A system and method for detecting the quality of a mid-beam pier column of a crawling device

Abstract

The application relates to the technical field of bridge quality detection, in particular to a middle tie beam pier column quality detection system and method for a crawling device, which comprises the following steps: collecting data through a multidimensional sensor group, constructing a three-dimensional point cloud model, extracting local features and time sequence dependence, dividing a sector area and generating a defect detection curve, and finally outputting a detection result through weighted fusion. The application can effectively improve the detection efficiency, expand the coverage range, reduce data lag, and provide efficient and accurate technical support for bridge pier column quality evaluation.

Description

Technical Field

[0001] This invention relates to the field of beam and pier inspection technology, and more specifically, to a quality inspection system and method for beam and piers of crawling equipment. Background Technology

[0002] During bridge or tunnel construction, the central beam piers, as crucial load-bearing and connecting structures, directly impact the overall stability and service life of the project due to the quality of their concrete pouring, dimensional accuracy, and surface integrity. Current quality inspection methods for central beam piers primarily rely on manual or semi-automated methods, such as localized inspections using manual ranging, laser rangefinders, rebound hammers, or ultrasonic testing equipment. While these methods can reflect the strength and surface quality of the piers to some extent, they still have many limitations.

[0003] First, manual inspection methods are limited by the working environment and the skill level of personnel, resulting in low inspection efficiency and poor consistency, making it difficult to complete large-area inspections in a short period of time. Second, the deployment of fixed or single-point inspection equipment is limited, making it impossible to continuously scan the entire surface of the pier, resulting in insufficient spatial coverage of the inspection results. Third, existing inspection systems often lack mobility and cannot move synchronously with crawling equipment, leading to a disconnect between the inspection process and the construction progress. In addition, traditional inspection results often require manual recording and offline analysis, resulting in data lag and untimely defect identification, which cannot meet the needs of modern bridge construction for real-time, accurate, and intelligent inspection.

[0004] Therefore, there is an urgent need for a quality inspection system and method that can work in conjunction with crawling equipment to achieve automatic acquisition and intelligent analysis of multi-dimensional parameters of beams and piers, so as to improve inspection efficiency, coverage and data accuracy, and promote the development of structural inspection towards automation and intelligence. Summary of the Invention

[0005] In view of this, the present invention proposes a quality inspection system and method for mid-beam piers of crawling equipment, aiming to solve the problems of low inspection efficiency, limited coverage, insufficient mobility and lagging data processing in the current technology of mid-beam pier inspection methods, which make it difficult to achieve real-time, comprehensive and intelligent quality inspection of pier structures.

[0006] This invention proposes a method for quality inspection of intermediate beam piers in crawling equipment, comprising:

[0007] Geometric parameters and material properties data of the pier surface are collected using a multi-dimensional sensor array. The collected data is preprocessed to eliminate noise and outliers. A three-dimensional point cloud model of the pier surface is constructed based on the preprocessed data. Key structural regions of the pier are extracted based on the point cloud segmentation algorithm.

[0008] Point cloud data of key structural regions are input into a deep learning network to extract local features. The local features are then input into a recurrent neural network to model time series dependencies and output global feature representations. Two independent fully connected branches are set up, and the global feature representations are input into the normal working condition branch and the extreme working condition branch, respectively, to generate prediction results under the corresponding working conditions.

[0009] Based on the directional information of the surface normal vector of the pier column, the global feature representation is divided into multiple sector regions, and a defect detection curve is constructed in each sector region;

[0010] Substitute the real-time monitoring data of the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas to obtain the detection results of multiple areas. Then, based on the inverse distance weighting strategy, fuse the detection results of adjacent areas to generate a comprehensive detection value.

[0011] The final detection result is output by weighting and fusing the prediction results of the normal working condition branch, the prediction results of the extreme working condition branch, and the comprehensive detection value.

[0012] Furthermore, when collecting geometric parameters and material property data of the pier surface based on the multi-dimensional sensor group mounted on the crawling device, the following are included:

[0013] A laser rangefinder, an ultrasonic probe, and an infrared thermal imager are installed on the crawling equipment to collect the distance distribution, internal density changes, and temperature field distribution on the surface of the pier column, respectively.

[0014] The collected raw data is filtered to remove high-frequency noise caused by environmental interference or equipment vibration.

[0015] The sliding window algorithm is used to perform segmented smoothing of the data and remove outliers that exceed a reasonable range;

[0016] The processed data is stored in a uniform format file, aligned with timestamps.

[0017] Furthermore, a three-dimensional point cloud model of the pier surface is constructed based on the preprocessed data, and key structural regions of the pier are extracted using a point cloud segmentation algorithm, including:

[0018] An initial point cloud model is generated based on distance distribution data collected by a laser rangefinder, and redundant data is reduced by voxelization downsampling.

[0019] The point cloud segmentation algorithm based on normal vector clustering divides the point cloud model into multiple sub-regions and identifies the key structural regions of the pier based on the geometry of the sub-regions.

[0020] Secondary refinement processing is performed on the point cloud data within the key structural region to supplement missing points and optimize the point cloud density distribution.

[0021] The optimized point cloud model is stored as a standard format file.

[0022] Furthermore, when inputting point cloud data of key structural regions into a deep learning network to extract local features, this includes:

[0023] The point cloud data of key structural regions is converted into a two-dimensional grid representation to form a data format suitable for input to deep learning networks;

[0024] A convolutional neural network is used to extract features from two-dimensional grid data, with the kernel size set to 3×3, 5×5, or 7×7.

[0025] Introducing dilated convolution operations into the intermediate layers of the network expands the receptive field and captures a wider range of contextual information;

[0026] Batch normalization layers and activation functions are added after each convolutional layer to suppress the gradient vanishing problem and improve feature representation capabilities.

[0027] The final output is the local feature matrix.

[0028] Furthermore, when inputting local features into a recurrent neural network to model time series dependencies, this includes:

[0029] The local feature matrix is ​​expanded in chronological order to form a time series feature vector;

[0030] Modeling time series feature vectors based on gated recurrent unit networks to capture the changing patterns of features over time;

[0031] An attention mechanism is introduced at the network output to calculate the importance score of the feature at each time step and obtain the attention weight by normalization through the Softmax function.

[0032] The time-step features are weighted and aggregated based on attention weights to generate a global feature representation.

[0033] Furthermore, two independent fully connected branch networks are set up, and the global feature representation is input into the normal operating condition branch and the extreme operating condition branch respectively, including:

[0034] Design two fully connected branch networks, namely the normal operating condition branch and the extreme operating condition branch, each branch containing at least two fully connected layers;

[0035] Adding non-linear activation functions between fully connected layers enhances the network's non-linear fitting ability;

[0036] During the training phase, the corresponding branch is selected based on the working condition label to update the parameters, while the parameters of the other branch are frozen.

[0037] During the prediction phase, the global feature representation is simultaneously input into two branches to generate prediction results for normal operating conditions and prediction results for extreme operating conditions, respectively.

[0038] Furthermore, when dividing the global feature representation into multiple sector regions based on the direction information of the pier surface normal vector, this includes:

[0039] Based on the direction information of the surface normal vector of the pier, the surface of the pier is divided into multiple fan-shaped regions with equal angle widths.

[0040] Within each sector, multiple sub-regions are further divided based on the distribution density of the point cloud data;

[0041] The global feature representation is mapped to each sector region to generate a region feature representation;

[0042] Within each sector, a defect detection curve is constructed based on the relationship between historical monitoring data and actual detection results.

[0043] Furthermore, when substituting real-time monitoring data for the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas, the following steps are taken:

[0044] Determine the target sector area and its adjacent sector areas corresponding to the direction of the normal vector on the pier surface during the target time period;

[0045] Substitute the real-time monitoring data for the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas to determine the detection results for each area;

[0046] The inverse distance weights are determined based on the angle between the center normal vector of each sector region and the target normal vector.

[0047] The detection results of each region are weighted and summed according to the inverse distance weight to generate a comprehensive detection value.

[0048] Furthermore, when weighted and fused with the prediction results of the normal operating condition branch, the prediction results of the extreme operating condition branch, and the comprehensive detection value, the following is included:

[0049] Performance evaluation indicators, including root mean square error and correlation coefficient, are calculated for the branch prediction results of normal working conditions, the branch prediction results of extreme working conditions, and the comprehensive test values.

[0050] Subjective and objective weights are determined based on performance evaluation indicators, and normalization methods are used to constrain the weights.

[0051] The prediction results of the normal working condition branch, the prediction results of the extreme working condition branch, and the comprehensive test value are weighted and summed according to the combined weights to generate the final test result, and the final test result is stored as a standardized report.

[0052] Compared with existing technologies, the advantages of this invention are as follows: By integrating a multi-dimensional sensor group into the crawling device, synchronous acquisition of geometric parameters and material property data of the pier surface is achieved. This enables high-density, continuous data acquisition during device movement, effectively improving detection coverage and spatiotemporal consistency. Preprocessing the acquired data to eliminate noise and outliers further enhances data accuracy and stability, providing a reliable foundation for subsequent modeling and analysis. A three-dimensional point cloud model of the pier surface is constructed based on the preprocessed data, and key structural regions are extracted using point cloud segmentation algorithms, enabling accurate identification of the overall pier shape and local details. Compared with traditional two-dimensional image detection methods, this approach has significant advantages in spatial information preservation and structural fidelity, providing a more comprehensive reflection of the pier's true state. In the feature extraction stage, the combination of deep learning networks and recurrent neural networks not only extracts spatial features of local pier areas but also establishes dependencies between temporal features, thereby improving the model's ability to identify surface defects and structural anomalies under different working conditions. By setting up two independent branch networks for normal and extreme working conditions, adaptive detection can be achieved for different construction or environmental conditions, improving the model's generalization ability and robustness. Finally, based on the normal vector direction information, the global features are divided into multiple fan-shaped regions, and defect detection curves are constructed within these regions. This enables regionalized and refined analysis of the detection results, facilitating the rapid location of potential defect areas. By introducing an inverse distance weighting strategy to fuse the detection results of adjacent regions, not only is the smoothness and continuity of the detection results improved, but the impact of local errors on the overall judgment is also effectively reduced.

[0053] On the other hand, this application also provides a quality inspection system for intermediate beam piers of crawling equipment, comprising:

[0054] The data acquisition module is configured to collect geometric parameters and material properties data of the pier surface through a multi-dimensional sensor group mounted on the crawling device, and to preprocess the collected data.

[0055] The modeling module is configured to construct a 3D point cloud model of the pier surface based on the preprocessed data, and extract key structural regions through a point cloud segmentation algorithm;

[0056] The feature extraction module is configured to input point cloud data of key structural regions into a deep learning network to extract local features, and input the local features into a recurrent neural network to model time series dependencies.

[0057] The load condition branch module is configured to set up two independent fully connected branch networks, inputting the global feature representation into the normal load condition branch and the extreme load condition branch respectively.

[0058] The region division module is configured to divide the global feature representation into multiple sector regions based on the direction information of the normal vector of the pier surface, and to construct a defect detection curve in each sector region;

[0059] The detection fusion module is configured to substitute real-time monitoring data of the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas to generate a comprehensive detection value.

[0060] The results output module is configured to output the final detection result by weighted fusion of the prediction results of the normal working condition branch, the prediction results of the extreme working condition branch, and the comprehensive detection value.

[0061] It is understood that the quality inspection system and method for the middle beam pier column of the crawling equipment in the above embodiments have the same beneficial effects, and will not be described in detail here. Attached Figure Description

[0062] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0063] Figure 1 A flowchart illustrating a method for quality inspection of intermediate beam piers in crawling equipment, provided by an embodiment of the present invention;

[0064] Figure 2 This is a schematic flowchart of a method for quality inspection of intermediate beam piers in crawling equipment, provided by an embodiment of the present invention.

[0065] Figure 3 This is a functional block diagram of a quality inspection system for a mid-beam pier column of a crawling device, provided in an embodiment of the present invention. Detailed Implementation

[0066] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0067] like Figures 1-2 As shown in some embodiments of this application, this embodiment provides a method for quality inspection of the middle beam pier column of a crawling device, including:

[0068] Step S100: Collect geometric parameters and material properties data of the pier surface based on a multi-dimensional sensor group, preprocess the collected data to eliminate noise and outliers, construct a three-dimensional point cloud model of the pier surface based on the preprocessed data, and extract the key structural regions of the pier based on the point cloud segmentation algorithm.

[0069] Specifically, when collecting geometric parameters and material properties data of the pier surface based on the multi-dimensional sensor group mounted on the crawling device, the process includes: installing a laser rangefinder, an ultrasonic probe, and an infrared thermal imager on the crawling device to collect distance distribution, internal density changes, and temperature field distribution on the pier surface, respectively; filtering the collected raw data to remove high-frequency noise caused by environmental interference or equipment vibration; using a sliding window algorithm to perform segmented smoothing of the data and remove outliers that exceed a reasonable range; and storing the processed data in a unified format file aligned with timestamps.

[0070] Specifically, the process of constructing a 3D point cloud model of the pier surface based on the preprocessed data and extracting the key structural regions of the pier using a point cloud segmentation algorithm includes: generating an initial point cloud model based on distance distribution data collected by a laser rangefinder and reducing redundant data through voxelization downsampling; dividing the point cloud model into multiple sub-regions based on a point cloud segmentation algorithm using normal vector clustering, and identifying the key structural regions of the pier based on the geometry of the sub-regions; performing secondary refinement processing on the point cloud data within the key structural regions to supplement missing points and optimize the point cloud density distribution; and storing the optimized point cloud model as a standard format file.

[0071] Understandably, by integrating multiple sensors such as laser rangefinders, ultrasonic probes, and infrared thermal imagers into the crawling device, simultaneous acquisition of multi-dimensional information such as geometric morphology, internal density, and temperature distribution is achieved. The laser rangefinder can accurately acquire spatial coordinate data of the pier surface, providing a foundation for 3D geometric reconstruction; the ultrasonic probe is used to identify internal density changes, thereby reflecting potential structural defects such as cracks and voids; and the infrared thermal imager is used to detect temperature anomalies, assisting in judging the uniformity of internal materials and potential deterioration areas. At the data processing level, this invention uses filtering and sliding window algorithms to perform multi-stage preprocessing of the acquired data, effectively eliminating high-frequency interference caused by factors such as equipment vibration and environmental noise, while also removing abnormal data to ensure the stability and reliability of the input data. The preprocessed multi-source data is synchronized with timestamps and standardized in format, achieving unified management of spatial information and time series, providing a precise spatiotemporal alignment foundation for subsequent point cloud modeling. In the 3D modeling stage, this invention constructs an initial point cloud model of the pier surface based on laser rangefinder data and uses voxelization downsampling technology to reduce redundant points, thereby improving data processing efficiency. Subsequently, a point cloud segmentation algorithm based on normal vector clustering was introduced to regionalize the point cloud. This algorithm analyzes the directional consistency and spatial neighborhood relationships of the point cloud normal vectors, subdividing the overall pier model into multiple geometrically significant sub-regions, and further identifying key structural parts of the pier (such as transition sections, connecting sections, and stress concentration areas). Finally, by performing secondary refinement and missing point compensation on the point cloud data of the key structural regions, adaptive optimization of point cloud density and enhancement of model integrity are achieved, thereby obtaining a high-fidelity 3D structural representation of the pier. This technology, by integrating multi-source sensing, noise suppression, point cloud segmentation, and refined modeling, constructs a detection foundation model that combines geometric accuracy with physical feature representation capabilities, providing reliable technical support for subsequent defect identification and structural analysis.

[0072] As can be seen, in practical implementation, the method for quality inspection of pier columns in bridge construction sites provided by this invention can be applied to pier column inspection tasks. The crawling device moves at a constant speed along the pier column surface. The laser rangefinder, ultrasonic probe, and infrared thermal imager integrated on the device work together to achieve multi-dimensional information acquisition of the pier column surface and interior. For example, when the device moves to the middle of the pier column, the laser rangefinder scans the surface contour in real time to obtain distance distribution data; the ultrasonic probe simultaneously emits sound waves into the concrete to detect areas of abnormal density; and the infrared thermal imager captures surface temperature distribution images to identify potential hollow or cracked areas.

[0073] The collected multi-source raw data is filtered in real time by an embedded processing unit to remove high-frequency noise signals caused by construction site noise, electromagnetic interference, or equipment micro-vibration. The system then uses a sliding window algorithm to perform segmented smoothing on the continuously sampled data to ensure stable curve trends, while removing outliers that deviate from the normal range in the detection results. The processed data is aligned and uniformly encoded according to timestamps and finally stored in a standardized format file for easy subsequent modeling and analysis.

[0074] During the data modeling phase, the system uses distance distribution data collected by a laser rangefinder to generate an initial 3D point cloud model of the pier surface. For example, a high-precision point cloud set containing millions of spatial points can be formed in a single scanning cycle. To reduce data volume and improve computational efficiency, the system first performs voxelization downsampling on the point cloud data, compressing dense point clusters into regular voxel units. Next, the system applies a point cloud segmentation algorithm based on normal vector clustering to analyze the normal vector direction and spatial neighborhood relationship of each point, dividing the point cloud into multiple sub-regions with geometric features, such as the main surface of the pier, edge transition areas, and bottom connecting sections.

[0075] Furthermore, for the identified key structural areas (such as the transition connection between piers and tie beams), the system performs secondary refinement of the point cloud data. Missing points are recovered through interpolation and density redistribution algorithms, improving the morphological continuity and detail fidelity of these areas. The final high-precision point cloud model is output in a standard file format and can be directly used for subsequent structural analysis, defect identification, or construction quality assessment.

[0076] This example demonstrates that the present invention not only achieves automated multidimensional detection in complex construction environments, but also generates structurally complete and high-resolution three-dimensional digital models, providing an operable technical path for the accurate detection and digital management of central beam piers and columns.

[0077] Step S200: Input the point cloud data of the key structural region into the deep learning network to extract local features, and input the local features into the recurrent neural network to model the time series dependency, output the global feature representation, and set up two independent fully connected branch networks, inputting the global feature representation into the normal working condition branch and the extreme working condition branch respectively to generate the prediction results under the corresponding working conditions.

[0078] Specifically, when inputting point cloud data of key structural regions into a deep learning network to extract local features, the process includes: converting the point cloud data of key structural regions into a two-dimensional grid representation to form a data format suitable for input to the deep learning network; using a convolutional neural network to extract features from the two-dimensional grid data, where the convolutional kernel size is set to 3×3, 5×5, or 7×7; introducing dilated convolution operations in the intermediate layers of the network to expand the receptive field and capture a wider range of contextual information; adding batch normalization layers and activation functions after each convolutional layer to suppress the gradient vanishing problem and improve feature representation capabilities; and finally outputting a local feature matrix.

[0079] Specifically, when inputting local features into a recurrent neural network to model time series dependencies, the process includes: expanding the local feature matrix in chronological order to form a time series feature vector; modeling the time series feature vector based on a gated recurrent unit network to capture the changing patterns of features over time; introducing an attention mechanism at the network output to calculate the importance score of each feature at each time step and normalizing it using the Softmax function to obtain the attention weights; and weighting and converging the time step features according to the attention weights to generate a global feature representation.

[0080] Specifically, setting up two independent fully connected branch networks and inputting the global feature representation into the normal operating condition branch and the extreme operating condition branch respectively includes: designing two fully connected branch networks, namely the normal operating condition branch and the extreme operating condition branch, each branch containing at least two fully connected layers; adding non-linear activation functions between the fully connected layers to enhance the non-linear fitting ability of the network; during the training phase, selecting the corresponding branch for parameter updates based on the operating condition label and freezing the parameters of the other branch; during the prediction phase, inputting the global feature representation into both branches simultaneously to generate prediction results for normal operating conditions and prediction results for extreme operating conditions respectively.

[0081] Understandably, high-precision quality prediction under complex working conditions can be achieved by extracting and fusing multi-level features from point cloud data of key structural areas of piers. This is achieved by organically integrating spatial structural features with temporal evolution features, enabling the model to capture subtle local anomalies on the pier surface and understand the dynamic changes in structural response under different working conditions. Firstly, in the local feature extraction stage, this invention solves the problem of traditional neural networks' inability to directly process irregular point cloud data by converting three-dimensional point cloud data into a two-dimensional grid representation. A convolutional neural network (CNN) is used for multi-scale convolution operations, utilizing convolutional kernels of different sizes (such as 3×3, 5×5, and 7×7) to achieve feature extraction from details to the whole. Dilated convolution is introduced in the intermediate layers of the network, which expands the receptive field without increasing computational cost, thereby capturing a wider range of spatial contextual information. Simultaneously, the combination of batch normalization and nonlinear activation functions (such as ReLU) ensures stable network training and efficient feature representation. The technical principles described here ensure the model's sensitive ability to identify geometric changes, microcracks, and local abnormal textures on the pier surface. Secondly, in the time-dependent modeling stage, this invention employs a recurrent neural network structure based on gated recurrent units (GRUs), unfolding local features into a time-series input in chronological order, thereby establishing a dynamic dependency relationship between structural features and time. GRUs effectively solve the gradient vanishing problem in long-sequence training, ensuring the model's ability to remember long-term temporal features. Furthermore, an attention mechanism is introduced at the network output. By calculating the importance score of features at each time step and performing Softmax normalization, the model can automatically focus on key time segments, highlighting the most significant features of structural state changes. This mechanism enables the detection system to have higher sensitivity and accuracy in judging sudden anomalies or changes in working conditions under complex environments. Finally, in the working condition branch prediction stage, this invention sets up two independent fully connected networks for ordinary working condition branches and extreme working condition branches to achieve differentiated modeling of pier quality under different environments or stress states. During the training phase, the two branches update parameters separately based on the operating condition labels, thereby forming a targeted feature mapping capability. During the prediction phase, they run simultaneously, providing a more comprehensive state assessment basis for the system by comparing the prediction results under the two operating conditions. The principle behind this parallel branch structure lies in utilizing the idea of ​​multi-task learning to achieve robust fusion of operating condition adaptation and prediction results.

[0082] It can be seen that the quality inspection method for tie beams and piers can be applied to automated inspection at bridge construction sites. After the crawling device collects point cloud data of key structural areas of the piers and piers, the original three-dimensional point cloud data is first converted into a two-dimensional mesh representation to adapt the data format to the input of a deep learning network. For example, when inspecting the transition connection area between the pier and the tie beam, the point cloud data is projected and interpolated to form a two-dimensional grayscale matrix, where the pixel grayscale values ​​correspond to point cloud density or reflection intensity. Subsequently, a convolutional neural network (CNN) performs multi-scale convolution operations on this two-dimensional mesh, with convolutional kernels of different sizes (3×3, 5×5, 7×7) used to capture multi-level features such as crack details, surface depressions, and overall contours. After introducing dilated convolutions in the intermediate layers, the network can expand its receptive field without increasing computational load, thereby extracting a wider range of spatial structural information. After each convolutional layer, batch normalization layers and activation functions are used to standardize and nonlinearly enhance the features, ultimately generating a feature matrix that reflects local geometric features and surface anomaly distribution.

[0083] Next, the local feature matrix is ​​expanded according to the time series and input into a gated recurrent unit (GRU) network for temporal modeling. For example, when the crawling device moves continuously along the pier surface, the point cloud features collected at different times reflect the dynamic changes in the structural state. The GRU retains historical feature information through a gating mechanism, capturing the evolution trend of the pier surface state over time, thereby enabling the identification of potential changes such as structural aging and crack propagation. At the network output, an attention mechanism is introduced, assigning different weights to the features at each time step, allowing the network to focus on time segments with significant abnormal changes. After weighted aggregation, a global feature representation is output, which integrates spatial morphology and temporal change information, providing a high-dimensional semantic description for working condition classification and defect prediction.

[0084] During the prediction phase, the global feature representation is simultaneously input into two independent fully connected branch networks, corresponding to the normal operating condition branch and the extreme operating condition branch, respectively. For example, when the detected ambient temperature and humidity are within the normal range, the model automatically uses the normal operating condition branch to output a normal state prediction; while when the equipment detects drastic changes in the external environment or abnormal structural stress, prediction is performed through the extreme operating condition branch. Each branch contains multiple fully connected layers and nonlinear activation functions for deep fitting and classification of complex features. During training, the corresponding branch is selected for parameter updates based on different operating condition labels to ensure that the model can achieve targeted optimization under various operating conditions; during prediction, both branches run simultaneously, outputting dual-channel prediction results to provide a more comprehensive basis for pier status assessment.

[0085] For example, in a field application, when inspecting the tie beam piers of a bridge, the normal working condition branch outputs results showing intact surfaces and uniform material density, while the extreme working condition branch identifies areas with abnormally high temperatures over a certain period. Combined with infrared thermal imaging information, this indicates a potential risk of cracking at these locations. In this way, the present invention achieves intelligent working condition identification and defect prediction based on multi-source feature fusion and temporal analysis, significantly improving the real-time performance and accuracy of pier quality inspection.

[0086] Step S300: Based on the direction information of the surface normal vector of the pier column, the global feature representation is divided into multiple fan-shaped regions, and a defect detection curve is constructed in each fan-shaped region.

[0087] Specifically, when dividing the global feature representation into multiple sector regions based on the direction information of the surface normal vector of the pier, the process includes: dividing the pier surface into multiple sector regions of equal angle width according to the direction information of the surface normal vector of the pier; further dividing each sector region into multiple sub-regions according to the distribution density of the point cloud data; mapping the global feature representation to each sector region to generate a regional feature representation; and constructing a defect detection curve based on the relationship between historical monitoring data and actual detection results within each sector region.

[0088] Step S400: Substitute the real-time monitoring data of the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas to obtain the detection results of multiple areas, and fuse the detection results of adjacent areas based on the inverse distance weighting strategy to generate a comprehensive detection value.

[0089] Specifically, when substituting the real-time monitoring data of the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas, the process includes: determining the target sector area and its adjacent sector areas corresponding to the normal vector direction of the pier surface within the target time period; substituting the real-time monitoring data of the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas to determine the detection result for each area; determining the inverse distance weight based on the angle between the center normal vector of each sector area and the target normal vector; and weighting and summing the detection results of each area according to the inverse distance weight to generate a comprehensive detection value.

[0090] It can be seen that by analyzing the distribution characteristics of the normal vector on the pier surface, the overall surface is divided into multiple fan-shaped regions with equal angles and widths, ensuring that the detection and analysis of different regions correspond to the spatial geometric directions of the pier surface. This partitioning method not only guarantees uniform spatial coverage of the detection data but also fully reflects the feature differences in different directions of the pier surface. Secondly, within each fan-shaped region, sub-regions are further divided based on the distribution density of the point cloud data, and the global feature representation is mapped to these regions to obtain regional feature representations. By establishing defect detection curves based on historical monitoring data and actual detection results within each fan-shaped region, regional adaptive learning of the detection model is achieved, enabling the system to independently optimize its detection and discrimination capabilities according to the feature patterns of different regions. Finally, in the detection phase, the system substitutes the real-time monitoring data of the target time period into the defect detection curves corresponding to the target fan-shaped region and its adjacent regions to obtain detection results for multiple regions. Subsequently, based on the angle between the center normal vector of each fan-shaped region and the target normal vector, the inverse distance weight is calculated, and the detection results of multiple regions are weighted and fused. This inverse distance weighting strategy ensures the continuous transition and smoothness of the detection results in space, effectively reducing abrupt errors caused by regional boundaries, thereby generating a comprehensive detection value with spatial consistency.

[0091] It can be seen that during the construction or maintenance of bridges, defects such as cracks, peeling, and bubbles on the outer surface of the piers are detected and evaluated in real time using intelligent technology.

[0092] For example, when inspecting a cylindrical bridge pier, the inspection first uses a point cloud scanning device to acquire three-dimensional structural data of the pier surface and calculates the normal vector information of each point on the surface. Based on the directional distribution of these normal vectors, the outer surface of the pier is divided into several equiangular sector regions (e.g., one region every 10°) to ensure the spatial uniformity and geometric correspondence of the inspection area.

[0093] Next, within each sector, several sub-regions are divided based on the distribution of point cloud density. For example, smaller sub-regions are set for areas with large surface curvature or dense point cloud distribution to improve detection accuracy. Global feature representations (such as reflection intensity, surface roughness, grayscale gradient, etc.) are mapped onto these regions to form feature expression vectors for each region. Subsequently, based on previous monitoring data (such as information on cracks or spalling at the same location in historical inspections) and manual verification results, a "defect detection curve" is established within each sector. This curve describes the correspondence between the feature values ​​of that region and the probability of defect occurrence.

[0094] In actual inspection, when the inspection equipment collects real-time monitoring data at a certain moment (i.e., the target time period), it first determines the target sector area and its adjacent sector areas based on the current posture and scanning direction of the pier. For example, when the inspection equipment is pointed due east at the pier, the corresponding target sector area might be area number 9, and the adjacent areas would be areas number 8 and 10. Subsequently, the real-time feature data collected at that moment is substituted into the defect detection curves of these areas to obtain the inspection results for each area, such as defect probability or damage score.

[0095] Finally, based on the angle between the center normal vector of each sector region and the target detection direction, the inverse distance weight is calculated, and the detection results of multiple regions are weighted and fused. For example, if the angle between the normal vector of the target region and the normal vector of the adjacent region is small, the detection result of the adjacent region will be assigned a higher weight. Through this weighting strategy, the generated comprehensive detection value can not only reflect the true state of the target region, but also compensate for the boundary errors caused by spatial partitioning, thus forming a continuous and smooth defect distribution curve on the pier surface.

[0096] Step S500: Based on the weighted fusion of the branch prediction results for normal working conditions, the branch prediction results for extreme working conditions, and the comprehensive detection value, the final detection result is output.

[0097] Specifically, when weighting and fusing the prediction results of the ordinary working condition branch, the prediction results of the extreme working condition branch, and the comprehensive test value, the process includes: calculating performance evaluation indicators, including root mean square error and correlation coefficient, for the prediction results of the ordinary working condition branch, the prediction results of the extreme working condition branch, and the comprehensive test value; determining subjective and objective weights based on the performance evaluation indicators and constraining the weights using a normalization method; weighting and summing the prediction results of the ordinary working condition branch, the prediction results of the extreme working condition branch, and the comprehensive test value according to the combined weights to generate the final test result, and storing the final test result as a standardized report.

[0098] Understandably, the model branches under different working conditions exhibit varying response characteristics to the pier structure. The ordinary working condition branch primarily reflects the structural state of the pier under normal environmental conditions (such as stable temperature and humidity, and relatively low load); the extreme working condition branch characterizes the changes in structural performance under special environmental conditions (such as heavy rain, strong earthquakes, and heavy impact); and the comprehensive detection value, derived from the aforementioned sector defect curve and the spatial weighted result generated by the inverse distance weighting strategy, represents the overall reflection of actual observation data. Therefore, fusing these three can simultaneously consider both the model's predictive ability and the reliability of real observations. Secondly, before fusion, to ensure the credibility and rationality of each result, performance evaluation indicators were calculated for the ordinary working condition prediction results, the extreme working condition prediction results, and the comprehensive detection value. The two core indicators used here are the root mean square error (RMSE) and the correlation coefficient (R). The former measures the magnitude of the deviation between the predicted value and the actual detection value, reflecting the model's accuracy; the latter measures the degree of linear correlation between the predicted value and the actual value, reflecting the model's stability and consistency. Through these two indicators, the reliability of different working condition branches and the comprehensive detection results can be quantitatively assessed. Next, subjective and objective weights are determined based on the aforementioned performance indicators. Subjective weights can be set according to expert experience, the importance of the operating conditions, or historical evaluation results to reflect human preferences; objective weights are adaptively determined based on the calculation results of the performance indicators, dynamically reflecting the performance of different model branches in specific scenarios. A normalization constraint method is used to normalize the subjective and objective weights, ensuring that the sum of all weights is 1, preventing any one branch from excessively dominating the fusion result. Finally, based on the normalized combined weights, the prediction results of the ordinary operating condition branch, the prediction results of the extreme operating condition branch, and the comprehensive detection value are weighted and summed to generate the final detection result. This fusion process not only integrates multi-source information at the numerical level but also achieves multi-dimensional coupling from operating condition characteristics to structural response at the model level. The generated final detection result is then stored as a standardized report for easy data tracking, trend analysis, and maintenance decisions.

[0099] As can be seen, by calculating performance evaluation indicators (such as root mean square error and correlation coefficient) for the prediction results of the normal working condition branch, the prediction results of the extreme working condition branch, and the comprehensive detection value, this method can quantify the prediction performance of different data sources or model branches, and achieve accurate evaluation of the reliability of the detection results. Compared with the traditional direct weighted average or fixed weight fusion method, this method can dynamically adjust the weight allocation based on performance, making the weight allocation more scientific and reasonable, thereby improving the accuracy and credibility of the final detection results. Secondly, the introduction of a collaborative adjustment mechanism of subjective and objective weights helps to achieve a balance between model data-driven and expert experience-guided approaches. Subjective weights can reflect the importance of actual working conditions or historical experience, while objective weights are automatically calculated by model performance indicators, ensuring that the fusion strategy has adaptability and flexibility in different detection environments. Through the normalization constraint method, the weight of a certain branch result is avoided from being too large or too small, ensuring the stability and numerical convergence of the fusion process, thereby improving the overall robustness and generalization ability of the system. Furthermore, by employing a weighted summation of prediction results for ordinary working conditions, extreme working conditions, and comprehensive detection values ​​using combined weights, multi-source detection information can be fully integrated. This compensates for the prediction bias of single-branch models under specific working conditions, enabling comprehensive judgment across multiple working conditions and scales. This fusion method maintains high detection stability and reliability even under complex construction conditions or sudden anomalies, significantly improving the system's practicality and intelligence. Finally, by storing the fused final detection results as a standardized report, not only is the format of the detection results unified and standardized management achieved, but it also facilitates subsequent trend analysis, historical data comparison, and maintenance decisions. This standardized output mechanism allows the detection results to be directly connected to the quality management system or engineering monitoring database, improving data utilization efficiency and project traceability.

[0100] In the above embodiments, by integrating a multi-dimensional sensor group into the crawling device, synchronous acquisition of geometric parameters and material property data of the surface of the central beam pier is achieved. This enables high-density, continuous data acquisition during device movement, effectively improving detection coverage and spatiotemporal consistency. Preprocessing the acquired data to eliminate noise and outliers further enhances the accuracy and stability of the data, providing a reliable foundation for subsequent modeling and analysis. A three-dimensional point cloud model of the pier surface is constructed based on the preprocessed data, and key structural regions are extracted using point cloud segmentation algorithms, enabling accurate identification of the overall morphology and local details of the pier. Compared to traditional two-dimensional image detection methods, this approach has significant advantages in spatial information preservation and structural fidelity, providing a more comprehensive reflection of the pier's true state. In the feature extraction stage, the combination of deep learning networks and recurrent neural networks not only extracts spatial features of local pier areas but also establishes dependencies between temporal features, thereby improving the model's ability to identify surface defects and structural anomalies under different working conditions. By setting two independent branch networks for normal and extreme working conditions, adaptive detection for different construction or environmental conditions can be achieved, improving the model's generalization ability and robustness. Finally, based on the normal vector direction information, the global features are divided into multiple sector regions, and defect detection curves are constructed within these regions. This enables regionalized and refined analysis of the detection results, facilitating the rapid location of potential defect areas. By introducing an inverse distance weighting strategy to fuse detection results from adjacent regions, not only is the smoothness and continuity of the detection results improved, but the impact of local errors on the overall judgment is also effectively reduced.

[0101] In another preferred embodiment based on the above embodiments, such as Figure 2 As shown, this embodiment provides a quality inspection system for intermediate beam piers of crawling equipment, including:

[0102] The data acquisition module is configured to collect geometric parameters and material properties data of the pier surface through a multi-dimensional sensor group mounted on the crawling device, and to preprocess the collected data.

[0103] The modeling module is configured to construct a 3D point cloud model of the pier surface based on the preprocessed data, and extract key structural regions through a point cloud segmentation algorithm;

[0104] The feature extraction module is configured to input point cloud data of key structural regions into a deep learning network to extract local features, and input the local features into a recurrent neural network to model time series dependencies.

[0105] The load condition branch module is configured to set up two independent fully connected branch networks, inputting the global feature representation into the normal load condition branch and the extreme load condition branch respectively.

[0106] The region division module is configured to divide the global feature representation into multiple sector regions based on the direction information of the normal vector of the pier surface, and to construct a defect detection curve in each sector region;

[0107] The detection fusion module is configured to substitute real-time monitoring data of the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas to generate a comprehensive detection value.

[0108] The results output module is configured to output the final detection result by weighted fusion of the prediction results of the normal working condition branch, the prediction results of the extreme working condition branch, and the comprehensive detection value.

[0109] It is understood that the quality inspection system and method for the middle beam pier column of the crawling equipment in the above embodiments have the same beneficial effects, and will not be described in detail here.

[0110] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0111] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0112] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0113] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for quality inspection of intermediate beam piers in crawling equipment, characterized in that, include: Geometric parameters and material properties data of the pier surface are collected using a multi-dimensional sensor array. The collected data is preprocessed to eliminate noise and outliers. A three-dimensional point cloud model of the pier surface is constructed based on the preprocessed data. Key structural regions of the pier are extracted based on the point cloud segmentation algorithm. Point cloud data of key structural regions are input into a deep learning network to extract local features. The local features are then input into a recurrent neural network to model time series dependencies and output global feature representations. Two independent fully connected branches are set up, and the global feature representations are input into the normal working condition branch and the extreme working condition branch, respectively, to generate prediction results under the corresponding working conditions. Based on the directional information of the surface normal vector of the pier column, the global feature representation is divided into multiple sector regions, and a defect detection curve is constructed in each sector region; Substitute the real-time monitoring data of the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas to obtain the detection results of multiple areas. Then, based on the inverse distance weighting strategy, fuse the detection results of adjacent areas to generate a comprehensive detection value. The final detection result is output by weighted fusion of the branch prediction results for normal working conditions, the branch prediction results for extreme working conditions, and the comprehensive detection value. When collecting geometric parameters and material property data of the pier surface using a multi-dimensional sensor array mounted on the crawling device, the following are included: A laser rangefinder, an ultrasonic probe, and an infrared thermal imager are installed on the crawling equipment to collect data on the surface of the pier column. The distance distribution, internal density variation, and temperature field distribution; The collected raw data is filtered to remove high-frequency noise caused by environmental interference or equipment vibration; a sliding window algorithm is used to smooth the data in segments and remove outliers that exceed the reasonable range. The processed data is stored in a uniform format file aligned with timestamps. When constructing a 3D point cloud model of the pier surface based on the preprocessed data, and extracting the key structural regions of the pier using a point cloud segmentation algorithm, the following steps are taken: An initial point cloud model is generated based on distance distribution data collected by a laser rangefinder, and redundant data is reduced by voxelization downsampling. The point cloud segmentation algorithm based on normal vector clustering divides the point cloud model into multiple sub-regions and identifies the key structural regions of the pier based on the geometry of the sub-regions. Secondary refinement processing is performed on the point cloud data within the key structural region to supplement missing points and optimize the point cloud density distribution. Save the optimized point cloud model as a standard format file; When inputting point cloud data of key structural regions into a deep learning network to extract local features, the following are included: The point cloud data of key structural regions is converted into a two-dimensional grid representation to form a data format suitable for input to deep learning networks; A convolutional neural network is used to extract features from two-dimensional grid data, with the kernel size set to 3×3, 5×5 or 7×7. Introducing dilated convolution operations into the intermediate layers of the network expands the receptive field and captures a wider range of contextual information; Batch normalization layers and activation functions are added after each convolutional layer to suppress the gradient vanishing problem and improve feature representation capabilities. The final output is the local feature matrix; When using local features as input to a recurrent neural network to model time series dependencies, the following are included: The local feature matrix is ​​expanded in chronological order to form a time series feature vector; Modeling time series feature vectors based on gated recurrent unit networks to capture the changing patterns of features over time; An attention mechanism is introduced at the network output to calculate the importance score of the feature at each time step, and then the score is calculated using a Softmax function. Number normalization yields attention weights; The time-step features are weighted and aggregated according to attention weights to generate a global feature representation; Two independent fully connected branch networks are configured, and the global feature representations are input into the normal operating condition branch and the extreme operating condition branch respectively, including: Design two fully connected branch networks, namely the normal operating condition branch and the extreme operating condition branch, each branch containing at least two fully connected layers; Adding non-linear activation functions between fully connected layers enhances the network's non-linear fitting ability; During the training phase, the corresponding branch is selected based on the working condition label to update the parameters, while the parameters of the other branch are frozen. During the prediction phase, the global feature representation is simultaneously input into two branches to generate prediction results for normal operating conditions and prediction results for extreme operating conditions, respectively. When the global feature representation is divided into multiple sector regions based on the direction information of the surface normal vector of the pier, it includes: Based on the direction information of the surface normal vector of the pier, the surface of the pier is divided into multiple fan-shaped regions with equal angle widths. Within each sector, multiple sub-regions are further divided based on the distribution density of the point cloud data; The global feature representation is mapped to each sector region to generate a region feature representation; Within each sector, a defect detection curve is constructed based on the relationship between historical monitoring data and actual detection results.

2. The method for quality inspection of intermediate beam piers for crawling equipment as described in claim 1, characterized in that, When substituting real-time monitoring data for the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas, the following applies: Determine the target sector area and its adjacent sector areas corresponding to the direction of the normal vector on the pier surface during the target time period; Substitute the real-time monitoring data for the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas to determine the detection results for each area; The inverse distance weights are determined based on the angle between the center normal vector of each sector region and the target normal vector. The detection results of each region are weighted and summed according to the inverse distance weight to generate a comprehensive detection value.

3. The method for quality inspection of intermediate beam piers for crawling equipment as described in claim 1, characterized in that, When weighted and fused with the prediction results of normal operating conditions, the prediction results of extreme operating conditions, and the comprehensive detection value, the following are included: Performance evaluation indicators, including root mean square error and correlation coefficient, are calculated for the branch prediction results of normal working conditions, the branch prediction results of extreme working conditions, and the comprehensive test values. Subjective and objective weights are determined based on performance evaluation indicators, and normalization methods are used to constrain the weights. The prediction results of the normal working condition branch, the prediction results of the extreme working condition branch, and the comprehensive test value are weighted and summed according to the combined weights to generate the final test result, and the final test result is stored as a standardized report.

4. A quality inspection system for intermediate beam piers of crawling equipment, applicable to the quality inspection method for intermediate beam piers of crawling equipment as described in any one of claims 1-3, characterized in that, include: The data acquisition module is configured to collect geometric parameters and material properties data of the pier surface through a multi-dimensional sensor group mounted on the crawling device, and to preprocess the collected data. The modeling module is configured to construct a 3D point cloud model of the pier surface based on the preprocessed data, and extract key structural regions through a point cloud segmentation algorithm; The feature extraction module is configured to input point cloud data of key structural regions into a deep learning network to extract local features, and input the local features into a recurrent neural network to model time series dependencies. The load condition branch module is configured to set up two independent fully connected branch networks, inputting the global feature representation into the normal load condition branch and the extreme load condition branch respectively. The region division module is configured to divide the global feature representation into multiple sector regions based on the direction information of the normal vector of the pier surface, and to construct a defect detection curve in each sector region; The detection fusion module is configured to substitute real-time monitoring data of the target time period into the defect detection curves corresponding to the target sector area and its adjacent areas to generate a comprehensive detection value. The results output module is configured to output the final detection result by weighted fusion of the prediction results of the normal working condition branch, the prediction results of the extreme working condition branch, and the comprehensive detection value.

Images