Method for measuring flatness of bridge steel formwork joint

Abstract

The present application relates to the technical field of bridge engineering detection, and discloses a kind of methods for measuring the flatness of bridge steel formwork joint.The method collects the three-dimensional point cloud data of target joint area according to the preset scanning path, and synchronously obtains the material property parameters of steel formwork;Integrate historical three-dimensional point cloud data and corresponding flatness detection results into a training sample set, and train to generate a first prediction model;Input current three-dimensional point cloud data into the first prediction model, output the predicted joint offset, and determine whether it exceeds the preset threshold;If it exceeds the threshold, activate the second control model matched with the material property parameters, which contains the dynamic mapping relationship between spatial coordinate distribution and flatness deviation;Generate path correction instructions based on the second control model, and return to collect data after adjusting the scanning path.The method realizes effective detection of the flatness of bridge steel formwork joint, and improves the adaptability and accuracy of detection.

Description

Technical Field

[0001] This invention relates to the field of bridge engineering testing technology, specifically a method for measuring the flatness of steel formwork joints in bridges. Background Technology

[0002] During bridge construction, the flatness of the steel formwork joints directly affects the structural appearance quality and load-bearing performance after concrete pouring, making it a crucial aspect of construction quality control. Currently, the flatness inspection of bridge steel formwork joints mainly relies on traditional manual inspection and partially automated inspection methods, but these methods have significant limitations. Traditional inspection methods often use tools such as straightedges and feeler gauges for manual measurement. The inspection process requires inspectors to measure and record each point, which is not only inefficient and unsuitable for inspecting large areas of steel formwork, but also susceptible to subjective factors such as the inspector's experience and sense of responsibility, posing a significant risk of error.

[0003] With the development of digital technology, technologies such as laser scanning and 3D modeling are gradually being applied to flatness inspection, enabling automated analysis through the acquisition of 3D point cloud data. However, existing 3D point cloud acquisition methods mostly employ fixed-path scanning, failing to fully consider the structural characteristics and material differences in the steel formwork splicing area. This often results in insufficient data acquisition or excessive redundant data at key splicing locations. Steel formwork is made of diverse materials, with variations in reflectivity and surface roughness that directly affect laser reflection intensity. Existing methods do not optimize data processing for material properties, leading to high noise and insufficient accuracy in the point cloud data.

[0004] Existing inspection models mostly rely on single spatial coordinate data for flatness assessment, failing to integrate key features such as reflection intensity distribution, making it difficult to accurately identify subtle misalignments at splice joints. The correlation between historical inspection data and results is not effectively utilized, and there is a lack of predictive models based on data accumulation, making it impossible to optimize the current inspection process using existing cases. When flatness exceeds standards, the scanning path cannot be dynamically adjusted to obtain more accurate data for key areas, requiring manual replanning of the scanning range, extending the inspection cycle and impacting construction progress. These issues make existing inspection methods insufficient in terms of accuracy, efficiency, and adaptability to meet the stringent quality control requirements of modern bridge construction for steel formwork splicing. Summary of the Invention

[0005] The purpose of this invention is to provide a method for measuring the flatness of the joints of steel formwork in bridges, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for measuring the flatness of the joints of steel formwork in bridges, the method comprising:

[0007] Step S1: Collect three-dimensional point cloud data of the target bridge steel formwork splicing area according to the preset scanning path. The three-dimensional point cloud data includes a set of spatial coordinates and reflection intensity distribution. Simultaneously, obtain the material property parameters of the steel formwork.

[0008] Step S2: Integrate the historically collected 3D point cloud data with the corresponding flatness detection results into a training sample set. Each training sample set contains an input feature set and an output label. The input feature set includes a set of spatial coordinates and a reflection intensity distribution. The output label is a preset splice seam offset threshold. Train and generate the first prediction model based on the training sample set.

[0009] Step S3: Input the currently acquired 3D point cloud data into the first prediction model, output the predicted splice seam offset, and identify whether the predicted splice seam offset exceeds the preset splice seam offset threshold.

[0010] Step S4: When the predicted splice seam offset exceeds the preset splice seam offset threshold, activate the second control model that matches the material property parameters of the steel formwork. The second control model includes the dynamic mapping relationship between spatial coordinate distribution and flatness deviation.

[0011] Step S5: Generate a 3D point cloud data acquisition path correction instruction based on the second control model, adjust the preset scanning path according to the correction instruction, and return to step S1.

[0012] Preferably, step S1 further includes:

[0013] Historical deformation characteristic data corresponding to the material attribute parameters of steel formwork are extracted from the preset template attribute library. The historical deformation characteristic data includes the distribution of thermal expansion coefficient and stress sensitivity curve.

[0014] Based on the peak intervals in the distribution of thermal expansion coefficients, the target splicing region is divided into several sub-region clusters.

[0015] Assign a corresponding 3D point cloud acquisition density benchmark value and reflection intensity sampling frequency to each sub-region cluster.

[0016] Preferably, step S2 includes:

[0017] The input feature set is extracted from the training sample set to extract highly mutated features, resulting in multiple highly mutated feature sets.

[0018] A centralized analysis was conducted on multiple sets of highly mutated features to determine the centralized values ​​of multiple highly mutated features.

[0019] Calculate the spatial correlation bandwidth sets corresponding to multiple highly abrupt feature set values ​​respectively.

[0020] The receptive field parameters of the network layer are configured based on the spatial correlation bandwidth set.

[0021] The input feature set is processed using the configured multi-scale analysis network layer to generate the first feature map set.

[0022] Input the first feature mapping set and output labels into the regression model, iteratively update the model parameters until convergence, and generate the first prediction model.

[0023] Preferably, step S4 includes:

[0024] Real-time monitoring and prediction of the changing trend of splice seam offset and the fluctuation range of highly abrupt characteristic concentration values.

[0025] When the predicted trend of the splice seam offset exceeds the first warning threshold and the fluctuation range of the highly abrupt feature concentration value is within the preset stable range, the first control model is activated.

[0026] When the fluctuation range of the highly abrupt feature set value exceeds the second warning threshold and the trend of the predicted splice seam offset is within the preset stable range, the second control model is activated.

[0027] If the predicted trend of splice seam offset and the fluctuation range of highly abrupt feature concentration value both exceed the warning threshold, the first control model will be activated first.

[0028] Preferably, step S5 includes:

[0029] Based on the spatial distribution rules in the second control model, the optimal point cloud acquisition density of each sub-region cluster is calculated.

[0030] Using the value of highly abrupt feature clusters as an index, spatial correlation bandwidth diffusion identification is performed within the corresponding sub-region clusters.

[0031] Based on the spatial correlation bandwidth diffusion identification results and the optimized point cloud acquisition density, a 3D point cloud data acquisition path correction instruction is generated.

[0032] Preferably, the method further includes:

[0033] After correcting the 3D point cloud data acquisition path, acquire the newly acquired 3D point cloud data.

[0034] The newly acquired 3D point cloud data is input into the first prediction model, and the updated predicted splice seam offset is output.

[0035] Determine whether the updated predicted seam offset is less than the preset seam offset threshold.

[0036] Otherwise, readjust the mapping relationship of the second control model and return to step S5.

[0037] Preferably, readjusting the mapping relationship of the second control model includes:

[0038] Extract the actual flatness deviation distribution from the newly acquired 3D point cloud data.

[0039] The actual flatness deviation distribution is compared with the predicted deviation range of the second control model.

[0040] The spatial coordinate distribution rules in the second control model are updated based on the comparison results.

[0041] Synchronize the updated spatial coordinate distribution rules to the template attribute library.

[0042] Preferably, the method further includes:

[0043] For sub-region clusters where the actual flatness deviation exceeds the threshold, perform local repair operations.

[0044] The repaired 3D point cloud data was collected for verification.

[0045] When the verification results meet the preset accuracy conditions, the parameters of the current second control model are locked.

[0046] If the verification result does not meet the preset accuracy condition, the first prediction model is retrained.

[0047] Preferably, triggering the retraining of the first prediction model includes:

[0048] The 3D point cloud data before and after the repair, along with the verification results, are integrated into new training samples.

[0049] Add the new training samples to the training sample set.

[0050] Adjust the receptive field parameters of the multi-scale analysis network layers based on the updated training sample set.

[0051] The regression model is retrained to generate the optimized first prediction model.

[0052] Preferably, the method further includes:

[0053] The first and second early warning thresholds are dynamically updated based on historical fluctuation records of highly mutated feature values.

[0054] The activation determination of the first control model or the second control model is performed based on the updated warning threshold.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] This measurement method demonstrates several advantages in detecting the flatness of steel formwork joints in bridges. During data acquisition, it not only obtains the spatial coordinate set and reflection intensity distribution from the 3D point cloud data but also simultaneously collects the material property parameters of the steel formwork, enriching the data dimensions. The spatial coordinates provide geometric location information of the joint area, the reflection intensity distribution reflects the surface material's response characteristics to the scanning signal, and the material property parameters provide a targeted basis for subsequent data processing and model application, avoiding data interpretation biases caused by material differences and laying a more comprehensive data foundation for subsequent analysis and judgment.

[0057] By constructing a training sample set based on historical data and training the first prediction model, the model can fully learn the patterns and features from past detections. The correspondence between input features and output labels in the sample set allows the model to make predictions based on existing learning experience when facing new detection data, reducing reliance on human experience. This data-driven prediction method can more objectively reflect the actual situation of seam offset, avoiding inconsistencies caused by subjective judgment in traditional detection, and making the prediction results more reliable.

[0058] When the predicted seam offset exceeds a preset threshold, a second control model matching the material property parameters is activated. This mechanism demonstrates the method's dynamic adaptability. Different steel template materials exhibit variations in reflectivity and surface condition during scanning. The dynamic mapping relationship between spatial coordinate distribution and flatness deviation included in the second control model allows for targeted analysis based on specific material properties. This targeted analysis ensures accurate identification of key factors affecting flatness when dealing with steel templates of different materials, providing a scientific basis for subsequent path correction.

[0059] Based on the second control model, a 3D point cloud data acquisition path correction command is generated, enabling dynamic adjustment of the scanning path. In cases where the initial scan may contain omissions or insufficient data, adjusting the preset scanning path using the correction command allows subsequent scans to more effectively focus on areas with excessive seam offset, supplementing crucial data. This dynamic adjustment process avoids ineffective repeated scanning and prevents data loss in critical areas, allowing the entire inspection process to be flexibly adjusted according to actual conditions, thus improving the effectiveness and efficiency of data acquisition.

[0060] The entire methodology forms a closed-loop inspection system, with each step—from data acquisition and model prediction to path correction—working in tandem. Through continuous iterative optimization, it ensures the comprehensiveness and accuracy of flatness inspection at bridge steel formwork joints, better adapting to inspection needs under different working conditions and addressing various complex situations that may arise during steel formwork assembly. This provides a practical technical solution for controlling the quality of steel formwork assembly in bridge construction. Compared to traditional static inspection methods, this systematic inspection approach is better suited to the stringent quality control requirements of modern bridge construction, effectively addressing various variables in practical applications and ensuring the continuous and effective conduct of inspection work. Attached Figure Description

[0061] Figure 1 This is a schematic diagram illustrating the working principle of the method for measuring the flatness of bridge steel formwork joints as described in this invention.

[0062] Figure 2 The flowchart for training the first prediction model.

[0063] Figure 3 To control the activation decision subprocess of the model.

[0064] Figure 4 Update the subprocess for mapping the second control model. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Please see Figure 1 This invention provides a method for measuring the flatness of steel formwork joints in bridges, the method comprising:

[0067] Three-dimensional point cloud data of the steel formwork splicing area of ​​the target bridge is collected according to a preset scanning path. This data includes a set of spatial coordinates and a reflection intensity distribution, and the material properties of the steel formwork are also acquired. Historically collected 3D point cloud data is integrated with corresponding flatness detection results to form a training sample set. The training sample set includes an input feature set and an output label. The input feature set includes the set of spatial coordinates and the reflection intensity distribution, and the output label is a preset splice seam offset threshold. A first prediction model is trained based on the training sample set. The currently collected 3D point cloud data is input into the first prediction model, which outputs the predicted splice seam offset. It is then determined whether this offset exceeds the preset threshold. If it exceeds the threshold, a second control model matching the material properties of the steel formwork is activated. This model includes a dynamic mapping relationship between the spatial coordinate distribution and the flatness deviation. A 3D point cloud data acquisition path correction command is generated based on the second control model, and the preset scanning path is adjusted before data acquisition is re-executed.

[0068] Example 1: See Figure 2 In the process of measuring the flatness of bridge steel formwork splicing joints, the acquisition and processing of 3D point cloud data is a core step. First, historical deformation characteristic data matching the current steel formwork material properties are extracted from a pre-set formwork attribute library. This historical data includes the distribution of thermal expansion coefficients and stress sensitivity curves, reflecting the deformation patterns of different materials under temperature changes and stress conditions. Based on the peak intervals in the thermal expansion coefficient distribution, the target splicing area is divided into several sub-region clusters. Each sub-region cluster represents a local area with similar thermal expansion characteristics, and its division is based on the spatial trend of thermal expansion coefficient variations. For each sub-region cluster, a corresponding 3D point cloud acquisition density benchmark value and reflection intensity sampling frequency are assigned. The criterion for determining the peak interval of the thermal expansion coefficient is that when the thermal expansion coefficient value of a certain area exceeds 120% of the average thermal expansion coefficient of the steel formwork material, it is determined to be a peak interval (e.g., the average thermal expansion coefficient of Q235 steel is 12×10⁻⁻⁻⁴). 6 / ℃, then >14.4×10⁻ 6 The area at / ℃ is the peak range); the baseline values ​​for the sampling parameters of the sub-region clusters are: a 3D point cloud acquisition density of 200 points / m² for the peak range sub-region clusters, and a reflection intensity sampling frequency of 20Hz; and an acquisition density of 100 points / m² for the non-peak range, with a sampling frequency of 10Hz. The acquisition density baseline values ​​determine the sampling interval of the laser scanner in this area, while the reflection intensity sampling frequency affects the accuracy of the optical sensor in capturing the surface reflection characteristics.

[0069] After data acquisition, the training sample set construction phase begins. The input feature set of the training sample set includes a set of spatial coordinates and a reflection intensity distribution; the output label is a preset seam offset threshold. After preprocessing, the input feature set undergoes height abrupt change feature extraction. This step identifies areas with potential seams or unevenness by analyzing the height differences between adjacent points in the point cloud data. The extracted height abrupt change feature set contains multiple local anomalous point clusters, each representing a potential flatness defect. These height abrupt change feature sets are then analyzed collectively, calculating the central value of each set, i.e., the average deviation of height changes within that region. The height abrupt change feature extraction uses an adjacent point cloud height difference differential algorithm, calculating the height difference ΔZ between two consecutive sampling points along the Z-axis (template height direction). When ΔZ > 0.5 mm, the point is marked as a height abrupt change point. The selection rule for height abrupt change feature sets is to group three or more consecutive height abrupt change points into one set; single or non-contiguous abrupt change points are not included in the set analysis.

[0070] Based on the concentrated values ​​of highly abrupt feature features, the corresponding spatial correlation bandwidth set is further calculated. The spatial correlation bandwidth reflects the spatial influence range of the highly abrupt feature and is used to guide the parameter configuration of subsequent multi-scale analysis network layers. The receptive field parameters of the multi-scale analysis network layers are adjusted according to the spatial correlation bandwidth, enabling the network to simultaneously capture local details and global trends. The adjusted network processes the input feature set to generate the first feature mapping set. This feature mapping set integrates flatness features at different scales, providing a more comprehensive input for subsequent regression models. The spatial correlation bandwidth is calculated as follows: Using the spatial coordinates corresponding to the concentrated value of the highly abrupt feature as the center, the spatial distribution standard deviation σ of the point cloud cluster to which this concentrated value belongs is calculated, and 2σ is taken as the spatial correlation bandwidth corresponding to this concentrated value (e.g., if σ = 2.5mm for a point cloud cluster with a concentrated value, then the spatial correlation bandwidth is 5mm). The receptive field parameter configuration relationship is as follows: when the spatial correlation bandwidth is 5mm, the receptive field size of the multi-scale analysis network layer is set to 10×10 pixels; when the bandwidth is 8mm, the receptive field size is set to 16×16 pixels; and when the bandwidth is 10mm, the receptive field size is set to 20×20 pixels.

[0071] The regression model is trained using an iterative optimization approach. The input is the first feature map set, and the output is the predicted seam offset. During training, the model parameters are continuously adjusted until the error between the prediction result and the true label converges to a stable range. The final generated first prediction model can quickly and accurately output the predicted seam offset based on new 3D point cloud data. This model fully learns the flatness variation patterns in historical data during the training phase, thus effectively identifying potential seam offset problems when facing new detection tasks.

[0072] During the data acquisition phase, the scanning strategy needs to be dynamically adjusted due to the differences in thermal expansion characteristics among different sub-region clusters. For regions with a high coefficient of thermal expansion, the sampling density and reflection intensity sampling frequency are increased accordingly to ensure that more subtle deformation features can be captured. Conversely, for regions with a low coefficient of thermal expansion, the sampling density can be appropriately reduced to improve overall scanning efficiency. This dynamic adjustment strategy makes the acquisition of 3D point cloud data more accurate while avoiding unnecessary resource waste.

[0073] Extraction and analysis of height abrupt changes are crucial steps in the entire method. By calculating the height differences between adjacent point cloud data, potentially problematic areas can be quickly located. The calculation of stagnation values ​​further quantifies the degree of flatness deviation in these areas, providing a basis for subsequent spatial correlation analysis. The introduction of spatial correlation bandwidth enables the multi-scale analysis network to better adapt to the detection needs of different regions, avoiding detection bias caused by a fixed receptive field.

[0074] The training process of the regression model relies on high-quality feature input. The first feature map set is generated through a multi-scale analysis network, fusing spatial information at different levels, enabling the model to more comprehensively understand the flatness of the stitched region. The iterative optimization process ensures the model's stability and generalization ability, allowing it to maintain high accuracy when facing detection tasks under different material or environmental conditions.

[0075] Example 2: See Figure 3 In the process of flatness inspection at the splice joints of bridge steel formwork, the dynamic response mechanism of the real-time monitoring system is a key factor in ensuring the accuracy of the inspection. This embodiment mainly involves the collaborative analysis between the predicted trend of splice joint offset and the fluctuation amplitude of the concentrated value of height abrupt change characteristics, as well as the model activation strategy based on this analysis.

[0076] The monitoring system continuously tracks and predicts the changing trend of splice joint offset, which reflects the dynamic deformation characteristics of the steel formwork under external loads or temperature changes. Simultaneously, the system calculates the fluctuation amplitude of the concentrated value of height abrupt change characteristics in real time, which characterizes the severity of local flatness anomalies. To quantify the dynamic relationship between the two, a trend-fluctuation coupling coefficient is introduced. Its expression is:

[0077]

[0078] in: This indicates the rate of change of the predicted seam offset. This indicates the fluctuation range of the concentrated values ​​of highly aberrant characteristic features. This is the thermal resistance coefficient of the steel formwork material. This coefficient comprehensively reflects the interaction strength between the deformation trend and local fluctuations.

[0079] When the predicted seam offset exceeds a preset first warning threshold, and the fluctuation range of the highly abrupt change characteristic concentration value is within a preset stable range, the system activates the first control model. The first control model primarily compensates for and corrects the overall deformation trend. Its core is to suppress systematic offsets by adjusting the spatial sampling strategy of the 3D point cloud data. Based on deformation patterns under similar working conditions in historical data, this model generates compensation parameters for the current trend. These parameters directly affect the path planning and sensor configuration for subsequent data acquisition.

[0080] If the fluctuation range of the highly abrupt change feature concentration exceeds the second warning threshold, while the predicted splice seam offset trend remains within a stable range, the system activates the second control model. The second control model focuses on fine-tuning local flatness anomalies. It establishes a dynamic mapping relationship between spatial coordinate distribution and flatness deviation to address problem areas specifically. The model first identifies sub-region clusters with excessive fluctuation ranges, and then generates local correction schemes based on the correlation analysis between material property parameters and historical deformation data. These schemes may include adjusting the focusing depth of the laser scanner or changing the exposure parameters of the optical sensor to obtain more accurate local point cloud data.

[0081] The first warning threshold (predicted splice joint offset trend) is exceeded when the hourly growth rate of the predicted splice joint offset is greater than 0.2 mm. The second warning threshold (fluctuation amplitude of height abrupt change characteristic concentration value) is exceeded when the standard deviation of the height abrupt change characteristic concentration value is greater than 0.3 mm. The allowable deviation of the flatness at the steel formwork splice is ≤2 mm / m. The warning threshold is set at 1 / 10 of the allowable deviation to ensure early identification of potential defects.

[0082] In special cases, when the predicted trend of seam offset and the fluctuation of the high abrupt change feature value both exceed their respective warning thresholds, the system prioritizes activating the first control model. This selection is based on the dominant influence of the overall deformation trend on the final detection result. After completing the compensation and correction of the first control model, the system re-evaluates the fluctuation of the high abrupt change feature; if it still exceeds the threshold, the second control model is activated. This phased processing strategy effectively avoids system instability caused by adjusting multiple parameters simultaneously. The matching logic between material properties and the second control model is as follows: automatic matching is achieved through the "material parameter - model type" mapping relationship in the preset template attribute library: ① Coefficient of thermal expansion 12-13×10⁻ 6 / ℃, surface roughness Ra1.6μm steel template (such as Q235 steel precision-machined parts), matched with the "low deformation sensitive type" second control model; ② thermal expansion coefficient >13×10⁻ 6A steel template with a surface roughness of Ra3.2μm and a temperature of / ℃ (such as a rough-machined Q345 steel part) is matched with a second control model of "high deformation sensitivity".

[0083] Trend-Volatility Coupling Coefficient The calculation process requires real-time acquisition of multiple dynamic parameters. This includes predicting the rate of change of the splice seam offset. This is obtained by differential calculation of the offsets output at consecutive time points, reflecting the speed of deformation development. The fluctuation amplitude of the concentrated values ​​of highly abrupt characteristic features is also shown. It is determined by the standard deviation of the concentrated values ​​within the sliding time window and is used to measure the degree of abnormal fluctuations in local smoothness. Thermal resistivity Extracted from the material property library, it demonstrates the steel formwork's resistance to temperature changes.

[0084] The warning thresholds are not fixed but dynamically adjusted based on the detection environment. Under conditions of drastic temperature changes, the system automatically raises the first warning threshold to accommodate normal deformation caused by material expansion and contraction. Similarly, when residual stress concentration is detected in the steel formwork, the second warning threshold is lowered to enhance sensitivity to localized defects. This dynamic threshold management mechanism significantly improves the system's adaptability to different environmental conditions.

[0085] The data processing flow after model activation features feedback regulation. After completing compensation correction, the first control model collects new point cloud data to verify the adjustment effect. If the corrected prediction offset still exceeds the allowable range, the system iteratively optimizes the compensation parameters until a stable state is reached. The second control model rescans the problem sub-region after each local adjustment, evaluating the correction effect by comparing the changes in feature set values ​​before and after the adjustment. This closed-loop regulation mechanism ensures the reliability of the detection results.

[0086] The real-time performance of the monitoring system relies on an efficient data processing architecture. The calculation of predicted seam offsets employs an incremental update algorithm, processing only the differences between newly added point cloud data and the previous state, significantly reducing the computational load. The analysis of highly abrupt feature cluster values ​​utilizes a parallel computing model, simultaneously performing feature extraction and fluctuation calculation for different sub-region clusters. This optimized design enables the system to meet real-time requirements while maintaining detection accuracy.

[0087] Example 3: In a bridge steel formwork splicing flatness detection system, dynamic optimization of the 3D point cloud data acquisition path is a key step in achieving high-precision measurement. This example details the specific method for correcting the data acquisition path based on the spatial distribution rules of the second control model, as well as the complete process for verifying the correction effect. The spatial distribution rules of the second control model are as follows: Based on the flatness deviation of sub-region clusters, the following classifications apply: ① Deviation > 1mm is a high-priority acquisition area; ② Deviation 0.5-1mm is a medium-priority acquisition area; ③ Deviation < 0.5mm is a low-priority acquisition area.

[0088] Once the second control model is activated, the system first calculates the optimized point cloud acquisition density for each sub-region cluster based on the material properties and deformation characteristics of the current steel formwork. This calculation process considers the thermal expansion characteristics of the material and historical deformation data, converting the theoretical density requirement into practically executable acquisition parameters through spatial distribution rules. The calculation of the optimized acquisition density uses a region weight allocation method, the mathematical expression of which is:

[0089]

[0090] in: This represents the optimized acquisition density of the i-th sub-region cluster. The weight representing the influence of the j-th deformation feature. It is the spatial distance between the current sub-region and the feature reference point. For the k-th material attribute parameter, This is the adjustment coefficient for the corresponding parameters, where n represents the total number of deformation features affecting the calculation of the current sub-region cluster acquisition density, and m represents the total number of deformation features affecting the calculation of the current sub-region cluster acquisition density. This formula comprehensively considers the influence of spatial location relationships and material properties on acquisition density.

[0091] Using the concentrated value of highly abrupt feature as an index, the system performs spatial correlation bandwidth diffusion identification within the corresponding sub-region cluster. This process analyzes the spatial propagation characteristics of the concentrated feature values ​​to determine the detection range that needs to be focused on. The identification algorithm adopts an adaptive radius search method, with the initial search radius automatically adjusted according to the magnitude of the concentrated feature values, dynamically converging to the optimal value during iterative calculation. In this way, the system can accurately capture the spatial influence range of highly abrupt features, avoiding over-collection or omission of key areas. The specific operation steps of spatial correlation bandwidth diffusion identification are as follows: ① Take the spatial coordinates (X0, Y0) where the concentrated value of highly abrupt feature is located as the starting point; ② Diffusion outward with a step size of 2mm along the splicing seam length direction (X-axis) and perpendicular to the splicing seam direction (Y-axis); ③ Monitor the rate of change of reflection intensity of the point cloud within the diffusion area in real time (rate of change = |current region reflection intensity - starting region reflection intensity| / starting region reflection intensity × 100%); ④ Stop diffusion when the rate of change of reflection intensity < 5%, and the diffusion range at this time (e.g., X0±8mm, Y0±8mm) is the final spatial correlation bandwidth.

[0092] Based on the spatial correlation bandwidth diffusion identification results and optimized point cloud acquisition density, the system generates 3D point cloud data acquisition path correction instructions. These instructions consist of two main parts: first, optimizing the scanning order of each sub-region cluster, prioritizing regions with higher feature concentration values ​​and larger spatial correlation bandwidths; and second, detailed settings for scanning parameters, including specific parameters such as laser power, scanning speed, and sampling interval. These instructions are transmitted to the scanning device via a standardized interface, enabling real-time adjustment of the acquisition path.

[0093] The selection of spatial coordinate points and the determination of flatness are achieved through specialized 3D laser scanning equipment. The scanning equipment operates along a corrected path, emitting a laser beam onto the surface of the steel template and recording the 3D coordinate data of each measurement point by receiving the reflected signals. These spatial coordinate points specifically include positional information in three dimensions: X, Y, and Z. The X-axis typically runs along the length of the steel template, the Y-axis along the width, and the Z-axis represents the height. During the scanning process, the equipment acquires a large amount of point cloud data at an extremely high frequency, collecting hundreds of spatial coordinate points per square meter, thus forming a dense point cloud matrix. Flatness determination is based on this massive amount of spatial coordinate point data. The system first preprocesses the acquired point cloud data, including noise filtering, coordinate unification, and data registration, to ensure data accuracy and consistency. Subsequently, it identifies height abrupt changes by calculating the height difference between adjacent points; these features directly reflect the surface unevenness. For the splicing seam area, the system extracts point cloud data from both sides of the seam, performs local plane fitting, and quantifies the flatness deviation by calculating the deviation angle between the fitted plane and the ideal reference plane, as well as the standard deviation of the distance between the point cloud and the fitted plane. This deviation value is a direct measure of the seam offset.

[0094] After correcting the 3D point cloud data acquisition path, the system immediately executes a new round of data acquisition. The newly acquired 3D point cloud data is first preprocessed, including noise filtering and coordinate unification, and then input into the first prediction model for processing. The first prediction model processes the input feature set through an internal multi-scale analysis network layer. This network layer has been trained and optimized based on historical data and can effectively extract height abrupt change features and their spatial distribution patterns in the point cloud data. The model output is the predicted seam offset, which is a quantified value that integrates local and global flatness information and is directly used as the flatness measurement result. If this value is lower than a preset threshold, the flatness is deemed acceptable; if it still exceeds the threshold, a new round of correction loop is triggered.

[0095] When the updated predicted seam offset still exceeds the preset threshold, the system initiates the mapping adjustment process of the second control model. The adjustment process employs incremental learning, using the latest collected data as training samples to update the spatial coordinate distribution rules without altering the original model architecture. Specific operations include: recalculating the spatial distribution characteristics of the feature set values, adjusting the regional weight allocation parameters, and optimizing the adjustment coefficients in the collection density calculation formula. These adjustments are implemented using a mini-batch gradient descent algorithm to ensure the model can quickly adapt to new detection requirements. Through this iterative optimization approach, the system gradually reduces the measurement uncertainty in key areas, ultimately obtaining accurate flatness measurement results. The entire process forms a closed-loop system from data acquisition, processing, analysis to decision-making, ensuring the accuracy and reliability of flatness measurement.

[0096] The data acquisition path modification is an iterative optimization process. Each modification generates new detection data, which in turn is used to improve the mapping relationship of the control model. The system establishes a modification history database to save the parameter settings and effect evaluation data of each modification, providing a reference for subsequent modification decisions. This adaptive optimization mechanism enables the system to gradually approach the optimal acquisition scheme through multiple iterations.

[0097] The verification of the correction effect employs a multi-index comprehensive evaluation method. In addition to the primary indicator of predicted seam offset, the system also monitors auxiliary indicators such as the integrity of point cloud data, the stability of feature extraction, and computational efficiency. These indicators are weighted and summed to obtain a comprehensive score, which is used to determine whether a next round of path correction is needed. The scoring threshold is dynamically adjusted according to the accuracy requirements of the detection task, maximizing work efficiency while ensuring detection quality.

[0098] In special cases, when multiple corrections fail to bring the predicted offset to the target level, the system will trigger a more in-depth analysis process. This process includes: thoroughly checking the accuracy of material property parameters, verifying sensor calibration status, and analyzing the impact of environmental factors on the measurement results. Based on the analysis results, the system may take measures such as adjusting the material parameter library, recalibrating the equipment, or changing the detection environment to fundamentally solve the detection deviation problem.

[0099] The entire implementation process demonstrates the intelligence and adaptive characteristics of the detection system. Through closed-loop control of real-time data acquisition, model prediction, path correction, and effect verification, the system can automatically adjust its working strategy for different detection scenarios. This method is not only applicable to the flatness detection of bridge steel formwork, but its core concept can also be extended to other engineering fields requiring high-precision three-dimensional measurement. The modular design of the system allows each functional component to be optimized independently, while simultaneously enabling collaborative work through standard interfaces, ensuring the stability and scalability of overall performance.

[0100] Example 4: See Figure 4 In the flatness detection system for bridge steel formwork splices, handling the deviation between actual detection data and model prediction results is a crucial step in ensuring long-term detection accuracy. This embodiment details how to update the mapping relationship of the second control model based on the actual flatness deviation distribution, as well as the subsequent local repair and verification process. The comparison standard between actual flatness deviation and predicted deviation is as follows: when the absolute value of the difference between the actual flatness deviation and the predicted deviation of the second control model is >0.3mm, it is judged as "deviation mismatch," and the model mapping relationship needs to be updated. The update method for the spatial coordinate distribution rule is as follows: the spatial coordinate weight coefficient corresponding to the sub-region cluster with excessive deviation is increased by 20% (e.g., the original weight coefficient is 1.0, and the updated weight coefficient is 1.2), while the weight coefficients of other regions remain unchanged, ensuring that the model pays more attention to the deviation areas.

[0101] When the system detects a significant deviation between the newly acquired 3D point cloud data and the predictions of the second control model, it initiates a model update procedure. First, it extracts the actual flatness deviation distribution from the new data. This distribution records the difference between the measured and predicted values ​​for each measurement point. The system then categorizes these discrepancies into corresponding sub-region clusters based on their spatial location, forming a regional deviation statistics table (see Table 1).

[0102] Table 1: Sub-region cluster bias analysis table.

[0103]

[0104] This table displays deviation analysis data for four typical sub-region clusters, including spatial distribution characteristics and material parameters. By comparing these actual measurement data with the original prediction deviation range of the second control model, the system identifies the mapping relationships that need adjustment. For example, the average deviation of sub-region A2 exceeds the prediction range, and its stress sensitivity level is high, indicating that the current model's prediction rules for this type of region need to be revised.

[0105] The process of updating the mapping relationship employs a hierarchical adjustment strategy. For sub-regions with small deviations (such as B1), only their spatial coordinate weight coefficients are fine-tuned; for regions with significant deviations (such as A2), the complete mapping rules for that region are reconstructed. The updated rules consider three dimensions of influencing factors: the distribution pattern of spatial coordinates, the combination characteristics of material properties, and the dynamic changes in environmental parameters. These new rules are synchronized to the template attribute library as benchmark data for subsequent detection.

[0106] After the model update is completed, the system performs local repair operations on sub-region clusters with deviations exceeding the threshold. The repair process adopts a hierarchical approach: for surface flatness deviations, micron-level trimming is performed using mechanical grinding equipment; for structural deformations, a combination of local heating and pressure correction is used. Each repair step is equipped with a corresponding quality checkpoint to monitor the repair effect in real time. The specific process for local repair is as follows: diamond wheel mechanical grinding is used, with the grinding wheel speed set to 3000 r / min and the grinding depth controlled between 0.1-0.3 mm. After repair, the surface roughness of the sub-region cluster must reach Ra≤1.6μm. The accuracy verification condition is: the "flatness error of the sub-region cluster ≤1mm / m" calculated from the 3D point cloud data collected after repair, and this error is stable within the threshold for three consecutive measurements, which is considered to meet the preset accuracy condition.

[0107] After the repair operation is completed, the system immediately acquires new 3D point cloud data of the repaired area for verification. The verification process employs a dual standard: first, it checks whether the absolute flatness of the repaired area meets the specification requirements; second, it evaluates the smoothness of the transition between this area and the surrounding areas. The verification data is then input into the updated second control model to check whether the prediction deviation has been reduced to within the allowable range.

[0108] When the verification results meet the preset accuracy conditions, the system locks the parameter state of the current second control model. The locked model parameters are marked as "verified" version, and a corresponding configuration fingerprint is generated. These fingerprints contain a summary of the model's key parameters for version management and traceability in subsequent detection tasks. Simultaneously, the system automatically creates a recovery point for this version to allow for quick rollback in case of anomalies during subsequent use.

[0109] If the verification results do not meet the requirements, the system triggers the retraining process of the first prediction model. Retraining not only uses the latest patched data but also retrieves detection records under similar operating conditions from the historical database to construct a more comprehensive training sample set. During training, special attention is paid to regional features with large prediction deviations in the past, and the model's local prediction ability is improved by enhancing the sample weights of these features.

[0110] The entire implementation process established a complete closed loop for deviation handling: from deviation detection and model updates to remediation and verification, each stage had corresponding quality control nodes. The system maintains a dynamically updated database of exception cases, recording the detailed process and final solution for each deviation handling. This case data provides a reference for handling similar issues in the future, forming a knowledge accumulation mechanism for continuous improvement.

[0111] In practical engineering applications, this implementation method demonstrates strong adaptability. For steel formwork joints made of different materials, the system can automatically select appropriate processing strategies based on specific deviation characteristics. At construction sites with significant ambient temperature variations, the system increases its sensitivity to thermal expansion-related deviations and adjusts the model update frequency accordingly. This targeted approach effectively balances the relationship between detection accuracy and system stability.

[0112] The data structure shown in Table 1 can be expanded according to specific project needs in practical applications. In large-scale bridge projects, fields such as construction batch and inspection timestamp may be added; for templates made of special alloy materials, corresponding metallurgical parameters will be supplemented. This flexible data organization method enables the system to adapt to diverse engineering inspection needs.

[0113] Through continuous iterative model updates and repair verification mechanisms, the system gradually improves its understanding of the flatness characteristics of various steel formwork joints. This self-optimization capability enables the detection accuracy to continuously improve with usage time, ultimately forming customized detection solutions for specific engineering environments.

[0114] Example 5: In the process of flatness detection at the joints of bridge steel formwork, the core of continuous system optimization lies in the iterative upgrading of the prediction model and the dynamic improvement of the monitoring mechanism. This example focuses on how to reconstruct the first prediction model using the complete dataset of repair operations and combine it with historical records to achieve adaptive adjustment of the early warning system.

[0115] After the repair operation is completed, the system automatically collects three types of key information: the original 3D point cloud data before repair, the operation parameter records during the repair process, and the detection results during the post-repair verification phase. This information is integrated into uniquely identified new training samples based on spatiotemporal dimensions. Each sample contains a complete input feature set and an output label verified on-site. When new samples are added to the training sample set, spatiotemporal stamps and regional location codes are retained, forming a dataset with spatiotemporal traceability capabilities. During data integration, the system performs feature alignment operations to ensure that the coordinate system of the point cloud remains consistent before and after repair, and that the spatial boundaries of the repaired area are accurately labeled. Threshold for the number of newly added training samples: Each time the first prediction model is retrained, at least 50 complete samples of "pre-patch 3D point cloud data + post-patch 3D point cloud data + validation results" must be integrated to ensure that the sample size meets the model optimization requirements; Adjustment range of receptive field parameters: Adjusted according to the rate of change of highly abrupt feature values ​​in the newly added samples: when the rate of change is <10%, the receptive field parameters are adjusted by +10%; when the rate of change is 10%-20%, they are adjusted by +20%; when the rate of change is >20%, they are adjusted by -10% (to avoid excessive parameter shift).

[0116] The expanded training sample set triggers the retraining process of the first prediction model. The first step in retraining is to reconfigure the receptive field parameters for the multi-scale analysis network layers. The system analyzes the distribution characteristics of the newly added samples in the spatial coordinate set and calculates the statistical trend of the values ​​in the highly abrupt feature set. Based on this trend distribution, the boundary values ​​of the spatial correlation bandwidth are redefined, and then the optimal receptive field coverage required for each analysis layer is derived. The parameter adjustment process uses gradient sensitivity analysis to implement differentiated configurations based on the differences in importance of different sub-regions. The final generated receptive field parameter matrix maps the weight relationships of various features in the current engineering environment.

[0117] Using the updated multi-scale analysis network layer, the system initiates the retraining process for the regression model. An incremental learning mechanism is introduced during training, prioritizing the optimization of neuron connection weights highly correlated with new samples while retaining the core parameters of the original model. Iterative training employs an adaptive learning rate strategy, initially applying higher learning weights to error terms in the areas requiring repair, gradually expanding to balance adjustments across the entire sample set. After each iteration, the model's performance on the historical validation set is verified to prevent overfitting. The convergence condition is set to a minimum validation error fluctuation over three consecutive iterations; at this point, the network parameters are locked, generating the optimized first prediction model.

[0118] The system also establishes a historical fluctuation database of highly abrupt characteristic values. This database continuously records monitoring values ​​for each sub-region cluster on a minute-by-minute basis, with additional annotations of operating parameters such as ambient temperature, detection time, and load status. Data storage employs a sliding time window mechanism, retaining valid records for the most recent 12 months. Based on this database, the system periodically performs dynamic revisions of the warning thresholds. The revision process first categorizes historical data by operating condition, calculating the long-term variation curve of the fluctuation amplitude for each type of operating condition. By statistically analyzing the standard deviation distribution within each time window, the boundary values ​​of the fluctuation intervals at different confidence levels are derived. The final first and second warning thresholds are set at the upper boundary of the specified confidence interval according to the engineering acceptance standards, with a safety margin reserved. The statistical period for the historical fluctuation records of highly mutated characteristic concentrated values ​​is based on natural weeks, and a "sub-region cluster fluctuation trend report" is generated every week. The report includes the maximum fluctuation value, average fluctuation value, and number of fluctuations exceeding the standard within the week. The early warning threshold is updated once a month based on the fluctuation report of the previous month. If the early warning threshold exceeds the standard 3 or more times in a month, a temporary update is triggered (not limited by the monthly cycle) to ensure that the threshold is adapted to the real-time operating conditions.

[0119] In practical applications of the early warning threshold, the system employs a hierarchical judgment strategy. When real-time monitoring data enters the threshold neighborhood, i.e., within the critical warning zone (below the formal warning value but above 120% of the baseline value), the system activates the preparatory analysis process. This process automatically extends the monitoring time window, increases the sampling frequency, and calls historical data from the same period to compare the current fluctuation pattern. The control model is only formally activated when sufficient judgment conditions are met. The model activation logic includes three levels of screening: primary screening is based on the absolute value of the current fluctuation amplitude; secondary screening analyzes the second derivative characteristics of the fluctuation trend curve; and the final decision is verified by combining the state correlation of neighboring areas.

[0120] The deployment process after model retraining adopts a dual-track parallel mode. The new model first runs in a shadow computing environment, comparing the prediction differences between the old and new versions on the same input data in real time. The difference analysis focuses on key indicators in key regions, setting a tolerance range for differences. If the differences in the comparison data remain within the tolerance range for 24 consecutive hours, the system automatically switches to the new version model as the master. During the transition period, the old version model is in a hot standby state, supporting seamless reverting if the new version exhibits abnormal predictions. When officially switching over, a version change report is generated, recording details of model parameter changes and expected performance changes.

[0121] The metadata management system records all operational processes throughout the implementation, including change logs of the training sample set, historical versions of model parameter adjustments, timelines of early warning threshold revisions, and changes in model deployment status. This data supports state reconstruction and root cause analysis at any point in time. The system has a regular review mechanism, conducting a technical evaluation of the entire process's effectiveness every quarter, with the review results driving optimization decisions for the next cycle.

[0122] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0123] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for measuring the flatness of the joints of steel formwork in bridges, characterized in that, Includes the following steps: Step S1: Collect three-dimensional point cloud data of the target bridge steel formwork splicing area according to the preset scanning path. The three-dimensional point cloud data includes a set of spatial coordinates and reflection intensity distribution. Simultaneously, obtain the material property parameters of the steel formwork. Step S2: Integrate the historically collected 3D point cloud data with the corresponding flatness detection results into a training sample set. Each training sample set contains an input feature set and an output label. The input feature set includes a set of spatial coordinates and a reflection intensity distribution. The output label is a preset splicing seam offset. Train and generate the first prediction model based on the training sample set. Step S3: Input the currently acquired 3D point cloud data into the first prediction model, output the predicted splicing seam offset, and identify whether the predicted splicing seam offset exceeds the preset splicing seam offset threshold. Step S4: Real-time monitoring and prediction of the changing trend of the splice seam offset and the fluctuation range of the high abrupt change characteristic concentration value; When the predicted trend of splice seam offset exceeds the first warning threshold and the fluctuation range of highly abrupt feature concentration value is within the preset stable range, the first control model is activated. When the fluctuation range of the highly abrupt feature set value exceeds the second warning threshold and the trend of the predicted splice seam offset is within the preset stable range, the second control model is activated. If the predicted trend of splice seam offset and the fluctuation range of highly abrupt feature concentration value both exceed the warning threshold, the first control model will be activated first. Step S5: Generate a 3D point cloud data acquisition path correction instruction based on the second control model, adjust the preset scanning path according to the correction instruction, and return to step S1.

2. The method for measuring the flatness of bridge steel formwork joints according to claim 1, characterized in that, Step S1 further includes: Extract historical deformation characteristic data corresponding to the material property parameters of steel formwork from the preset template attribute library. The historical deformation characteristic data includes the distribution of thermal expansion coefficient and stress sensitivity curve. Based on the peak intervals in the distribution of thermal expansion coefficients, the target splicing region is divided into several sub-region clusters; Assign a corresponding 3D point cloud acquisition density benchmark value and reflection intensity sampling frequency to each sub-region cluster.

3. The method for measuring the flatness of bridge steel formwork joints according to claim 2, characterized in that, Step S2 includes: Extract the input feature set from the training sample set to perform high mutation feature extraction, and obtain multiple high mutation feature sets; A centralized analysis was performed on multiple sets of highly mutational features to determine the centralized values ​​of multiple highly mutational features; Calculate the set of spatial correlation bandwidths corresponding to multiple highly abrupt feature values; The method for calculating the spatial correlation bandwidth is as follows: taking the spatial coordinates corresponding to the concentrated value of height change feature as the center, calculate the spatial distribution standard deviation σ of the point cloud cluster to which the concentrated value of height change feature belongs, and take 2σ as the spatial correlation bandwidth corresponding to the concentrated value. Based on the configuration of spatial correlation bandwidth sets, multi-scale analysis of network layer receptive field parameters is performed. The configured multi-scale analysis network layer is used to process the input feature set and generate the first feature mapping set. Input the first feature mapping set and output labels into the regression model, iteratively update the model parameters until convergence, and generate the first prediction model.

4. The method for measuring the flatness of bridge steel formwork joints according to claim 1, characterized in that, Step S5 includes: Based on the spatial distribution rules in the second control model, calculate the optimal point cloud acquisition density for each sub-region cluster; Using the concentrated value of highly abrupt feature as an index, spatial correlation bandwidth diffusion identification is performed within the corresponding sub-region cluster; Based on the results of spatial correlation bandwidth diffusion identification and the optimization of point cloud acquisition density, a 3D point cloud data acquisition path correction instruction is generated.

5. The method for measuring the flatness of bridge steel formwork joints according to claim 4, characterized in that, Also includes: After completing the 3D point cloud data acquisition path correction command, acquire the newly acquired 3D point cloud data; The newly acquired 3D point cloud data is input into the first prediction model, and the updated predicted splice seam offset is output. Determine whether the updated predicted seam offset is less than the preset seam offset threshold; Otherwise, readjust the dynamic mapping relationship between the spatial coordinate distribution of the second control model and the flatness deviation, and return to step S5.

6. The method for measuring the flatness of bridge steel formwork joints according to claim 5, characterized in that, The dynamic mapping relationship between the spatial coordinate distribution of the second control model and the flatness deviation is readjusted, including: Extract the actual flatness deviation distribution from the newly acquired 3D point cloud data; The actual flatness deviation distribution is compared with the predicted deviation range of the second control model; Update the spatial coordinate distribution rules in the second control model based on the comparison results; Synchronize the updated spatial coordinate distribution rules to the template attribute library.

7. The method for measuring the flatness of bridge steel formwork joints according to claim 6, characterized in that, Also includes: For sub-region clusters where the actual flatness deviation exceeds the threshold, perform local repair operations; Collect and verify the repaired 3D point cloud data; When the verification results meet the preset accuracy conditions, the parameters of the current second control model are locked. If the verification result does not meet the preset accuracy condition, the first prediction model is retrained.

8. The method for measuring the flatness of bridge steel formwork joints according to claim 7, characterized in that, The following factors trigger the retraining of the first prediction model: The 3D point cloud data before and after the repair and the verification results are integrated into new training samples; Add the new training samples to the training sample set; Adjust the receptive field parameters of the multi-scale analysis network layers based on the updated training sample set; The regression model is retrained to generate the optimized first prediction model.

9. The method for measuring the flatness of bridge steel formwork joints according to claim 8, characterized in that, Also includes: Based on the historical fluctuation records of highly mutated feature values, the first and second early warning thresholds are dynamically updated. The activation determination of the first control model or the second control model is performed based on the updated warning threshold.

Images