0020] The timing control module consists of a trigger determination unit, a master clock unit, a synchronous trigger distribution unit, an acquisition and feedback unit, and a timing management unit, wherein: The trigger determination unit is used to receive the temperature status signal output by the temperature monitoring module and determine whether the current target steel box girder assembly welding section meets the preset acquisition trigger conditions. When the conditions are met, it outputs an acquisition enable signal to the main clock unit. The master clock unit is used to provide a unified system time base and sampling cycle, and after receiving the acquisition enable signal, it locks the current acquisition cycle and generates the corresponding synchronization trigger time. The synchronous trigger distribution unit is used to simultaneously output a unified trigger pulse to each high-resolution industrial camera according to the synchronous trigger time, so as to control each industrial camera to complete exposure acquisition under the same time reference. The acquisition feedback unit is used to receive exposure completion signals, image reception status signals or abnormal status signals returned by each high-resolution industrial camera, and to determine whether the synchronous imaging acquisition has been effectively completed. The timing management unit is used to assign a unified batch number and timestamp to the current acquisition batch, and bind the batch number and timestamp to the post-weld cross-sectional images acquired by each industrial camera for identification in subsequent image alignment, stitching and curvature release stages.

[0021] After welding is completed, the current temperature of the welded section of the target steel box girder is acquired in real time. When the section temperature cools from high temperature to within the preset detectable temperature window, and does not fall below the lower limit of the image capture cutoff temperature, image acquisition is triggered. This triggers a high-resolution industrial camera array to simultaneously acquire images of the welded section. At this acquisition moment, the section may still be in the curvature release stage corresponding to elastic recovery, plastic compression, or residual thermal shrinkage. The preset detectable temperature window is represented as follows: ; in, For the welded cross section at time The current temperature; The upper limit of the temperature allowed for photography is set to avoid image contour instability caused by excessively high cross-sectional temperature and strong thermal disturbance. The lower limit of the temperature at which photography is to be taken, and The temperature is higher than the remedial correction cutoff temperature to ensure that the component is not completely stable when the data is collected, and to retain the operation window for mechanical adjustment or top pressure correction.

[0022] The upper limit of the allowable temperature for taking pictures was obtained through post-weld cooling test calibration. Cross-sectional images and contour data were continuously collected during the process of the weld temperature changing from high to low. When the contour extraction was unstable due to thermal radiation, surface reflection, air disturbance or thermal deformation, the corresponding temperature was not suitable as the acquisition temperature. When the temperature continued to cool until the contour edge could be stably identified and the contour change at adjacent sampling times entered the calculable range, this temperature was taken as the upper limit of the allowable temperature for taking pictures. The lower limit of the allowable temperature for taking photos is determined based on the remedial correction cutoff temperature. That is, it is calibrated in combination with the latest operable temperature required for on-site mechanical adjustment devices, jacking fixtures or jacks for correction. When the residual stress and post-weld morphology of the component are basically solidified after the component is below a certain temperature, and the difficulty of jacking or forced adjustment increases significantly, this temperature is the remedial correction cutoff temperature. On this basis, a certain safety margin is added, which is 5℃-30℃, preferably 10℃-15℃, to obtain the lower limit of the allowable temperature for taking photos.

[0023] It can determine the stable linear trend in advance before the welded section is completely cured, leaving a time window for subsequent correction measures and avoiding difficulties in remediation when the deviation is discovered after cooling is completed.

[0024] S12. The post-weld cross-sectional image is processed by grayscale conversion and edge enhancement preprocessing to obtain a transition deformation contour map, which is used to characterize the deformation characteristics of the weld reinforcement area and the flange plates on both sides.

[0025] Grayscale processing is represented as: ;in, For the post-weld cross-section image at pixel points grayscale value at that location , , These are the red, green, and blue channel components of that pixel, respectively. , , These are the grayscale weighting coefficients; Edge enhancement of a grayscale image is represented as: ;in, For pixels Edge enhancement value at the location, This is the result of horizontal gray-level difference. This is the result of gray-level difference in the vertical direction; Based on the edge enhancement value, a threshold is applied to filter the candidate edge points, which are represented as follows: ; in, The binary results for candidate edges, To determine the edge filtering threshold, the edge enhancement values ​​of all pixels in the edge enhancement results are statistically analyzed to obtain the grayscale change amplitude distribution. Then, the boundary value that can clearly distinguish between strong edges and weak noise is selected as the initial threshold. Subsequently, the threshold is verified using labeled weld reinforcement edge and flange deformation edge samples. The threshold is then adjusted to ensure that the real structural edges are completely preserved, while eliminating as many false edges as possible caused by reflection, noise, and surface texture. The final determined threshold is the edge filtering threshold, used to distinguish between real structural edges and weak noise edges. Combining the structural screening masks of the weld reinforcement area and the flange area, the transition deformation profile is shown as follows: ; in, This is a transitional deformation profile diagram. The structure screening mask is used to retain the effective contour points of the weld reinforcement and the corresponding areas of the two side flanges, and to remove background noise, reflective edges and irrelevant textures. Based on the design geometry of the steel box girder section, the cross-sectional coordinate system after camera calibration and the weld position, the effective detection area in the image is defined, and the weld reinforcement area, the left flange area and the right flange area are initially determined. Then, combined with the edge enhancement results, edge points located in these areas with continuity, length and direction characteristics are retained, while the background area, tooling fixture area, reflective noise area and irrelevant texture area are set to invalid. Finally, the effective structure area is assigned a value of 1 and the non-structure area is assigned a value of 0, which is the structure screening mask.

[0026] The grayscale variation amplitudes in the edge enhancement results are statistically analyzed, and their overall distribution is calculated. Amplitudes exceeding a certain quantile are used as a threshold for initial pixel screening, retaining edge points with larger variation amplitudes as candidate contour points. This process eliminates most false edges generated by noise or weak textures. Subsequently, connectivity analysis is performed on the retained candidate edge points, aggregating spatially adjacent and directionally continuous edge points to form several candidate edge curves. After obtaining the candidate edge curves, further screening based on geometric location and structural features is required: edge curves located in the middle of the cross-section, exhibiting obvious convexity or thickness variation characteristics, typically correspond to weld reinforcement contours, while those distributed... The edge curves on both sides of the cross section, which have a gentle overall trend but exhibit bending or offset, correspond to the flange deformation profile. By judging the spatial position, length, curvature change trend, and continuity of the curves, discrete edges or short noise segments that do not conform to the structural characteristics are eliminated, and only the main profiles that can stably describe the deformation of the weld area and flange area are retained. The retained structural profiles are smoothed and the breakpoints are repaired to make the profiles spatially continuous without obvious jagged edges or jumps. Finally, a complete, continuous set of profiles with clear physical meaning is obtained, which constitutes the transition deformation profile diagram, used to characterize the weld reinforcement and the overall deformation characteristics of the flange in the post-weld state.

[0027] S13, the convolutional deep network consists of an input layer, a shared convolutional feature extraction backbone, a staged parallel discriminant branch, a staged curvature representation output layer, and a parameter decoding interface. The input layer receives the transition deformation contour map obtained through grayscale conversion, edge enhancement, and contour extraction, and performs size unification, grayscale normalization, and channel normalization on it. The shared convolutional feature extraction backbone consists of multi-level convolutional layers, batch normalization layers, nonlinear activation layers, and pooling layers, used to progressively extract multi-scale features such as weld reinforcement boundaries, local flange bending, overall section offset, local springback marks, and residual thermal shrinkage morphology, outputting a shared feature tensor. After the shared feature tensor, parameters are set to... The network has three parallel output branches, corresponding to the elastic recovery stage, the plastic compression stage, and the thermal contraction stage, respectively. Each branch includes a stage feature convolutional layer, a feature compression layer, and a fully connected discriminant layer to output the response value corresponding to that stage. The three response values ​​are normalized to form a curvature release stage probability distribution, and the stage corresponding to the highest probability is taken as the current curvature release stage. At the same time, the hidden layer feature tensor before the final discriminant layer is extracted from the selected stage branch, and then fully connected to generate stage curvature representation units. This enables the network to not only determine the current stage, but also output staged features for subsequent analysis of the lateral displacement vector field and local curvature radius distribution coefficients.

[0028] The convolutional deep network is constructed in the following order: input specification determination, backbone network construction, stage branch configuration, representation output connection, and output rule setting. First, based on the resolution of the images acquired by the industrial camera and the size of the transition deformation contour map, the network input size is determined, and the grayscale range, contour width, and coordinate normalization method of the input image are specified. Second, a shared convolutional feature extraction backbone is constructed, enabling it to progressively extract high-level morphological features from low-level edge features. For example, low-level convolutions are used to identify contour boundaries and local bulges in weld seam excess height; mid-level convolutions are used to identify flange bending trends and local compression morphology; and high-level convolutions are used for… The system identifies overall springback, residual thermal shrinkage, and cross-sectional shape release patterns. Three parallel stage branches are then connected to the output of the backbone network, each corresponding to a curvature release stage. Independent weight parameters are set for different branches, enabling them to extract more sensitive discriminative features from the same shared features, focusing on elastic recovery, plastic compression, and thermal shrinkage. Finally, the response values ​​output by each branch are converted into stage probabilities. Simultaneously, a stage curvature representation output layer is set within the corresponding stage branch, compressing the hidden features of that stage into fixed-dimensional stage curvature representation units, which serve as input for subsequent stage correction and stable curvature contour reconstruction.

[0029] The training of the convolutional deep network consists of five parts: sample acquisition, label construction, network forward computation, error backpropagation, and model solidification. First, post-weld cross-sectional images of steel box girders with different specifications, welding processes, post-weld times, and clamp release states are acquired, and corresponding transition deformation profiles are extracted. Simultaneously, welding completion time, cooling temperature curves, clamp release time, profile change sequence, and final stable profile are recorded. Then, based on temperature change trends, profile change rates, springback performance after clamp release, and final stable form, the samples are labeled as elastic recovery stage, plastic compression stage, or thermal contraction stage. Furthermore, based on the difference between the transition profile and the stable profile, a curvature representation for the supervision stage is established. The displacement compensation relationship and curvature distribution relationship of the unit are considered. During training, the transition deformation contour map is input into the network. The classification results of the three parallel branches are constrained by the stage label, so that the network can accurately identify the curvature release stage. At the same time, the lateral displacement vector field and local curvature radius distribution information obtained by back-inferring the stable contour are used to assist in constraining the stage curvature representation unit, so that it not only has the ability to classify and distinguish, but also has the ability to express the deformation compensation characteristics of the current stage. Through multiple rounds of iterative training, the weight parameters and bias terms of the convolutional layer, branch layer and fully connected layer are continuously updated until the stage recognition error, displacement compensation error and curvature distribution error all meet the preset convergence conditions. Finally, a convolutional depth network that can be used for actual detection is obtained.

[0030] When constructing training labels, the transition profile and the target stable profile must first establish a correspondence of the same material location: for finite element samples, the same finite element node number is directly used to correspond to the transition and stable states; for measured samples, corresponding points are established using high-temperature resistant markers, fixed feature points on the plate edge, or structural features that have been verified to be stably trackable; before calculating the displacement difference between the two states, the design baseline of the steel box girder, fixed reference holes, or reference points that do not participate in local welding deformation are used as registration references to perform rigid body registration on the transition and stable profiles, removing overall translation and overall rotation, and only retaining the cross-sectional deformation differences to be learned; the reference temperature for the stable state is taken as the reference ambient temperature set on site, and the final constraint state is defined according to the actual construction process. The process involves releasing all necessary tooling while retaining the support constraints required by design and construction. When the maximum corresponding point coordinate change at at least three consecutive sampling times is not greater than a preset stability threshold, and the difference between the average cross-sectional temperature and the reference ambient temperature is not greater than a preset temperature difference threshold, the stability criterion is satisfied. The stable profile coordinates satisfying the above stability criterion are used as the geometric true value, and the curvature calculation method consistent with S31 is used to calculate the curvature true value on the stable profile. The transverse displacement vector field monitoring label takes the complete geometric displacement difference between the registered transition profile and the corresponding stable profile. Therefore, this label reflects the total deformation difference between the two states caused by thermal shrinkage, residual stress release, constraint changes, and plastic effects.

[0031] Transition deformation profile The pre-trained convolutional deep network is used for feature extraction and stage recognition. The convolutional deep network includes a shared feature extraction backbone network and at least three parallel output branches, which correspond to: elastic recovery stage branch; plastic compression stage branch; and thermal contraction stage branch, respectively.

[0032] The convolutional deep network has a shared structure of convolutional feature extraction backbone, multi-stage parallel identification branches, and stage curvature representation output layers. Specifically, it first collects post-weld cross-sectional images of steel box girders under different welding processes, cooling times, and jig release states, and extracts transition deformation contour maps as training samples. Then, based on synchronously recorded temperature changes, jig release times, subsequent stable linearity measurement results, and manual verification results, it labels the samples with elastic recovery stage, plastic compression stage, or thermal shrinkage stage, and establishes a correspondence between transition deformation contours and stable curvature contours. During training, the samples are input into the network, and the stage classification loss constrains the network to correctly identify the curvature release stage. The representation mapping loss constrains the stage curvature representation units to reflect the corresponding displacement compensation and curvature distribution characteristics. After multiple rounds of backpropagation, the parameters of the convolutional layers, parallel branches, and fully connected layers are updated, and finally, a network model that can be used for actual detection is obtained.

[0033] The correspondence between the transition deformation profile and the stable curvature profile is established according to the aforementioned finite element node number or actual identical mark, and displacement monitoring labels are formed after removing rigid body translation and rotation under a unified registration datum.

[0034] The transition deformation contour map, after grayscale conversion, edge enhancement, and contour extraction, is used as the input layer data, and the image size and grayscale range are unified. Then, a shared feature extraction backbone is constructed through multiple convolutional layers, activation layers, and downsampling layers to extract multi-scale contour features such as weld reinforcement, flange bending, local springback, and overall contraction. At least three parallel output branches are set after the shared backbone, corresponding to the elastic recovery stage, plastic compression stage, and thermal contraction stage, respectively. Each branch performs differential mapping on the same feature tensor, outputs the response value of the corresponding stage, and obtains the probability distribution of each curvature release stage after normalization. At the same time, the hidden layer feature tensor is extracted from the stage branch corresponding to the highest probability, and then outputs the stage curvature representation unit through fully connected mapping, which is used for subsequent analysis of the lateral displacement vector field and local curvature radius distribution coefficient.

[0035] This convolutional deep network is not a typical image classification or curvature extraction network. Instead, it introduces parallel branches for curvature release stages and stage curvature representation output layers in its structure. This allows the network to not only determine which post-weld release stage the current image belongs to, but also output stage features for stable linear reconstruction. This improves the traditional method of directly measuring the current curvature by first identifying the post-weld release state, then reconstructing the stable curvature contour, and finally evaluating the linear accuracy.

[0036] S14. The transition deformation contour map is uniformly processed to meet the input requirements of the convolutional deep network. The processed contour map is input into the network in matrix form, with the value of each pixel serving as the basic input data for the network. The image first enters the front-end layer of the convolutional deep network, i.e., the feature extraction part. The convolutional deep network scans and processes the input image layer by layer through multiple convolutional layers, gradually extracting key deformation information from the contour from the local to the global level, such as edge direction, local convexity, and overall bending trend. The above information is continuously combined and abstracted within the convolutional deep network, ultimately forming a set of high-dimensional feature representations that can characterize the deformation state of the cross-section, i.e., the feature tensor. The obtained unified feature representation is simultaneously input into three parallel output branches, each corresponding to one of the three curvature release stages: elastic recovery, plastic compression, and thermal contraction. Each branch performs weighted accumulation of the input features and compresses or amplifies them nonlinearly, outputting a response value that characterizes the degree of matching between the input profile and the corresponding stage. The response values ​​of the three branches are uniformly normalized to convert them into probabilistic form, obtaining the probability values ​​of the profile belonging to the elastic recovery, plastic compression, and thermal contraction stages, respectively. Finally, by comparing the magnitudes of the three probability values, the stage corresponding to the highest probability is selected as the output result, thus completing the determination of the curvature release stage of the current post-weld section.

[0037] The probability distribution obtained after normalization is expressed as follows: First, the differences are amplified using exponential operations, then normalized by the sum of the exponential values ​​from all stages, finally obtaining... It is a value between 0 and 1, and the sum of the probabilities of all stages is 1. Therefore, It can be used directly to compare sizes. The larger the value, the more the current cross section of the surface matches the characteristics of that stage, thus providing a basis for determining the final curvature release stage.

[0038] in, For the first The probability value of each curvature release phase. For the first The response value output by each branch. In this embodiment, the number of stage categories is [number]. , This is the sum of all stage response values ​​after exponential transformation, used for normalization.

[0039] The current curvature release phase, determined according to the maximum probability criterion, is represented as follows: ;in, To identify the current curvature release phase category.

[0040] S15, determining the current curvature release phase Then, the hidden layer feature tensors of the corresponding branches are extracted from the convolutional deep network. The stage curvature representation unit is generated through fully connected mapping and represented as follows: ; in, As a stage curvature characterization unit, This represents the hidden layer feature tensor for the corresponding stage. The weight parameters of the fully connected layer are obtained by inputting a large number of labeled transition deformation contour maps of the curvature release stage into the convolutional deep network. The prediction result is obtained through forward computation, and the error is calculated with the true stage label. Then, the error is propagated layer by layer to the fully connected layer using the backpropagation mechanism, and its weight parameters are updated with gradients. This continuously adjusts the contribution of each feature component to the output result. After multiple rounds of iterative training, the weight parameters gradually converge, and finally form a set of stable parameters that can effectively represent the mapping relationship between the input features and the curvature release stage. These are the weight parameters of the fully connected layer. The bias term of the fully connected layer is acquired synchronously with the weight parameters during training. In each training round, the bias term is also updated through backpropagation of the error. Under different feature distributions, it can output the corresponding stage results more accurately. After multiple rounds of sample training, the bias term will also converge into a set of stable values, which are used to correct and compensate the fully connected mapping results in practical applications.

[0041] The stage curvature characterization unit is used to comprehensively characterize the deformation feature distribution of the current post-weld section under the curvature release stage, providing input basis for the stage correction of the subsequent transition deformation profile.

[0042] After completing the curvature release stage determination, the target branch corresponding to the current input has been determined, such as the elastic recovery or plastic compression branch. There is no need to use other branches. Directly within the target branch, select the hidden feature layer before its final discrimination output as the extraction object. This layer of features still retains relatively complete spatial or channel information and can reflect the key pattern of contour deformation under this stage. By reading the output result of this layer, the corresponding hidden layer feature tensor can be obtained. The hidden layer feature tensor is a high-dimensional expression of the deformation characteristics of the current section under this stage.

[0043] The extracted hidden layer feature tensors are input into a fully connected mapping structure, where they are flattened and weighted. This involves linearly combining each feature component using weight parameters and adding bias terms to compress the high-dimensional features into a structured low-dimensional representation. This low-dimensional result is the stage curvature representation unit, which expresses the overall deformation characteristics of the current section in the corresponding curvature release stage in a unified numerical form, and is used for subsequent transition deformation correction and stable curvature reconstruction.

[0044] S2, based on the stage curvature characterization unit, performs stage correction on the transition deformation profile in the post-weld section image to reconstruct the stable curvature profile of the target steel box girder assembly welded section.

[0045] S21, analyze the stage curvature characterization unit and map it into physical quantity parameters describing the deformation characteristics of the current curvature release stage, including the lateral displacement vector field and the local curvature radius distribution coefficient.

[0046] Utilizing the feature and deformation correspondence learned during the training phase of the convolutional deep network, the components in the stage curvature representation unit are weighted, combined, and grouped for interpretation, mapping them to two types of physical quantities: The transverse displacement vector field, specifically, provides a displacement vector for each position on the cross section, indicating how much the point has shifted relative to the steady state and the direction of the shift. These displacement vectors are continuously distributed across the entire cross section, forming a displacement field used to describe the predicted deformation shift distribution of the current transition profile relative to the target stable profile. The local radius of curvature distribution coefficient, specifically, provides a quantitative index of the degree of curvature at each location on the cross section, which describes the strength of the curvature of the profile at that location. Through this distribution information, the curvature variation law of the entire cross section can be reflected, rather than a simple overall curvature.

[0047] The above mapping is accomplished through a parameter decoding structure, which takes a stage curvature representation unit as input. In this embodiment, the stage curvature representation unit is a 128-dimensional vector. The decoding structure sequentially includes two fully connected mapping layers, a spatial rearrangement layer, four levels of progressively upsampling convolutional layers, and two parallel output heads. The two fully connected layers first map the 128-dimensional input to 512-dimensional intermediate features, and then map it to low-resolution feature blocks corresponding to the input size of the convolutional depth network. The spatial size of the low-resolution feature blocks is one-sixteenth of the height and width of the input contour map, and the number of channels is 64. After spatial rearrangement, each level of upsampling convolutional layer amplifies the spatial resolution by a factor of 2 and performs convolutional feature fusion until it is restored to the shape corresponding to the network input contour. Figure 1 The H×W output grid is consistent.

[0048] Two parallel output heads are set on the H×W output grid: the displacement output head outputs a two-dimensional displacement field of H×W×2, with the two channels corresponding to the lateral displacement component and the vertical displacement component in the unified cross-section physical coordinate system, respectively; the curvature output head outputs a local curvature radius distribution coefficient field of H×W×1. The decoder outputs a full-pixel grid in form, but only uses the effective contour position in the structure screening mask as the effective output position for subsequent stage correction. Non-contour positions do not participate in the compensation calculation.

[0049] The first in the original transition deformation profile For each contour point, its corresponding position in the H×W output grid is first determined using the pixel-section physical coordinate mapping established by S11. When the contour point is located between grid nodes, bilinear interpolation is used on the displacement field and the local radius of curvature distribution coefficient field to obtain the two-dimensional displacement and curvature coefficient of the contour point, and then compared with the original contour point index. The points are bound together to form a one-to-one correspondence between the original contour points and the decoded output.

[0050] The transverse displacement vector field can be expressed as: ; Reflecting pixels The offset direction and offset magnitude, and the displacement vectors of all positions are combined to form a continuously distributed lateral displacement vector field, which is used to describe the predicted total deformation offset distribution of the entire cross-sectional profile relative to the target stable profile.

[0051] in, For pixels The lateral displacement vector at that location. For the horizontal displacement component, the stage curvature representation unit is input into a preset displacement decoding model. Based on the feature-displacement mapping relationship learned during the training phase, the displacement decoding model weights and combines each feature component to generate displacement vectors at each position of the corresponding cross section. Then, the horizontal component of this displacement vector is extracted at each position, thus obtaining... , The vertical displacement component is obtained in the same way as the horizontal displacement component. After obtaining the displacement vector at each position, the vector is decomposed, and its vertical component is extracted. .

[0052] The local radius of curvature distribution coefficient can be expressed as: ; in, For pixels The curvature coefficient at that point To determine the local radius of curvature at the corresponding location, i.e., the radius of the contour bend, the representation unit is input into the curvature decoding module. Based on the feature-curvature mapping relationship established during the training phase, the distribution of curvature coefficients at each location is calculated. Subsequently, according to the correspondence between curvature and radius of curvature, the curvature coefficients are converted into radii of curvature, thus obtaining the curve. .

[0053] The lateral displacement vector field is used to characterize the predicted total deformation offset of each position of the cross-sectional profile relative to the target stable profile under the current curvature release stage; the local curvature radius distribution coefficient is used to characterize the spatial curvature distribution characteristics of the cross-sectional profile.

[0054] S22, based on the transverse displacement vector field, performs reverse displacement compensation on the discrete sampling points on the transition deformation profile extracted from the post-weld cross-sectional image to eliminate the instantaneous elastic deformation component generated during the current curvature release stage.

[0055] Based on the edge enhancement and contour extraction results described above, one or more continuous transition deformation contour curves are obtained. Then, points are uniformly sampled along these contour curves at fixed intervals. During the sampling process, the starting point of the contour is set as a reference point, and positions satisfying the spacing or feature conditions are selected sequentially along the contour direction, recording their pixel positions in the image coordinate system. For each selected sampling position, its coordinate values ​​in the image are directly read, thus obtaining the corresponding discrete sampling point. At the same time, these sampling points are numbered according to the contour order, the first... The selected point is defined as the nth point. The nth discrete sampling point, let the nth point on the transition deformation profile be... Each discrete sampling point is represented as: Then the compensated point coordinates are expressed as: ;in, For the compensated contour points, As a compensation coefficient, it is set to 1 in the basic implementation method that uses the complete displacement difference between the current contour and the target stable contour as the supervision label, and the predicted deformation offset is not scaled. The lateral displacement vector is the lateral displacement vector field at the corresponding position. The lateral displacement vector field represents the predicted total deformation offset of the transition deformation profile point relative to the target stable curvature profile. The predicted total deformation offset is used to characterize the geometric difference between the two states formed by thermal contraction, residual stress release, constraint change and plastic effect. It is not limited to elastic components, and the displacement vector and the profile point coordinates are in the same cross-sectional coordinate system.

[0056] Through the above-mentioned reverse displacement compensation, the predicted total deformation offset of the original transition deformation profile relative to the target stable curvature profile is compensated, so that the compensated profile directly forms the predicted result of the target stable profile, instead of just stripping the elastic deformation component.

[0057] S23. A smoothing constraint function is constructed using the local radius of curvature distribution coefficient. The local radius of curvature distribution coefficient essentially reflects the bending strength at different locations on the cross-section. Therefore, when constructing the smoothing constraint function, it's not simply required that the curve be as smooth as possible, but rather that its smoothness conforms to the curvature distribution law corresponding to the current stage. Specifically, the local radius of curvature distribution coefficient is first converted into the corresponding curvature intensity distribution. The conversion method is that curvature intensity and radius of curvature are inversely related; that is, the smaller the radius of curvature, the greater the corresponding curvature intensity, indicating more severe bending at that location; conversely, the larger the radius of curvature, the smaller the curvature intensity, indicating that the location tends to be straighter. After obtaining the curvature intensity at each location, it needs to be spatially organized to form a continuous distribution, that is, the curvature intensity of each point is arranged according to the positional order of the cross-sectional contour, thus obtaining the curvature intensity distribution that varies along the entire contour. Then, when constructing the constraint, the bending variation of the curve at each location is penalized and controlled. That is, for areas with small radii of curvature, a certain degree of curvature variation is allowed; for areas with large radii of curvature, a certain degree of curvature variation is allowed. For regions with larger radii, stronger constraints are applied to curvature changes to suppress unreasonable bends. By using position-differentiated constraints, the constraint function not only ensures overall smoothness but also reflects the curvature distribution characteristics that the real structure should have. Spline curve fitting is performed on the compensated contour point set. Specifically, after obtaining the compensated contour point set, the discrete points are first parameterized according to the contour order, that is, each point is assigned a continuously changing parameter value so that it mathematically corresponds to the parameter domain of a curve. Then, a spline function is selected as the fitting model. By adjusting the control points of the spline curve, the curve is made to pass through or approach these discrete points as much as possible. During the fitting process, the curve is iteratively optimized with the dual objectives of minimizing the error of the fitting point set and satisfying the smoothness constraint. On the one hand, it ensures that the overall direction of the curve is consistent with the compensated contour, and on the other hand, it avoids local oscillations or unreasonable bends through smoothness constraints. The final continuous curve is a stable curvature contour that conforms to both the distribution of the compensated point set and the curvature change law. Finally, a stable curvature contour is generated.

[0058] The smoothing constraint function is expressed as: ; in, To fit and optimize the objective function, For the first The compensated contour points, Let be the spline curve function to be determined. For parameterized variables, For spline curves in parameters curvature at that point To smooth the weighting coefficients, the magnitudes of the fitting error term and the smoothing constraint term are calculated on the sample data, and then adjusted... This ensures that the two parts are on the same order of magnitude or in a reasonable proportion during the optimization process, guaranteeing both contour fitting accuracy and avoiding excessive curve bending. When the two reach a relatively optimal balance, the corresponding... That is, the appropriate value. The number of contour points, It is the integral of the overall curvature of the curve, used to measure the smoothness of the curve.

[0059] By minimizing the above objective function, the fitted curve can maintain the smoothness of the overall curvature distribution while conforming to the set of compensation points, thereby obtaining a stable curvature profile that conforms to the actual shape of the structure. The final stable curvature profile is used for subsequent lateral curvature detection and accuracy analysis.

[0060] S3 extracts the transverse curvature detection value of the welded section of the target steel box girder based on the stable curvature profile, compares the transverse curvature detection value with the preset design curvature, and outputs the transverse curvature accuracy result.

[0061] S31, the stable curvature profile is sampled discretely with equal arc length intervals throughout the entire process. The total arc length of the profile is divided into M equal length intervals and M+1 sampling points including the two endpoints are obtained. The sampling arc length interval remains unchanged during one sampling process of the same cross section. The sampling point sequence of the transverse cross section profile is obtained, and the local transverse curvature value at each sampling point is calculated.

[0062] Assume the stable curvature profile is determined by the parametric equations This means that we take countless tiny segments on the curve, and the length of each segment is determined by the changes in that point in both the horizontal and vertical directions, i.e., the real spatial distance formed by the combination of changes in these two directions. Then, we sum up the lengths of all the tiny segments along the curve from the starting point to the current position to obtain the total length of the curve. The arc length function is then expressed as: ;in, For the curve in parameters The corresponding coordinate function is below. For the starting point parameters To the current parameter The corresponding curve arc length; In the arc length range Inner spacing Sampling was performed to obtain the first Each sampling point is represented as: ; in, For the first Coordinates of each sampling point The total arc length of the profile. To determine the sampling arc length interval, a preliminary estimate is made based on the total arc length of the stable curvature profile. The entire profile is then divided into several equal-length segments. If a target number of sampling points is preset, the sampling arc length interval can be taken as the total arc length divided by the number of sampling points, thus obtaining a basic interval value. Secondly, adjustments are made based on the curvature changes. If there are areas with drastic curvature changes in the profile, the sampling arc length interval is appropriately reduced to ensure the accuracy of curvature calculation. In relatively straight areas, the interval can be appropriately increased to reduce redundant sampling points. Thirdly, constraints are also required based on image resolution and actual measurement accuracy. Generally, the sampling interval is required to be no less than the actual length corresponding to pixel-level resolution, but it should not be too large to avoid missing local deformation features.

[0063] The local lateral curvature value at each sampling point is represented as follows: The local transverse curvature value is calculated by combining the first and second derivatives of the curve at that point. The numerator reflects the degree to which the curve deviates from a straight line at that point, while the denominator normalizes the degree of curvature so that the calculation result is not affected by the rate of change of the curve parameters.

[0064] in, For the first The local lateral curvature value of each sampling point The first derivative is used to describe the rate of change of a curve at a certain point or the direction of the tangent. By selecting adjacent sampling points before and after the point, the first rate of change at that point is obtained by dividing the coordinate difference by the parameter interval. The second derivative, used to describe the change in the curvature of a curve, is calculated by differencing the changes in the first derivative at adjacent points. Specifically, the second derivative at a point is obtained by dividing the change in the first derivative between two consecutive segments by the parameter interval. The parameter interval is defined as the target arc length interval. Then, starting from the starting point, gradually search for the condition that the arc length increases to reach a certain value. The position of the sampling point corresponds to the parameter value of the next sampling point, and the difference between the parameter values ​​of two adjacent sampling points is the parameter interval.

[0065] S32 uses the arc length position as the lateral expansion coordinate. Then, according to the magnitude of this coordinate, all sampling points are sorted to form an ordered sequence from one side of the cross-section to the other. The lateral position coordinate is then used as the horizontal axis, and the corresponding local curvature value as the vertical axis, mapping each sampling point to a two-dimensional point. These points are then connected sequentially to form a continuous curve. To make the curve smoother and easier to analyze, interpolation or smoothing can be performed on these discrete points to make the curvature change show a continuous trend, thus obtaining a complete lateral curvature distribution curve. Furthermore, the lateral curvature distribution curve is represented as follows: ;in, For the first The position of each sampling point on the horizontal coordinate axis.

[0066] S33, compare the local transverse curvature value with the design curvature value of the corresponding section position in the preset design curvature database point by point, and calculate the curvature deviation.

[0067] Let the design curvature value be The curvature deviation is then expressed as: ;in, For the first Curvature deviation at each sampling point This is the actual curvature value. To obtain the design curvature value for the corresponding location, the design alignment data of the target steel box girder section is obtained from the engineering design documents. This design alignment data is typically provided in the form of design drawings, BIM models, or CAD curves, describing the geometric contour of the section under ideal conditions. This design contour is then parameterized in a manner consistent with the measured contour, and discrete sampling with equal arc lengths is performed to obtain a sequence of design contour points corresponding one-to-one with the actual sampling points. Curvature calculation is then performed on the design contour; that is, at each design sampling point, the corresponding curvature value is calculated using the same method as the actual curvature calculation, ensuring consistency between the design curvature and the measured curvature in the calculation method. Finally, the design curvature values ​​are stored according to the transverse coordinates of the section, forming a preset design curvature database. During actual inspection, the design curvature value at the corresponding location is read from the database based on the current sampling point's position, thus obtaining the design curvature value. .

[0068] S34, based on a preset graded accuracy evaluation threshold, performs a pass / fail assessment on the curvature deviation and generates a lateral curvature accuracy result report. The preset threshold is set to... The accuracy level determination rule is expressed as follows: ; in, The first precision grading threshold is used to define the high-precision acceptable range. In the sample project, the actual curvature deviation distribution under stable construction conditions is statistically analyzed, and the smaller range, such as the upper limit of the normal fluctuation range, is used as the first-level threshold to ensure that the first precision grading threshold corresponds to the ideal or excellent construction quality level. The second precision grading threshold is used to define the upper limit of acceptable error. Through statistical analysis of historical engineering data, the upper limit of curvature deviation of most qualified components is used as the threshold. , making more than The situation can be clearly identified as non-compliant or requiring correction.

[0069] The final lateral curvature accuracy result report includes: the curvature deviation value of each sampling point, the corresponding position coordinates, and the overall accuracy level evaluation result.

[0070] Example 2 like Figure 2 As shown, the steel box girder welding transverse alignment accuracy detection system based on depth convolution includes the following modules: The image acquisition module is used to acquire the post-weld cross-sectional image of the welded section of the target steel box girder assembly, and to complete image correction, pixel-to-section physical coordinate mapping, and multi-view depth recovery and target section cropping when stereo measurement is required, based on the camera's intrinsic and extrinsic parameters and distortion parameters. The stage recognition module is used to input the post-weld cross-sectional image into the convolutional depth network and identify the corresponding curvature release stage. The characterization output module is used to output the stage curvature characterization unit corresponding to the post-weld cross-sectional image. The stage correction module is used to perform stage correction on the transition deformation profile in the post-weld cross-sectional image based on the stage curvature characterization unit. The curvature reconstruction module is used to obtain the stable curvature profile of the welded section of the target steel box girder assembly. The accuracy detection module is used to extract the lateral curvature detection value based on the stable curvature profile, compare the lateral curvature detection value with the preset design curvature, and output the lateral curvature accuracy result.

[0071] Example 3 To verify the effectiveness and superiority of the method of the present invention, a finite element simulation model of steel box girder assembly and welding based on the actual engineering parameters of a cross-river bridge was constructed, with the specific settings as follows: Steel box girder cross-section specifications: top plate width 18.0m, bottom plate width 12.0m, beam height 3.5m, top plate thickness 24mm, web thickness 16mm, material Q345qD; Welding process parameters: CO2 gas shielded welding, current 280A, voltage 29V, welding speed 6mm / s, double-sided submerged arc welding cover. Simulation of physical field coupling: Thermo-elastic-plastic finite element analysis was performed using ANSYS software to simulate the temperature field, stress field and transient deformation of the entire welding process. The mesh size was controlled at 2mm in the weld zone and 20mm in the far field. Data acquisition simulation: After the simulation model is welded and cooled to 250℃, 200℃, 150℃, 100℃ and room temperature, contour point cloud data is extracted along the transverse section. Gaussian noise and uneven lighting are superimposed to simulate the imaging noise of real industrial cameras, generating a simulated post-weld cross-section image dataset. A total of 1200 sets of cross-section image samples with different welding sequences and different constraints are generated, which are divided into training set, validation set and test set in an 8:1:1 ratio.

[0072] Truth value calibration: The geometric truth value is the stable contour coordinates of the finite element model at the reference ambient temperature, after all the release tools to be released are completed according to the actual construction process and the support constraints required by the design are retained, and the stability criterion is met. The stability criterion is that the maximum node coordinate change of the cross section at no less than three consecutive calculation or sampling times is not greater than the preset stability threshold, and the difference between the average temperature of the cross section and the reference ambient temperature is not greater than the preset temperature difference threshold. The transition contour and the stable contour are established with the same finite element node number to establish the material point correspondence, and rigid body registration is performed based on the reference nodes that do not participate in local welding deformation before comparison to remove the overall translation and overall rotation. The curvature value of each corresponding position is calculated on the stable contour using the curvature calculation method consistent with S31, and the calculation result is used as the curvature truth value for algorithm accuracy evaluation.

[0073] The following three representative existing technical solutions were selected as comparison benchmarks: Table 1 is a table explaining the comparison methods. Mean absolute error of lateral curvature detection: the average absolute value of the difference between the detected curvature and the true curvature; Root mean square error of lateral curvature detection: detects the root mean square of the deviation between the curvature and the true curvature, and is more sensitive to large deviations; Out-of-tolerance rate: The percentage of cross-sections whose maximum curvature deviation exceeds the industry standard acceptable threshold out of the total number of cross-sections tested; Average processing time per cross section: The total time taken from input image to output accuracy result report.

[0074] Comparative analysis of simulation results and effects: Table 2 shows the simulation results analysis. The table shows a significant improvement in detection accuracy: By distinguishing the curvature release stage and performing targeted reverse correction, the method of this invention reduces the mean absolute error (MAE) of lateral curvature detection from 0.228m, the best existing method. -1 Reduced to 0.086m -1 The error was reduced by 62.3%; the root mean square error (RMSE) decreased from 0.297m. -1 Reduced to 0.112m-1 This indicates a stronger ability to suppress larger deviations; Significantly reduced tolerance rate: In engineering applications, cross-sectional deviations mean the need for secondary correction or even rework. The method of this invention reduces the tolerance rate from 19.2% to 3.8%, a relative reduction of 80.2%, greatly reducing the risk of misjudgment and avoiding false positive tolerance alarms caused by transient deformation due to welding heat. Computational efficiency remains superior: Due to the lightweight design of both the stage classification branch and the subsequent correction module, the processing time for a single section in this invention is only 29.4ms, which is even slightly better than that of a single regression network, meeting the needs of online real-time detection in welding production lines.

[0075] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0076] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution, characterized in that, Includes the following steps: Obtain the post-weld cross-sectional image of the welded section of the target steel box girder assembly, and input it into a convolutional depth network to identify the curvature release stage corresponding to the post-weld cross-sectional image; Output the stage curvature characterization unit corresponding to the post-weld cross-sectional image; Based on the stage curvature characterization unit, the transition deformation profile in the post-weld cross-sectional image is corrected in stages. Obtain the stable curvature profile of the welded section of the target steel box girder assembly; Based on the stable curvature profile, the transverse curvature detection value of the welded section of the target steel box girder is extracted, and the transverse curvature detection value is compared with the preset design curvature to output the transverse curvature accuracy result.

2. The method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution according to claim 1, characterized in that, A high-resolution industrial camera array, mounted directly above the steel box girder assembly station, acquires images of the welded section of the target steel box girder after welding and cooling to a preset detectable temperature window. Image acquisition is triggered when the section temperature cools from a high temperature into the preset detectable temperature window, but does not fall below the lower limit of the image capture cutoff temperature. The preset detectable temperature window is defined as follows: ； in, For the welded cross section at time The current temperature; The upper limit of the permitted temperature for taking photos; The lower limit of the temperature at which photography is to be taken; The post-weld cross-sectional image is subjected to grayscale conversion and edge enhancement preprocessing to extract a transition deformation contour map containing the weld reinforcement and deformation features of the two side flanges. ; Grayscale processing is represented as: ;in, For the post-weld cross-section image at pixel points grayscale value at that location , , These are the red, green, and blue channel components of that pixel, respectively. , , These are the grayscale weighting coefficients; Edge enhancement of a grayscale image is represented as: ;in, For pixels Edge enhancement value at the location, This is the result of horizontal gray-level difference. This is the result of gray-level difference in the vertical direction; Based on the edge enhancement value, a threshold is applied to filter the candidate edge points, which are represented as follows: ； in, The binary results for candidate edges, The threshold for edge filtering; Combining the structural screening masks of the weld reinforcement area and the flange area, the transition deformation profile is shown as follows: ； in, This is a transitional deformation profile diagram. For structural screening masks.

3. The method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution according to claim 2, characterized in that, The transition deformation profile map is input into a pre-trained convolutional deep network, which contains at least three parallel output branches, corresponding to the elastic recovery stage, the plastic compression stage and the thermal contraction stage, respectively. The convolutional depth network calculates the probability distribution of the transition deformation contour map corresponding to each curvature release stage, and the probability distribution obtained through normalization is expressed as follows: ;in, For the first The probability value of each curvature release phase. For the first The response value output by each branch. For the number of stage categories, The sum of all stage response values ​​after exponential transformation is used, and the current curvature release stage is identified based on the maximum probability value. The curvature release stage is represented as: ;in, To identify the current curvature release phase category.

4. The method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution according to claim 3, characterized in that, Extract the hidden layer feature tensors corresponding to the curvature release stage from the convolutional deep network, and output them after fully connected mapping as the stage curvature representation unit. The stage curvature representation unit is represented as follows: ； in, As a stage curvature characterization unit, This represents the hidden layer feature tensor for the corresponding stage. These are the weight parameters of the fully connected layer. This is the bias term for the fully connected layer.

5. The method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution according to claim 1, characterized in that, The curvature characterization unit of the stage is analyzed to obtain the lateral displacement vector field and local curvature radius distribution coefficient corresponding to the curvature release stage. The lateral displacement vector field can be expressed as: ,in, For pixels The lateral displacement vector at that location. This represents the horizontal displacement component. For the vertical displacement component, the local radius of curvature distribution coefficient is expressed as: ;in, For pixels The curvature coefficient at that point denoted as the local radius of curvature at the corresponding location.

6. The method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution according to claim 5, characterized in that, Based on the lateral displacement vector field, reverse displacement compensation is performed on the discrete sampling points on the transition deformation profile extracted from the post-weld cross-sectional image to eliminate the instantaneous elastic deformation component generated during the current curvature release stage. Let the first sampling point on the transition deformation profile be... Each discrete sampling point is represented as: Then the compensated point coordinates are expressed as: ;in, For the compensated contour points, For compensation coefficient, The lateral displacement vector is the lateral displacement vector field at the corresponding position. The lateral displacement vector field represents the instantaneous offset of the transition deformation profile point relative to the stable curvature profile, and the displacement vector and the profile point coordinates are in the same cross-sectional coordinate system.

7. The method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution according to claim 6, characterized in that, Using the local radius of curvature distribution coefficient, a smoothing constraint function is constructed. Spline curve fitting is then performed on the compensated contour point set to generate the stable curvature contour of the target steel box girder assembly welded section. The smoothing constraint function can be expressed as: ;in, To fit and optimize the objective function, For the first The compensated contour points, Let be the spline curve function to be determined. For parameterized variables, For spline curves in parameters curvature at that point For smoothing weighting coefficients, This represents the number of contour points.

8. The method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution according to claim 1, characterized in that, The stable curvature profile is subjected to equal-arc-length discrete sampling to obtain a sequence of sampling points for the transverse cross-section profile. The local transverse curvature value at each sampling point is calculated, and the local transverse curvature value is expressed as: ;in, For the first The local lateral curvature value of each sampling point It is the first derivative; It is the second derivative.

9. The method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution according to claim 8, characterized in that, The local transverse curvature value is compared point-by-point with the design curvature value at the corresponding section position in the preset design curvature database to calculate the curvature deviation. Let the design curvature value be... The curvature deviation is then expressed as: ;in, For the first Curvature deviation at each sampling point This is the actual curvature value. The design curvature value for the corresponding position; Based on a preset graded accuracy evaluation threshold, the curvature deviation is assessed for compliance, and a lateral curvature accuracy result report is generated, including the deviation value, deviation location, and overall accuracy level.

10. A system for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution, used to implement the method for detecting the transverse alignment accuracy of welded steel box girders based on depth convolution as described in any one of claims 1-9, characterized in that, Includes the following modules: Image acquisition module, which is used to acquire post-weld cross-sectional images of the welded section of the target steel box girder assembly; A stage recognition module is used to input the post-weld cross-sectional image into a convolutional depth network and identify the corresponding curvature release stage. A characterization output module, which is used to output a stage curvature characterization unit corresponding to the post-weld cross-sectional image; A stage correction module is used to perform stage correction on the transition deformation profile in the post-weld cross-sectional image based on the stage curvature characterization unit. A curvature reconstruction module is used to obtain the stable curvature profile of the welded section of the target steel box girder assembly; The accuracy detection module is used to extract the lateral curvature detection value based on the stable curvature profile, compare the lateral curvature detection value with the preset design curvature, and output the lateral curvature accuracy result.

Images