Method and system for predicting potential damage of large-aperture corrugated pipe culvert based on multi-source isomerism

Abstract

The invention relates to the technical field of large-aperture corrugated pipe culvert potential damage prediction, in particular to a large-aperture corrugated pipe culvert potential damage prediction method and system based on multi-source isomerism. The method comprises the following steps: preprocessing acquired multi-source heterogeneous data; constructing a large-aperture corrugated pipe culvert potential damage prediction model based on multi-source isomerism; performing feature extraction on the data by using a multi-source heterogeneous large-aperture corrugated pipe culvert potential damage prediction model, and training the model based on the extracted features; and performing prediction by using the trained model. Through fusion of the fiber grating technology and the Internet of Things technology, real-time acquisition and intelligent analysis of parameters such as structural strain, displacement and soil pressure in the construction period are realized, and the damage identification rate reaches 98% or above.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of potential damage prediction of large-aperture corrugated pipe culvert, and particularly relates to a potential damage prediction method and system of large-aperture corrugated pipe culvert based on multi-source heterogeneity. BACKGROUND

[0002] In highway engineering, culvert engineering accounts for about 60%-70% of the total number of bridge and culvert, and the engineering cost accounts for 40% of the total amount of bridge and culvert. The construction quality directly affects the safety and performance of the highway. The uneven settlement of the culvert is the main cause of structural damage. In particular, in the areas of permafrost, soft soil, expansive soil, collapsible loess and other poor engineering geotechnical areas, the culverts constructed by traditional concrete masonry structures are prone to cracking, dislocation, leakage and other damage problems, and the maintenance cost is high and there are hidden dangers of driving safety. As a flexible structure, the steel corrugated pipe culvert has the following advantages: strong adaptability to foundation deformation, which can alleviate the damage caused by uneven settlement of rigid foundation; axial corrugation design can disperse load stress concentration and fully utilize the high strength characteristics of steel structure; it can be standardized in design and production, and assembled on site, with short construction period and convenience; it can improve the "dislocation and bumping" phenomenon and meet the concept of low carbon and environmental protection (reduce the use of cement and gravel). However, the existing steel corrugated pipe culvert mainly focuses on medium and small aperture of 0.5-5m, and faces many technical bottlenecks when developing towards large corrugation, large aperture (such as 6m and above) and thin wall. In addition, the design theory of large-aperture corrugated pipe culvert is not perfect. Unlike medium and small aperture steel corrugated pipe culvert, the material nonlinearity and geometric nonlinearity of large corrugation, large aperture and thin wall steel corrugated pipe are more significant. The traditional ring pressure theory based on elastic assumption is not suitable for the design of large-size steel corrugated pipe culvert, and the imperfect design method and related specification seriously restrict the construction and application of large-span and large-size corrugated pipe culvert.

[0003] The steel corrugated pipe belongs to a thin-walled flexible structure, and the larger the size, the greater the structural flexibility coefficient, and the more sensitive to deformation during construction. Large disturbance during construction may cause buckling deformation and damage of the structure, and quality control is difficult. It is particularly important to accurately monitor and reasonably evaluate the mechanical parameters of the weak link of the steel corrugated pipe culvert during the construction process such as assembly and backfill.

[0004] The large-aperture thin-walled steel corrugated pipe culvert has many assembly segments and high installation precision requirements. The assembly error during construction will accumulate with the increase of segments, and excessive segment assembly error will also affect the waterproof performance of the pipe culvert. How to improve the assembly precision of the large-aperture thin-walled steel corrugated pipe culvert through construction control is a key technical problem. In order to solve the above-mentioned problems, the present application provides a potential damage prediction method and system of large-aperture corrugated pipe culvert based on multi-source heterogeneity.

[0005] The first aspect of the present application provides a potential damage prediction method for a large-aperture corrugated pipe culvert based on multi-source heterogeneity. A potential damage prediction method for a large-aperture corrugated pipe culvert based on multi-source heterogeneity, comprising: Multi-source heterogeneous data collaborative collection; Pretreatment of the obtained multi-source heterogeneous data; Construction of a potential damage prediction model for a large-aperture corrugated pipe culvert based on multi-source heterogeneity; Feature extraction of data using the potential damage prediction model for a large-aperture corrugated pipe culvert based on multi-source heterogeneity, and training of the model based on the extracted features; Prediction using the trained model.

[0006] Further, the multi-source heterogeneous data collaborative collection comprises arranging fiber Bragg strain sensors and inclination sensors in the axial, circumferential and key splicing joints of the corrugated pipe culvert, burying LoRa wireless pressure sensors in the soil, collecting strain, displacement and soil pressure data; using a total station + three-dimensional laser scanning joint positioning, obtaining point cloud data through a spiral scanning path, and detecting the splicing joint based on binocular vision; using an adjustable frequency seismic source and ground penetrating radar for detection, and using IEEE1588 clock protocol to realize multi-source data synchronization.

[0007] Further, the pretreatment of the obtained multi-source heterogeneous data comprises filtering and denoising of fiber sensing data, data calibration and outlier processing, wherein the fiber Bragg strain sensors, inclination sensors and the like are calibrated periodically using a standard pressure source and an inclination reference, and the sensor measurement error is corrected by fitting a calibration curve through the least squares method; and the collected strain and soil pressure data are subjected to outlier identification, and abnormal data points caused by construction interference or sensor failure are removed.

[0008] Further, the construction of the potential damage prediction model for a large-aperture corrugated pipe culvert based on multi-source heterogeneity comprises using a Transformer architecture as a basic model, combining attention mechanism and multi-modal fusion mechanism to realize prediction of potential damage of a large-aperture corrugated pipe culvert, wherein the Transformer captures long-distance dependencies in the sequence by calculating the correlation of each element in the input sequence with other elements, first maps the input feature vector to a query, key and value three vector space, then calculates the dot product similarity of the query vector and the key vector, and normalizes it through the Softmax function to obtain the attention weight.

[0009] Further, the construction of the potential damage prediction model based on multi-source heterogeneous large-aperture corrugated pipe culvert includes introducing a multi-modal fusion mechanism in the Transformer architecture to capture the interaction between different modal data. The multi-modal fusion mechanism is composed of a cross-attention layer and a feature fusion layer. In the cross-attention layer, information interaction between features of different modalities is allowed. The difference between the calculation process and the self-attention mechanism is that the query vector comes from one modality, and the key vector and the value vector come from another modality. In the feature fusion layer, multi-layer perception (MLP) and residual connection are used to fuse features of different modalities to obtain comprehensive feature representation.

[0010] Further, the feature extraction of the data using the potential damage prediction model based on multi-source heterogeneous large-aperture corrugated pipe culvert includes using a feature-level fusion method to extract features from preprocessed multi-source heterogeneous data, and then fusing the extracted features to construct a comprehensive feature vector. For strain data, the extracted features include mean, standard deviation, maximum value, minimum value, peak frequency, and energy entropy. For inclination data, the extracted features include inclination rate of change, inclination fluctuation range, and inclination trend term. For soil pressure data, the extracted features include soil pressure mean, soil pressure gradient, and soil pressure change rate. For point cloud data, the extracted features include point cloud density, point cloud curvature, joint width, and joint offset. For seismic wave and ground penetrating radar data, the extracted features include wave speed, amplitude, frequency, and reflection coefficient.

[0011] Further, the training of the model based on the extracted features includes training the model in stages based on a multi-task joint training strategy by constructing a loss function system and a Ranger optimization algorithm. In the pure data-driven training, normal operating condition data during construction are used and a first learning rate of 1e4 is set to make the model learn basic feature representation. In the mechanism data fusion training, buckling theory and steel-soil interaction mechanism constraints are injected, and the learning rate is linearly increased to 5e4 while data augmentation is performed. In the fault injection training, five typical damage data are generated through finite element simulation, the learning rate is cosine annealing decayed to 1e5, and damage pattern recognition is strengthened.

[0012] Further, the training of the model based on the extracted features also includes searching for hyperparameters using Bayesian optimization. The search space includes learning rate, number of Transformer layers, mechanism loss weight, and focal loss parameter. The evaluation index is confirmed by using a weighted comprehensive score, and the optimal parameter combination is obtained through 5-fold cross-validation on historical data of the bid section according to the evaluation index.

[0013] In a second aspect, a potential damage prediction system for a large-aperture corrugated pipe culvert based on multi-source heterogeneous data includes: The data acquisition module is configured to cooperatively collect multi-source heterogeneous data. a preprocessing module configured to preprocess the acquired multi-source heterogeneous data; a model construction module configured to construct a multi-source heterogeneous large-diameter corrugated pipe culvert potential damage prediction model; a training module configured to perform feature extraction on data by using the multi-source heterogeneous large-diameter corrugated pipe culvert potential damage prediction model, and train the model based on the extracted features; a prediction module configured to perform prediction by using the trained model.

[0014] In a third aspect, the present application provides a computer readable storage medium, wherein a plurality of instructions are stored, the instructions being adapted to be loaded and executed by a processor of a terminal device to implement the multi-source heterogeneous large-diameter corrugated pipe culvert potential damage prediction method.

[0015] In a fourth aspect, the present application provides a terminal device, comprising a processor and a computer readable storage medium, the processor being configured to implement instructions; and the computer readable storage medium being configured to store a plurality of instructions, the instructions being adapted to be loaded and executed by the processor to implement the multi-source heterogeneous large-diameter corrugated pipe culvert potential damage prediction method.

[0016] In summary, the present application has the following beneficial technical effects: The technical effects of the present application are closely related to the research objectives and technical indicators of the declaration, and are as follows: Through the integration of fiber grating technology and Internet of Things technology, real-time acquisition and intelligent analysis of construction period structure strain, displacement, soil pressure and other parameters are realized, and the damage identification rate reaches more than 98%. For example, in the 6m diameter corrugated pipe culvert of JLSG-2 section of Jinan-Linqing Expressway, the fiber grating strain sensor can capture a small strain change of 0.5με, combined with the point cloud data (accuracy of 1mm) obtained by three-dimensional laser scanning, the damage position such as joint offset and pipe wall buckling can be accurately positioned and spliced, the detection efficiency is improved by 300% compared with traditional manual detection, and the error is reduced by 85%. The digital pre-assembly process based on three-dimensional laser scanning technology controls the axis offset error to be less than or equal to 5mm, and the assembly accuracy is improved by 180% compared with the traditional method. At the same time, the improved Rayleigh wave method is used for backfill soil compaction effect detection, and through the inversion interpretation optimization algorithm, the backfill compactness and compaction degree are greater than or equal to 98%, solving the problem of insufficient compaction of large-diameter corrugated pipe culverts due to the inability to use large-scale rolling equipment. For example, in the K30+120 test section, this technology improves the one-time acceptance qualification rate of backfill construction quality from 75% to 99%, avoiding the cost of rework. The developed damage perception type large aperture corrugated pipe culvert structure has a size increase of 85% and a span capacity increase of 260% compared with traditional structures, and can meet the engineering requirements of apertures of 6 m and above. Meanwhile, by constructing a real-time monitoring cloud platform and a potential damage prediction model, intelligent evaluation of the structure state in the construction stage is realized, and technical support is provided for the industrialization and intelligent development of highway engineering pipe culverts in China. As described in the declaration, the achievement can enrich the construction theory of road and bridge culverts, fill the blank of related technical standards, and is expected to save 60 days of construction period and reduce the engineering cost by 24.5 million yuan, with significant economic and social benefits. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a schematic diagram of a potential damage prediction method for a large-aperture corrugated pipe culvert based on multi-source heterogeneity according to Embodiment 1 of the present application. DETAILED DESCRIPTION

[0018] The present application will be further described in detail below with reference to the accompanying drawings.

[0019] Embodiment 1 Referring to Figure 1 , a potential damage prediction method for a large-aperture corrugated pipe culvert based on multi-source heterogeneity according to the present embodiment includes: Potential damage prediction technical solution for large-aperture corrugated pipe culvert based on multi-source heterogeneity Specifically, the following steps are included: I. Multi-source heterogeneous data collaborative collection 1. Sensor array layout and data collection In the corrugated pipe culvert project of JLSG2 section of Jinan-Linqing Expressway, for structures with apertures of 3 m, 4 m and 6 m, optical fiber grating strain sensors are arranged in a grid pattern in the axial and circumferential directions, with a spacing adjusted according to the pipe diameter: 6 m pipe diameter is arranged at a spacing of 0.8 m, 4 m pipe diameter is arranged at a spacing of 0.6 m, 3 m pipe diameter is arranged at a spacing of 0.5 m, and the key splicing joints are encrypted to 0.3 m. The sensor is based on the principle of fiber Bragg grating. When the structure is strained, the grating period changes, causing the reflected light wavelength to shift. The strain value is calculated by the formula , wherein is the Bragg wavelength (central wavelength 1550 nm), Pe is the effective photoelastic coefficient (0.22), and the wavelength demodulator has an accuracy of 1 pm, corresponding to a strain resolution of 0.5 με.

[0020] The tilt sensor adopts MEMS silicon micro-mechanical structure, and is symmetrically arranged at the top, 1 / 4 height and 3 / 4 height of the corrugated pipe culvert, with 46 holes arranged in each hole, a measurement range of ± 15°, an accuracy of 0.01°, and a relationship between output voltage and tilt angle of V = 0.1θ + 2.5 (V). The LoRa wireless pressure sensor is buried in three layers in the earth: 0.5 m from the top of the pipe (1 / 3 of the depth of the earth), 1.0 m (2 / 3 of the depth of the earth) and the pipe side backfill layer, with 3 arranged in each layer. The pressure resistance type sensing element is used, and the relationship between the pressure resistance is (MPa), the communication distance reaches 1 km, and the sampling frequency is 1 Hz.

[0021] The three types of sensors are synchronized through the IEEE1588 clock protocol, and the time delay compensation mechanism of optical fiber transmission is used to control the timestamp error within 500 ns, ensuring the time sequence consistency of multi-source data during the construction period (such as assembly, backfilling, etc.), and providing accurate space-time coordinates for subsequent damage analysis.

[0022] 2. Three-dimensional spatial information and visual detection The total station uses Leica TS60 model, with an angle measurement accuracy of 0.5" and a distance measurement accuracy of 1 mm + 1.5 ppm. A global coordinate system containing 6 control points is established in the corrugated pipe culvert construction area. The control points are buried in stable bedrock to avoid the influence of construction disturbance. The three-dimensional laser scanner selects FARO Focus S150, with a scanning speed of 1.22 million points / second and a distance accuracy of 1 mm. It scans in a spiral path: spirally ascending from the bottom to the top of the pipe, with a layer spacing of 8 mm and an angular resolution of 0.067°. About 50 million point cloud data are obtained by single-hole scanning, which are spliced into a complete three-dimensional model through point cloud registration technology (iterative closest point algorithm ICP).

[0023] The binocular visual detection system uses two Basler acA2500 14gm cameras with a baseline distance of 40 cm, a resolution of 2592x1944, and a frame rate of 14 fps. It is installed 1.5 m in front of the assembly trolley. The three-dimensional coordinates of the assembly joint are calculated through the parallax formula (f is the focal length, B is the baseline distance, and D is the object distance). The joint width detection accuracy reaches 0.3 mm, and the assembly deviation of ≥1 mm can be identified.

[0024] 3. Geological and structural mechanical property detection The adjustable frequency source uses L22 type land low frequency source, with a frequency range of 101000 Hz. The source energy increases with the decrease of frequency, which is suitable for detecting soil bodies at different depths: a 10 Hz source can detect a depth of 510 m, and a 1000 Hz source focuses on a 0.52 m surface layer. The ground penetrating radar selects SIR3000 type, equipped with 200 MHz and 500 MHz antennas. The 200 MHz antenna has a detection depth of 35 m and a resolution of 0.3 m; the 500 MHz antenna has a detection depth of 12 m and a resolution of 0.1 m.

[0025] During seismic wave detection, 8 geophones were symmetrically arranged on both sides of the pipe culvert with a spacing of 2 m. The empirical relationship between Rayleigh wave velocity vR and soil compaction degree \gamma (unit: g / cm³) was analyzed to evaluate the backfill quality. Ground penetrating radar identified loose soil areas through dielectric constant differences: loose soil dielectric constant ε ≈ 46, dense soil ε ≈ 812, and structural damage ε ≈ 23. The abnormal area was located by changes in reflected wave amplitude and phase.

[0026] II. Preprocessing of multi-source heterogeneous data 1. Fiber sensing data filtering and calibration The original strain data was decomposed into approximate components and detail components using db5 wavelet basis for 5-layer decomposition. The soft threshold function was used to denoise the detail components. Threshold value (σ is the noise standard deviation, n is the signal length). The strain sensor was calibrated every 15 days using a YBT501 standard pressure source (accuracy 0.05% FS). Five levels of standard strain were applied: 0με, 500με, 1000με, 1500με, and 2000με. The corresponding wavelength drift was recorded, and the calibration curve was fitted by the least squares method. (unit: pm), the linear error of the calibrated sensor was ≤0.3% FS.

[0027] The inclination sensor calibration used a high-precision inclination reference (resolution 0.001°). Seven calibration points were taken in the range of 10° to +10°. The calibration equation V=0.102θ +2.48 was obtained by the least squares method. The corrected zero offset error was ≤0.02°, and the sensitivity error was ≤0.5%.

[0028] 2. Identification of outliers and data repair The improved 3σ principle was used for strain and soil pressure data: first, the trend item was removed by local weighted regression (LOESS) method, then the mean μ and standard deviation σ of the remaining data were calculated, when |xi\mu|>3.5\sigma was determined as an outlier. The piecewise cubic Hermite interpolation (PCHIP) was used for repair, which ensured the monotonicity of the interpolation curve at the nodes and avoided the overshoot phenomenon of traditional spline interpolation. For example, for a certain strain data x(t), the interpolation formula at the abnormal point t0 is: where yi is the node value, mi is the node derivative, hi(t) and are the basis functions, which ensure that the repaired data is consistent with the actual physical law.

[0029] 3. Cross-modal data alignment and standardization ​For three-dimensional point cloud coordinates (x, y, z) and sensor data (strain, pressure, etc.), ZScore standardization is performed first: where μ and σ are the mean and standard deviation of the training set data, respectively. For data with different sampling frequencies (e.g., strain 10 Hz, soil pressure 1 Hz), a synchronous synchrosqueezing transform (SST) is used for time alignment: the low-frequency soil pressure data is mapped to the high-frequency strain time axis, and through the time-frequency resolution adaptive characteristics of SST, the sub-sampling point alignment is realized while maintaining the signal time-frequency characteristics, with a time error ≤5 ms.

[0030] For the spatial alignment of point cloud data and sensor data, a registration method based on feature points is used: 1015 feature points (such as joint end points, sensor layout positions) are selected on the corrugated pipe culvert surface, and their three-dimensional coordinates are measured by a total station as the spatial reference for point cloud data and sensor data. The transformation matrix [R|t] is used to convert the point cloud coordinates to the sensor data coordinate system, with a conversion error ≤2 mm.

[0031] III. Construction of multi-source heterogeneous damage prediction model 1. Feature extraction and cross-modal fusion Time series features: kurtosis and skewness are extracted from strain data to reflect load impact characteristics; marginal spectral energy distribution is obtained through Hilbert-Huang transform (HHT) to identify structural natural frequency changes. Gradient features are extracted from soil pressure data to characterize soil stress diffusion patterns.

[0032] Spatial features: point cloud data is processed by moving least squares (MLS) to calculate local curvature , where n(p) is the normal vector of point p and d(p, q) is the distance between points), and curvature changes indicate structural deformation; RANSAC algorithm is used to fit the straight line equation of the joint, and the offset \Deltal between the actual joint line and the design joint line is calculated. Visual image features are extracted by Canny edge detection, and joint angle is identified by Hough transform with an accuracy of 0.5°.

[0033] Cross-modal fusion: construct steel-soil coupling feature matrix, such as strain-soil pressure phase difference , reflecting the time delay of structural and soil dynamic response; define joint error strain correlation degree (ei is the joint error), quantifying the influence of geometric defects on mechanical response. Calculate the weight of each modal feature through self-attention mechanism: where Fj is the jth modal feature, m=5 (strain, inclination, soil pressure, point cloud, vision), MLP is a 2-layer fully connected network, and the output is the fused feature .

[0034] 2. Model Architecture and Damage Mechanism Embedding Design an architecture of "Spatiotemporal Transformer + Damage Mechanism Module": The spatiotemporal Transformer encoder consists of 8 layers, each containing a multi-head self-attention network (12 heads) and a feedforward neural network. The self-attention calculation formula is as follows: PosEmb is a spatiotemporal location encoding that combines timestamps and spatial coordinates (culvert axial mileage, circumferential angle) to enable the model to perceive the spatiotemporal correlation of the data, and dk=128 is the key dimension.

[0035] Damage mechanism module embedded in buckling mechanics model of corrugated steel pipe culvert: deriving critical buckling stress through elastoplastic theory. (E is the elastic modulus, nu is Poisson's ratio, t is the wall thickness, and D is the pipe diameter), integrating theoretical buckling characteristics with data-driven characteristics to form a dual damage criterion of "theoretical threshold + data mapping".

[0036] The multi-task decoder includes: Strain prediction decoder: 3 fully connected layers, with Swish activation function and Huber loss function. Where δ=50με is the optimization weight that balances large and small errors.

[0037] Damage Classification Decoder: Employs a hierarchical classification structure, first distinguishing between "normal / abnormal," then further subdividing into "material damage / assembly damage / soil damage," with the loss function being the focal loss. For minor damage (accounting for about 70%) that is common during the construction period, α=0.3 and γ=2 are set for severe damage.

[0038] IV. Model Training Optimization 1. Multi-task joint training strategy Loss function system: Basic task loss: L0 = 0.4Lhub + 0.6Lfocal; Mechanistic constraint loss: Introducing buckling theory loss , ( To predict stress, ensure that the prediction results do not violate the principles of mechanics; Cross-modal consistency loss: (F{mech} represents the mechanistic characteristics, and F{data} represents the data characteristics), strengthening the consistency between physical mechanisms and data characteristics; Total loss: L = L0 + 0.2Lcr + 0.1Lcon.

[0039] Optimization algorithm: Ranger optimizer (RAdam + LookAhead + GC) RAdam dynamically adjusts the learning rate: where d = 60 is the preheating step number, η = 1e-3; LookAhead updates the weights slowly: , α = 0.5, k = 5; Gradient centralization (GC): , to suppress gradient explosion.

[0040] 2. Training in stages and data augmentation Stage 1 (0-40 rounds): Pure data-driven training, using normal working condition data during construction (60%), learning rate 1e4, to make the model learn basic feature representation; Stage 2 (41-90 rounds): Mechanism data fusion training, injecting buckling theory, steel-soil interaction, etc. Mechanism constraints, learning rate linearly increased to 5e4, while data augmentation: Strain data: Add ±10% Gaussian noise to simulate sensor error; Point cloud data: Randomly delete 510% points to test noise resistance; Soil pressure data: Introduce ±15% scale transformation to simulate soil parameter fluctuations; Stage 3 (91-150 rounds): Fault injection training, generate 5 typical damage data (joint disconnection, local soil loosening, pipe wall buckling, insufficient backfill compaction, temperature stress concentration) through finite element simulation, accounting for 30% of the training set, learning rate using cosine annealing decay to 1e5, \lambda{cr} from 0.2 to 0.3, to strengthen damage pattern recognition.

[0041] 3. Hyperparameter optimization and model validation Use Bayesian optimization to search for hyperparameters, search space includes: Learning rate: (log scale); Number of Transformer layers: ; Mechanism loss weight: ; Focal loss parameter: .

[0042] Evaluation index uses a weighted comprehensive score: where Corr is the correlation coefficient of predicted damage and actual damage. On the historical data of JLSG2 section, 5-fold cross-validation is performed, and the optimal parameter combination is: L = 8, η = 3e-4, λcr= 0.25, α = 0.5, γ = 2.5. At this time, the F1 score of the model on the validation set is 0.96, RMSE = 8.7με, and Corr = 0.92.

[0043] V. Model prediction and engineering application verification 1. Real-time prediction and multi-level warning mechanism Prediction process: real-time data is input into the model after preprocessing, and the construction condition knowledge base (such as backfill thickness, compaction machinery type) is called at the same time. The prediction accuracy is improved through conditional normalization processing (adjusting feature weight according to current working condition). The soil pressure feature weight is increased to 0.4 in the backfill stage, and the assembly error feature weight is increased to 0.5 in the assembly stage.

[0044] Warning classification: Yellow warning: damage probability ≥ 0.6 or strain ≥ 120με, prompting to strengthen monitoring; Orange warning: damage probability ≥ 0.8 or strain ≥ 150με accompanied by point cloud curvature anomaly, suggesting to suspend construction for inspection; Red warning: damage probability ≥ 0.95 or strain ≥ 200με and soil pressure-strain correlation r < 0.3, triggering emergency shutdown mechanism.

[0045] 2. Uncertainty quantification and engineering verification Based on MCDropout (100 times forward propagation) to calculate the prediction variance, combined with the risk matrix in the construction stage to determine the disposal strategy.

[0046] Embodiment 2 The embodiment provides a potential damage prediction system for a large-aperture corrugated pipe culvert based on multi-source heterogeneity, comprising: The data acquisition module is configured to A computer readable storage medium, wherein a plurality of instructions are stored, the instructions are suitable for being loaded and executed by a processor of a terminal device, and the instructions are suitable for being loaded and executed by a processor of a terminal device.

[0047] A terminal device, comprising a processor and a computer readable storage medium, the processor is used to implement instructions; the computer readable storage medium is used to store a plurality of instructions, the instructions are suitable for being loaded and executed by the processor, and the instructions are suitable for being loaded and executed by the processor.

[0048] The above are preferred embodiments of the present application, and do not limit the protection scope of the present application, so: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A potential damage prediction method based on multi-source heterogeneous large-diameter corrugated pipe culvert, characterized by, The method comprises the following steps: Cooperative collection of multi-source heterogeneous data; Preprocessing of the collected multi-source heterogeneous data; Construction of a potential damage prediction model for large-diameter corrugated pipe culverts based on multi-source heterogeneous data; Feature extraction from the data using the potential damage prediction model for large-diameter corrugated pipe culverts based on multi-source heterogeneous data, and training of the model based on the extracted features; Prediction using the trained model.

2. The method according to claim 1, wherein, The cooperative collection of multi-source heterogeneous data comprises arranging fiber Bragg grating strain sensors and inclination sensors on the axial, circumferential and key joint of the corrugated pipe culvert, and burying LoRa wireless pressure sensors in the soil to collect strain, displacement and soil pressure data; point cloud data is obtained through a spiral scanning path using a total station + three-dimensional laser scanning joint positioning, and the joint is detected based on binocular vision; detection is performed using an adjustable frequency seismic source and a ground penetrating radar, and IEEE1588 clock protocol is used to realize multi-source data synchronization.

3. The method according to claim 2, wherein, The preprocessing of the collected multi-source heterogeneous data comprises filtering and denoising of the fiber sensing data, data calibration and outlier processing, wherein the fiber Bragg grating strain sensors, inclination sensors and the like are calibrated periodically using a standard pressure source and an inclination reference, and the sensor measurement error is corrected by fitting a calibration curve using the least squares method; and the collected strain and soil pressure data are subjected to outlier identification to eliminate abnormal data points caused by construction interference or sensor failure.

4. The method according to claim 3, wherein, The construction of the potential damage prediction model for large-diameter corrugated pipe culverts based on multi-source heterogeneous data comprises using a Transformer architecture as the base model, combining attention mechanism and multi-modal fusion mechanism to realize prediction of the potential damage of large-diameter corrugated pipe culverts, wherein the Transformer captures long-range dependencies in the sequence by calculating the correlation between each element and other elements in the input sequence, first maps the input feature vector to a query, key and value vector space, then calculates the dot product similarity of the query vector and the key vector, and normalizes it through the Softmax function to obtain the attention weight.

5. The method according to claim 4, wherein, The construction of the potential damage prediction model for large-diameter corrugated pipe culverts based on multi-source heterogeneous data also comprises introducing a multi-modal fusion mechanism in the Transformer architecture to capture the interaction between different modal data, which comprises a cross-attention layer and a feature fusion layer, the cross-attention layer allows information exchange between features of different modalities, and the difference between the calculation process and the self-attention mechanism is that the query vector comes from one modality, and the key vector and the value vector come from another modality; The feature fusion layer adopts a multi-layer perceptron MLP and residual connection mode to fuse the features of different modalities to obtain a comprehensive feature representation.

6. The method according to claim 5, wherein, The potential damage prediction model based on multi-source heterogeneous large-diameter corrugated pipe culvert is used for feature extraction of data, including using a feature-level fusion method to extract features from preprocessed multi-source heterogeneous data, then fusing the extracted features to construct a comprehensive feature vector, wherein for strain data, the extracted features include mean, standard deviation, maximum value, minimum value, peak frequency and energy entropy; for inclination data, the extracted features include inclination change rate, inclination fluctuation range and inclination trend item; for soil pressure data, the extracted features include soil pressure mean, soil pressure gradient and soil pressure change rate; for point cloud data, the extracted features include point cloud density, point cloud curvature, splicing joint width and splicing joint offset; for seismic wave and ground penetrating radar data, the extracted features include wave speed, amplitude, frequency and reflection coefficient.

7. The method according to claim 6, wherein, The model is trained based on the extracted features, including training the model in stages based on a multi-task joint training strategy by constructing a loss function system and a Ranger optimization algorithm, wherein pure data-driven training uses construction period normal operating condition data and sets a first learning rate 1e4 to make the model learn basic feature representation; mechanism data fusion training injects buckling theory and steel-soil interaction mechanism constraints, and the learning rate is linearly increased to 5e4 while data augmentation is performed; fault injection training generates five typical damage data through finite element simulation, the learning rate adopts cosine annealing decay to 1e5, and damage mode recognition is strengthened.

8. The method according to claim 7, wherein, The model is trained based on the extracted features, and further includes searching for hyperparameters by using Bayesian optimization, the search space includes learning rate, number of Transformer layers, mechanism loss weight and focal loss parameter, a weighted comprehensive score is used to confirm the evaluation index, 5-fold cross-validation is performed on historical data of the bid section according to the evaluation index, and the optimal parameter combination is obtained.

9. A potential damage prediction system based on multi-source heterogeneous large aperture corrugated pipe culvert, characterized by, It comprises: a data acquisition module configured to cooperatively acquire multi-source heterogeneous data; a preprocessing module configured to preprocess the acquired multi-source heterogeneous data; a model construction module configured to construct a potential damage prediction model based on multi-source heterogeneous large-diameter corrugated pipe culvert; a training module configured to extract features from data using the potential damage prediction model based on multi-source heterogeneous large-diameter corrugated pipe culvert, and train the model based on the extracted features; a prediction module configured to use the trained model for prediction.

10. A computer readable storage medium having stored therein a plurality of instructions, wherein the instructions, when executed by a processor, cause the processor to perform the method of any one of claims 1-9. The instructions are suitable for being loaded and executed by a processor of a terminal device to implement the potential damage prediction method based on multi-source heterogeneous large-diameter corrugated pipe culvert.

Images