A method and system for identifying structural damage of high-pile wharf

Abstract

This invention proposes a method and system for identifying structural damage in high-pile wharves, relating to the technical field of wharf damage identification. This invention overcomes the blind spots of single sensor monitoring by collaboratively collecting multi-source data (strain, deformation, and defects). This multi-source data is integrated into multi-source heterogeneous data aligned in time and space, providing high-dimensional input for deep learning models. The damage feature vectors output by the model are matched with a dynamic threshold library, achieving a closed-loop process from "data collection-feature fusion-damage decision-making." Specifically, multi-dimensional data collection improves detection coverage, standardization and fusion processing enhance anti-interference capabilities, deep learning models improve damage analysis accuracy, and dynamic threshold libraries enable adaptive decision-making.

Description

Technical Field

[0001] The present invention relates to the technical field of dock damage identification, and in particular to a method and system for identifying damage to a high-pile dock structure. Background Art

[0002] As the core facility of port projects, high-pile docks are subjected to complex effects such as ship loads, water flow impact, and environmental corrosion for a long time, which can easily cause structural damage. Existing damage detection technologies have the following problems:

[0003] First, the detection dimension is limited: Traditional methods rely on a single sensor (such as a strain gauge or vibration sensor) to collect local data, making it difficult to fully reflect the internal and external damage status of the structure. For example, vibrating wire strain gauges can only monitor surface strain and cannot detect internal cracks or hidden defects in underwater pile foundations. Second, real-time performance and interference immunity are insufficient: Existing technologies rely on manual inspections or offline analysis, which cannot achieve real-time monitoring. Moreover, changes in ambient temperature and humidity can easily lead to misjudgments. For example, transient damage caused by overloading and heap loading may be missed due to detection delays. Third, multi-source data fusion capabilities are weak: Deformation, strain, and internal defect data are analyzed in isolation, lacking temporal and spatial correlation, leading to ambiguous damage location. For example, surface deformation data and internal defect data are not analyzed in a coordinated manner, making it difficult to distinguish damage causes (such as ship impact or material degradation). Fourth, the threshold library is static: Traditional threshold libraries are statically set based on historical data and cannot adapt to long-term terminal performance degradation and dynamic environmental changes, resulting in increased false alarm rates. Summary of the Invention

[0004] In view of the above problems existing in the prior art, the first aspect of the present invention provides a method for identifying damage to a high-pile wharf structure, comprising:

[0005] Step S1: collecting structural strain data and temperature data of the high-pile wharf in real time based on a distributed optical fiber sensing network to generate a first data set;

[0006] Step S2: acquiring surface deformation data of the high-pile wharf by a three-dimensional laser scanner to generate a second data set;

[0007] Step S3: using an ultrasonic tomography scanner to perform non-destructive testing on the internal structure of the high-pile wharf, obtaining ultrasonic tomography raw signal data, and generating a third data set;

[0008] Step S4: performing feature extraction and standardization on the first dataset, the second dataset, and the third dataset, including dimensionality reduction through principal component analysis, eliminating dimensional differences using a normalization algorithm, and aligning feature vectors based on geographic information system coordinates to generate multi-source heterogeneous data;

[0009] Step S5: Input the multi-source heterogeneous data into the pre-trained deep learning model, extract spatial features through the convolutional neural network, and analyze the temporal association through the long short-term memory network to generate a fused damage feature vector;

[0010] Step S6: Based on the damage feature vector, similarity matching is performed in combination with the dynamically updated damage threshold library to output the damage location and damage level.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, step S1 includes:

[0012] Step S1-1: Using distributed optical fiber sensors symmetrically arranged along the axis of the pile foundation, the strain distribution data on the pile foundation surface and the ambient temperature change data are collected to generate an original strain data set;

[0013] Step S1-2: inputting the original strain data set into the Kalman filter module to remove noise interference data to generate a first data set;

[0014] Step S2 includes:

[0015] Step S2-1: Scan the pile cap, pile body, and panel of the high-pile wharf at multiple angles using a 3D laser scanner to obtain surface deformation point cloud data;

[0016] Step S2-2: Input the surface deformation point cloud data into the coordinate conversion module of the geographic information system, align it to the unified spatial reference, and generate a second data set.

[0017] In conjunction with the first aspect, in some implementations of the first aspect, step S4 includes:

[0018] Step S4-1: Input the structural strain data in the first data set into the Hilbert-Huang transform module, decompose it into multiple intrinsic mode functions, extract the main frequency components and transient energy characteristics, and generate frequency domain feature vectors;

[0019] Step S4-2: Input the surface deformation data in the second data set into the iterative closest point algorithm module, remove noise points and perform registration based on geographic information system coordinates, extract the pile cap displacement and pile body inclination angle, and generate a deformation gradient vector;

[0020] Step S4-3: Inputting the ultrasonic tomography raw signal data in the third data set into the back-projection reconstruction algorithm module to analyze the geometric parameters and position information of the internal defects and generate a tomographic imaging map;

[0021] Step S4-4: Input the frequency domain feature vector, deformation gradient vector and tomographic imaging map into the normalization module, eliminate the dimensionality difference, and perform dimensionality reduction through principal component analysis to generate standardized multi-source heterogeneous data.

[0022] In conjunction with the first aspect, in some implementations of the first aspect, the dimensionality reduction by principal component analysis in step S4-4 includes:

[0023] Step S4-4-1: Input the frequency domain eigenvector, deformation gradient vector and tomographic imaging map into the covariance matrix calculation module to generate the covariance matrix of the multi-source data;

[0024] Step S4-4-2: Perform eigenvalue decomposition on the covariance matrix, select the top three principal components with the highest contribution rate, and generate the eigenvector after dimensionality reduction.

[0025] In conjunction with the first aspect, in some implementations of the first aspect, in step S4-3, analyzing the geometric parameters and location information of the internal defect includes:

[0026] According to the relationship between the acoustic wave propagation time and amplitude attenuation, the three-dimensional spatial coordinates and equivalent aperture of the internal defects are reconstructed.

[0027] In conjunction with the first aspect, in some implementations of the first aspect, step S5 includes:

[0028] Step S5-1: Input multi-source heterogeneous data into the convolutional neural network, extract spatial local features through multi-layer convolution kernels, and generate a damage-sensitive feature map;

[0029] Step S5-2: Input the damage-sensitive feature map into the long short-term memory network, analyze the temporal variation pattern, and generate a damage feature vector that integrates spatiotemporal correlation.

[0030] In conjunction with the first aspect, in some implementations of the first aspect, in step S6, generating the dynamically updated damage threshold library includes:

[0031] Step S6-1: training the initial damage threshold library using sample data in the historical damage database to generate benchmark matching parameters;

[0032] Step S6-2: Input the first data set, the second data set, and the third data set into the adaptive update module, dynamically adjust the benchmark matching parameters based on the real-time collected structural strain data, surface deformation data, and ultrasonic tomography raw signal data, and generate a dynamically updated damage threshold library.

[0033] In conjunction with the first aspect, in some implementations of the first aspect, the deploying of the distributed optical fiber sensor network in step S1 includes:

[0034] Two measuring lines are marked symmetrically along the axial direction on the pile foundation surface, one of which is located on the water side and the other on the slope side, to generate measuring line positioning information;

[0035] The distributed optical fiber is laid along the survey line to locate the information and fixed to the pile surface in three sections using underwater epoxy resin, located at the top, middle and bottom of the pile respectively, forming an optical fiber sensing network for continuous monitoring of the entire pile length.

[0036] In conjunction with the first aspect, in some implementations of the first aspect, the training process of the deep learning model includes:

[0037] Finite element simulation is used to simulate the structural strain data, surface deformation data and ultrasonic fault raw signal data of high-pile docks under different damage conditions to generate virtual training samples;

[0038] Integrate virtual training samples with historical monitoring data and field loading test data to construct a training sample set covering multiple scenarios;

[0039] Input the training sample set into the adversarial generative network to generate an enhanced training set;

[0040] The convolutional neural network and long short-term memory network are trained by enhancing the training set to generate a deep learning model.

[0041] In a second aspect, the present invention provides a high-pile wharf structure damage identification system, which adopts the method provided in any of the above embodiments and includes:

[0042] The data acquisition module includes a distributed optical fiber sensor network, a three-dimensional laser scanner, and an ultrasonic tomography scanner, which is used to collect structural strain data, temperature data, surface deformation data, and ultrasonic tomography raw signal data of the high-pile wharf in real time to generate a first data set, a second data set, and a third data set respectively;

[0043] a data processing module connected to the data acquisition module, configured to perform feature extraction and standardization processing on the first data set, the second data set, and the third data set, including dimensionality reduction through principal component analysis, elimination of dimensional differences using a normalization algorithm, and alignment of feature vectors based on geographic information system coordinates to generate multi-source heterogeneous data;

[0044] The model processing module is connected to the data processing module and includes a pre-trained deep learning model composed of a convolutional neural network and a long short-term memory network. The deep learning model is used to extract spatial features and perform temporal correlation analysis on multi-source heterogeneous data to generate a fused damage feature vector.

[0045] The damage identification module is connected to the model processing module and includes a dynamically updated damage threshold library and a matching calculation unit, which is used to perform similarity matching based on the damage feature vector and output the damage location and damage level.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The detection coverage is improved through multi-dimensional data collection from step S1 to step S3. Specifically, in step S1, the distributed fiber optic sensing network collects structural strain data and temperature data in real time, and covers the above-water and underwater areas through continuous monitoring of the entire pile length, solving the blind spot problem of traditional local monitoring; in step S2, the three-dimensional laser scanner obtains surface deformation data, accurately captures macroscopic deformation features such as pile cap displacement and pile body inclination, and makes up for the shortcomings of single strain data; in step S3, the ultrasonic tomography scanner analyzes internal defect data through acoustic wave signals to achieve visual detection of hidden damage (such as cracks and voids), breaking through the limitations of surface monitoring.

[0048] In addition, in step S4, principal component analysis reduces the dimensionality of multi-source data and eliminates redundant features; the normalization algorithm unifies the dimensions of strain, deformation, and defect data to reduce the magnitude deviation caused by ambient temperature fluctuations; and the geographic information system coordinates align the feature vectors to ensure the consistency of spatiotemporal data and avoid misjudgment.

[0049] In step S5, the convolutional neural network extracts the spatial local features of the strain and deformation data (such as crack propagation patterns), and the long-short-term memory network analyzes the temporal correlation patterns (such as the cumulative effect of overload and stacking). The two are fused to generate a damage feature vector, which solves the problem that the traditional threshold method has poor adaptability to complex damage patterns and improves the accuracy of damage analysis.

[0050] In step S6, the dynamically updated damage threshold library adaptively adjusts matching parameters based on real-time monitoring data, overcoming the benchmark failure problem of the static threshold library due to material aging and environmental corrosion, and improving the reliability of long-term monitoring.

[0051] The synergy between steps S1 through S6 is demonstrated by the integration of multi-source data (strain, deformation, and defects) into spatiotemporally aligned, heterogeneous data in step S4, providing high-dimensional input for the deep learning model in step S5. The damage feature vectors output by the model are matched with a dynamic threshold library, completing a closed-loop process from "data acquisition-feature fusion-damage decision-making." For example, fiber optic data captures instantaneous overload strain, laser scanning data reflects the cumulative effects of deformation, and ultrasonic data locates internal defects. These three collaboratively determine damage severity (e.g., minor cracks or severe fractures), reducing the risk of misjudgment from a single data source. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 FIG2 is a flow chart of a method for identifying structural damage of a high-pile wharf provided by an embodiment of the present invention.

[0054] Figure 2a FIG2 is a schematic diagram of a distributed optical fiber sensing network provided by an embodiment of the present invention.

[0055] Figure 2b FIG2 is a schematic diagram of a distributed optical fiber sensing network provided by another embodiment of the present invention.

[0056] Figure 3 FIG. 1 is a schematic diagram of a scanned image provided by an embodiment of the present invention.

[0057] Figure 4 FIG2 is a schematic structural diagram of a high-pile wharf structure damage identification system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0059] The specific embodiments of the present invention are described below.

[0060] Example 1

[0061] like Figure 1 As shown, the present invention proposes a method for identifying damage to a high-pile wharf structure, comprising:

[0062] Step S1: Based on a distributed optical fiber sensor network (such as Figure 2a and Figure 2b (as shown) collecting structural strain data and temperature data of the high-pile wharf in real time to generate a first data set;

[0063] Step S2: acquiring surface deformation data of the high-pile wharf by a three-dimensional laser scanner to generate a second data set;

[0064] in, Figure 3 The figure shows a scanning image of a high-pile pier obtained by a 3D laser scanner, through which surface deformation data of the high-pile pier is obtained;

[0065] Step S3: using an ultrasonic tomography scanner to perform nondestructive testing on the internal structure of the high-pile wharf, obtaining ultrasonic tomography raw signal data, and generating a third data set;

[0066] Step S4: performing feature extraction and standardization on the first dataset, the second dataset, and the third dataset, including dimensionality reduction through principal component analysis, eliminating dimensional differences using a normalization algorithm, and aligning feature vectors based on geographic information system coordinates to generate multi-source heterogeneous data;

[0067] Step S5: Input the multi-source heterogeneous data into the pre-trained deep learning model, extract spatial features through the convolutional neural network, and analyze the temporal association through the long short-term memory network to generate a fused damage feature vector;

[0068] Step S6: Based on the damage feature vector, similarity matching is performed in combination with the dynamically updated damage threshold library to output the damage location and damage level.

[0069] This structural damage identification method for high-pile wharfs achieves precise damage location through multi-source data collection, fusion processing, and intelligent analysis. A distributed fiber-optic sensor network is deployed along the surface of the pile foundation, collecting real-time structural strain and temperature data to form a first dataset covering the entire length of the pile. A 3D laser scanner scans the pile cap, pile body, and face plate from multiple angles, generating surface deformation point cloud data. This data is then aligned with the geographic information system (GIS) coordinates to form a second dataset. An ultrasonic tomography scanner detects internal defects using acoustic reflection signals, generating a third dataset. In the feature extraction and normalization phase, principal component analysis reduces the dimensionality of the multi-source data and eliminates redundant features. A normalization algorithm unifies the dimensional differences in strain, deformation, and defect data. GIS coordinate alignment ensures spatiotemporal consistency, generating multi-source heterogeneous data. In the deep learning model, a convolutional neural network extracts local spatial features (such as crack propagation patterns), while a long-short-term memory network analyzes temporal correlations (such as the cumulative effects of overloads) to generate a damage feature vector that incorporates spatiotemporal characteristics. A dynamically updated damage threshold library adjusts matching parameters based on real-time monitoring data, combines the damage feature vectors for similarity matching, and outputs damage locations and damage levels.

[0070] The present invention overcomes the blind spots of single sensor monitoring by collaboratively collecting multi-source data (strain, deformation, and defects). Standardized processing eliminates environmental interference, a deep learning model enhances the ability to analyze complex damage patterns, and a dynamic threshold library adapts to long-term terminal performance degradation. For example, distributed optical fiber captures instantaneous overload strain, laser scanning data reflects cumulative deformation trends, and ultrasonic tomography data locates internal cracks. The fusion of these three data sources can distinguish between ship impact (instantaneous high-frequency strain) and material degradation (long-term deformation accumulation), avoiding the misjudgment caused by traditional data isolation methods.

[0071] Among the alternative solutions, the distributed fiber optic sensing network can adopt fiber optic Bragg grating sensors or distributed acoustic wave sensing technology; the normalization algorithm can be replaced by Z-score standardization or decimal scaling standardization; and the deep learning model can choose graph convolutional network instead of convolutional neural network to better process non-Euclidean spatial data.

[0072] In conjunction with the first aspect, in some implementations of the first aspect, step S1 includes:

[0073] Step S1-1: Using distributed optical fiber sensors symmetrically arranged along the axis of the pile foundation, the strain distribution data on the pile foundation surface and the ambient temperature change data are collected to generate an original strain data set;

[0074] Step S1-2: inputting the original strain data set into the Kalman filter module to remove noise interference data to generate a first data set;

[0075] Step S2 includes:

[0076] Step S2-1: Scan the pile cap, pile body, and panel of the high-pile wharf at multiple angles using a 3D laser scanner to obtain surface deformation point cloud data;

[0077] Step S2-2: Input the surface deformation point cloud data into the coordinate conversion module of the geographic information system, align it to the unified spatial reference, and generate a second data set.

[0078] In step S1-1, distributed fiber optic sensors are arranged symmetrically along the axis of the pile foundation, with one measuring line located on the side facing the water to monitor the effects of ship loads, and the other on the side facing the slope to detect the effects of slope soil pressure. The sensor network covers the top, middle and bottom areas of the pile, and is fixed in sections with underwater epoxy resin to ensure continuous monitoring of the entire pile length. The original strain data set contains temporally and spatially correlated strain distribution and temperature change data. The Kalman filter module eliminates noise interference caused by wave impact or temperature fluctuations based on the state-space model to generate a first data set with a high signal-to-noise ratio. In step S2-1, the three-dimensional laser scanner obtains surface deformation point cloud data using multi-perspective scanning. The geographic information system coordinate conversion module aligns the point cloud data to the global coordinate system of the terminal, eliminates equipment posture errors, and generates a second data set with a unified spatial reference.

[0079] Symmetrically arranged fiber optic sensors enhance data representativeness, Kalman filtering improves strain data reliability, and coordinate registration ensures spatial consistency of deformation data. For example, survey lines facing the waterside prioritize capturing sudden strain changes caused by ship impacts, while Kalman filtering mitigates the impact of diurnal temperature fluctuations on long-term monitoring. Coordinate registration of laser scanning data reduces errors in the calculation of pile cap displacement and pile inclination angle, preventing misjudgment of deformation due to coordinate system offset.

[0080] Among the alternative solutions, Kalman filtering can be replaced by wavelet transform denoising; geographic information system coordinate conversion can use ICP (iterative closest point) algorithm or feature matching algorithm; 3D laser scanner can be replaced by structured light scanner, but the resolution and scanning speed need to be balanced.

[0081] In conjunction with the first aspect, in some implementations of the first aspect, step S4 includes:

[0082] Step S4-1: Input the structural strain data in the first data set into the Hilbert-Huang transform module, decompose it into multiple intrinsic mode functions, extract the main frequency components and transient energy characteristics, and generate frequency domain feature vectors;

[0083] Step S4-2: Input the surface deformation data in the second data set into the iterative closest point algorithm module, remove noise points and perform registration based on geographic information system coordinates, extract the pile cap displacement and pile body inclination angle, and generate a deformation gradient vector;

[0084] Step S4-3: Inputting the ultrasonic tomography raw signal data in the third data set into the back-projection reconstruction algorithm module to analyze the geometric parameters and position information of the internal defects and generate a tomographic imaging map;

[0085] Step S4-4: Input the frequency domain feature vector, deformation gradient vector and tomographic imaging map into the normalization module, eliminate the dimensionality difference, and perform dimensionality reduction through principal component analysis to generate standardized multi-source heterogeneous data.

[0086] In step S4-1, the Hilbert-Huang transform module performs empirical mode decomposition on the structural strain data in the first dataset, decomposing the non-stationary signal into multiple intrinsic mode functions (IMFs). The dominant frequency components and transient energy characteristics of the top three IMF components with the highest energy contribution are extracted to generate frequency domain eigenvectors. In step S4-2, the iterative closest point algorithm module registers the surface deformation point cloud data in the second dataset, removes outlier noise points, and calculates the pile cap displacement vector (including horizontal displacement and vertical settlement) and pile body inclination angle based on geographic information system coordinates to generate a deformation gradient vector. In step S4-3, the back-projection reconstruction algorithm module analyzes the raw ultrasonic tomography signal data in the third dataset. Based on the acoustic wave propagation time difference and amplitude attenuation coefficient, the three-dimensional spatial coordinates and equivalent aperture of the internal defects are reconstructed to generate a tomographic image spectrum. In step S4-4, the normalization module unifies the dimensionality of the frequency domain eigenvectors, deformation gradient vectors, and tomographic image spectrum. Principal component analysis is then used to select the top three principal components with the highest contribution rates to generate standardized multi-source heterogeneous data.

[0087] The technical benefits are as follows: the Hilbert-Huang transform effectively processes non-stationary strain signals, the iterative closest point algorithm improves the accuracy of deformation data registration, back-projection reconstruction enables 3D visualization of defects, and principal component analysis reduces data dimensionality. For example, transient energy signatures can identify instantaneous damage caused by overload, deformation gradient vectors quantify pile tilt trends, and tomographic imaging maps locate internal cracks. Normalization of these three methods eliminates dimensional differences, providing high-quality input for deep learning models.

[0088] Among the alternative solutions, the Hilbert-Huang transform can be replaced by the wavelet transform; the iterative closest point algorithm can be replaced by the NDT (normal distribution transform) registration; and the back projection reconstruction algorithm can be replaced by the filtered back projection or iterative reconstruction algorithm.

[0089] In conjunction with the first aspect, in some implementations of the first aspect, the dimensionality reduction by principal component analysis in step S4-4 includes:

[0090] Step S4-4-1: Input the frequency domain eigenvector, deformation gradient vector and tomographic imaging map into the covariance matrix calculation module to generate the covariance matrix of the multi-source data;

[0091] Step S4-4-2: Perform eigenvalue decomposition on the covariance matrix, select the top three principal components with the highest contribution rate, and generate the eigenvector after dimensionality reduction.

[0092] In step S4-4-1, the covariance matrix calculation module receives frequency-domain eigenvectors, deformation gradient vectors, and tomographic images, calculates the covariance matrix between multi-source data, and quantifies the correlations between each feature dimension. In step S4-4-2, the eigenvalue decomposition module decomposes the covariance matrix, sorts by eigenvalue, and selects the top three principal components with a cumulative contribution rate of at least 85% to generate a reduced-dimensional eigenvector. For example, the main frequency component in the frequency-domain eigenvector may show a high correlation with the pile inclination angle in the deformation gradient vector. Principal component analysis can fuse these into a single-dimensional feature to reduce data redundancy.

[0093] The technical benefits are as follows: the covariance matrix reveals the inherent correlations between multi-source data, eigenvalue decomposition retains key information, and the dimensionality-reduced eigenvectors improve model training efficiency. For example, if strain frequency domain features are highly correlated with deformation gradients, principal component analysis can combine them into a comprehensive damage index to avoid model overfitting. Contribution rate thresholds are set to balance data compression and information integrity, adapting to the monitoring needs of different terminals.

[0094] In alternative solutions, the contribution rate threshold can be adjusted to 80% or 90%; the number of principal components can be dynamically adjusted according to the actual data characteristics.

[0095] In conjunction with the first aspect, in some implementations of the first aspect, in step S4-3, analyzing the geometric parameters and location information of the internal defect includes:

[0096] According to the relationship between the acoustic wave propagation time and amplitude attenuation, the three-dimensional spatial coordinates and equivalent aperture of the internal defects are reconstructed.

[0097] In step S4-3, the back-projection reconstruction algorithm calculates the acoustic wave propagation time between the transmitter and receiver based on the difference in acoustic wave propagation speed in concrete. Combined with the amplitude attenuation coefficient (which is negatively correlated with defect size), it reconstructs the three-dimensional spatial coordinates of the internal defect. The equivalent aperture is calculated based on the acoustic wave diffraction effect and reflects the equivalent size of the defect. For example, cracks extend the acoustic wave propagation path and increase amplitude attenuation. By cross-validating multiple sets of sensor signals, the crack depth and width can be accurately calculated. Hole defects manifest as a missing acoustic wave reflection signal, and their spatial location is determined through triangulation.

[0098] The technical benefits are as follows: dual-parameter analysis of acoustic wave propagation time and amplitude attenuation improves defect detection accuracy, while equivalent aperture quantifies defect severity. For example, propagation time differences can distinguish surface cracks from deep cavities, while the amplitude attenuation coefficient combined with equivalent aperture assesses crack propagation risk, providing a quantitative basis for repair decisions.

[0099] In alternative solutions, the acoustic wave parameters can be supplemented with frequency component analysis; the equivalent aperture calculation can adopt geometric diffraction theory or numerical simulation methods.

[0100] In conjunction with the first aspect, in some implementations of the first aspect, step S5 includes:

[0101] Step S5-1: Input multi-source heterogeneous data into the convolutional neural network, extract spatial local features through multi-layer convolution kernels, and generate a damage-sensitive feature map;

[0102] Step S5-2: Input the damage-sensitive feature map into the long short-term memory network, analyze the temporal variation pattern, and generate a damage feature vector that integrates spatiotemporal correlation.

[0103] In step S5-1, a convolutional neural network receives multi-source heterogeneous data as input and extracts local spatial features layer by layer through multiple convolutional kernels. For example, the first convolutional kernel layer can identify local abrupt changes in strain data (such as stress concentration at the crack tip), the second convolutional kernel layer captures macroscopic displacement patterns in deformation data (such as the tilt trend of the pile cap), and the third convolutional kernel layer interprets the defect geometry (such as crack length and direction) in the tomographic image. In step S5-2, a long-short-term memory network receives the damage-sensitive feature map output by the convolutional neural network and analyzes temporal correlation patterns through time-step expansion. For example, the cumulative strain effect caused by overloading is manifested as an increasing gradient in the feature map, while ship impacts exhibit instantaneous pulse characteristics. The long-short-term memory network uses a gating mechanism to select key temporal nodes and generate a damage feature vector that integrates temporal and spatial correlations.

[0104] The technical benefits are reflected in the combination of the spatial feature extraction capabilities of convolutional neural networks and the temporal modeling capabilities of long-short-term memory networks, enabling precise analysis of complex damage patterns. For example, spatial features can locate crack propagation areas, while temporal features can distinguish between transient impacts and long-term degradation. The fusion of these two allows for accurate determination of the damage development stage (e.g., initial microcracks or severe fractures).

[0105] Among the alternative solutions, the convolutional neural network can be replaced by a visual model with a Transformer architecture, and the long short-term memory network can be replaced by a temporal convolutional network; the generation of damage-sensitive feature maps can use an attention mechanism to enhance the feature weights of key areas.

[0106] In conjunction with the first aspect, in some implementations of the first aspect, in step S6, generating the dynamically updated damage threshold library includes:

[0107] Step S6-1: training the initial damage threshold library using sample data in the historical damage database to generate benchmark matching parameters;

[0108] Step S6-2: Input the first data set, the second data set, and the third data set into the adaptive update module, dynamically adjust the benchmark matching parameters based on the real-time collected structural strain data, surface deformation data, and ultrasonic tomography raw signal data, and generate a dynamically updated damage threshold library.

[0109] In step S6-1, the historical damage database contains sample data for a variety of typical damage conditions, such as ship impact, overloading, and material corrosion. The initial damage threshold library is generated through supervised learning training, and the benchmark matching parameters include damage location confidence thresholds and damage level classification boundaries. In step S6-2, the adaptive update module receives the first dataset (strain data), the second dataset (deformation data), and the third dataset (defect data) collected in real time, and dynamically adjusts the benchmark matching parameters using an online learning algorithm (such as an incremental support vector machine). For example, if long-term monitoring reveals that the material elastic modulus has decreased due to corrosion, the adaptive module will lower the absolute value of the strain threshold to match the degraded structural state.

[0110] The technical benefits are as follows: the dynamic threshold library adapts to evolving dock structural performance and environmental changes through online learning, preventing false positives or omissions caused by static threshold failures. For example, when strain baselines drift due to material aging, dynamically adjusted thresholds can still accurately identify abnormal strains. New damage modes (such as new types of ship impacts) can be incorporated into the threshold library through incremental learning, improving the system's generalization capabilities.

[0111] In alternative solutions, the online learning algorithm can adopt random forest incremental training or Bayesian probability model; the benchmark matching parameters can be expanded to distance measurement parameters of multidimensional feature space.

[0112] In conjunction with the first aspect, in some implementations of the first aspect, the deploying of the distributed optical fiber sensor network in step S1 includes:

[0113] Two measuring lines are marked symmetrically along the axial direction on the pile foundation surface, one of which is located on the side facing the water and the other on the side facing the slope, to generate measuring line positioning information;

[0114] The distributed optical fiber is laid along the survey line to locate the information and fixed to the pile surface in three sections using underwater epoxy resin, located at the top, middle and bottom of the pile respectively, forming an optical fiber sensing network for continuous monitoring of the entire pile length.

[0115] During the survey line positioning phase, the survey line facing the water side is laid on the waterfront side of the pile foundation to monitor dynamic strain caused by ship berthing and wave impact. The survey line facing the bank side is laid on the backwater side of the pile foundation to detect static strain offsets caused by bank soil settlement or slip. After the distributed optical fiber is laid along the survey line, it is fixed in three sections with underwater epoxy resin: high-viscosity epoxy resin is used at the top of the pile to resist erosion by water flow, flexible epoxy resin is used in the middle of the pile to accommodate micro-bending deformation, and corrosion-resistant epoxy resin is used at the bottom of the pile to prevent the attachment of marine organisms. The fiber-optic sensing network, which continuously monitors the entire length of the pile, covers both the surface and underwater areas, and uses wavelength division multiplexing technology to achieve parallel data collection at multiple measurement points.

[0116] The technical benefits are as follows: symmetrical measurement line layout enhances data comparability, and the segmented fixation strategy improves the environmental adaptability of the sensor network. For example, the measurement line facing the water side prioritizes capturing ship impact signals, while the segmented fixation design prevents fiber loss caused by pile deformation. Wavelength division multiplexing technology supports a high-density measurement point layout, achieving high-resolution monitoring of strain distribution.

[0117] In alternative solutions, the number of measuring lines can be expanded to four (cross-symmetrically arranged), but the hardware cost will increase; underwater epoxy resin can be replaced with mechanical fixtures, but this may introduce stress concentration interference.

[0118] In conjunction with the first aspect, in some implementations of the first aspect, the training process of the deep learning model includes:

[0119] Finite element simulation is used to simulate the structural strain data, surface deformation data and ultrasonic fault raw signal data of high-pile docks under different damage conditions to generate virtual training samples;

[0120] Integrate virtual training samples with historical monitoring data and field loading test data to construct a training sample set covering multiple scenarios;

[0121] Input the training sample set into the adversarial generative network to generate an enhanced training set;

[0122] The convolutional neural network and long short-term memory network are trained by enhancing the training set to generate a deep learning model.

[0123] During the finite element simulation phase, a parametric model of a high-pile wharf was established. Virtual training samples were generated by varying the load type (concentrated force, distributed force), damage location (pile top, pile middle, pile bottom), and damage size (crack length, hole diameter). Historical monitoring data was used to screen typical damage events (such as caisson displacement data after a typhoon). Field loading tests were conducted by applying stepped loads and recording the structural responses to construct a training sample set covering multiple scenarios. A generative adversarial network (GAN) synthesizes rare damage patterns (such as explosive silt-squeezing damage at the pile bottom) through a generator, while a discriminator distinguishes between real and synthesized data, generating a training set with enhanced diversity. A convolutional neural network and a long short-term memory network were trained end-to-end jointly. The loss function combines spatial reconstruction error with temporal prediction error to optimize model parameters.

[0124] The technical benefits are as follows: virtual training samples fill the gap in measured data, and adversarial generation enhances the model's ability to identify rare damage. For example, finite element simulation can generate damage data for the underwater area at the bottom of the pile (a traditional blind spot for detection), while adversarial generation network-generated data for extreme working conditions (such as earthquakes) enhances the model's robustness.

[0125] Among the alternative solutions, finite element simulation can be replaced by discrete element model; generative adversarial network can be replaced by variational autoencoder; and the joint training strategy can be changed to staged pre-training and fine-tuning.

[0126] Example 2

[0127] like Figure 4 As shown, in a second aspect, the present invention provides a high-pile wharf structure damage identification system, the system adopts the method provided in any of the above embodiments, and the system includes:

[0128] The data acquisition module includes a distributed optical fiber sensor network, a three-dimensional laser scanner, and an ultrasonic tomography scanner, which is used to collect structural strain data, temperature data, surface deformation data, and ultrasonic tomography raw signal data of the high-pile wharf in real time to generate a first data set, a second data set, and a third data set respectively;

[0129] a data processing module connected to the data acquisition module, configured to perform feature extraction and standardization processing on the first data set, the second data set, and the third data set, including dimensionality reduction through principal component analysis, elimination of dimensional differences using a normalization algorithm, and alignment of feature vectors based on geographic information system coordinates to generate multi-source heterogeneous data;

[0130] The model processing module is connected to the data processing module and includes a pre-trained deep learning model composed of a convolutional neural network and a long short-term memory network. The deep learning model is used to extract spatial features and perform temporal correlation analysis on multi-source heterogeneous data to generate a fused damage feature vector.

[0131] The damage identification module is connected to the model processing module and includes a dynamically updated damage threshold library and a matching calculation unit, which is used to perform similarity matching based on the damage feature vector and output the damage location and damage level.

[0132] The recognition system provided in this embodiment of the present invention corresponds to the recognition method provided in Example 1. Multi-source data (strain, deformation, defect) is integrated into multi-source heterogeneous data aligned in time and space, providing high-dimensional input for the deep learning model; the damage feature vector output by the model is matched with the dynamic threshold library, realizing a full-link closed loop from "data acquisition-feature fusion-damage decision".

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than 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 they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying structural damage of a high-pile wharf, characterized in that: include: Step S1: collecting structural strain data and temperature data of the high-pile wharf in real time based on a distributed optical fiber sensing network to generate a first data set; Step S2: acquiring surface deformation data of the high-pile wharf by a three-dimensional laser scanner to generate a second data set; Step S3: using an ultrasonic tomography scanner to perform non-destructive testing on the internal structure of the high-pile wharf, obtaining ultrasonic tomography raw signal data, and generating a third data set; Step S4: performing feature extraction and normalization processing on the first dataset, the second dataset, and the third dataset, including dimensionality reduction through principal component analysis, eliminating dimensional differences using a normalization algorithm, and aligning feature vectors based on geographic information system coordinates to generate multi-source heterogeneous data; Step S4 includes: Step S4-1: inputting the structural strain data in the first data set into a Hilbert-Huang transform module, decomposing it into multiple intrinsic mode functions, extracting the main frequency components and transient energy characteristics, and generating a frequency domain feature vector; Step S4-2: Inputting the surface deformation data in the second data set into an iterative closest point algorithm module, removing noise points and performing registration based on the geographic information system coordinates, extracting the pile cap displacement and the pile body inclination angle, and generating a deformation gradient vector; Step S4-3: inputting the ultrasonic tomography raw signal data in the third data set into a back-projection reconstruction algorithm module, analyzing the geometric parameters and position information of the internal defects, and generating a tomographic imaging map; Step S4-4: Inputting the frequency domain feature vector, the deformation gradient vector, and the tomographic imaging map into a normalization module, eliminating dimensional differences, and performing dimensionality reduction through principal component analysis to generate standardized multi-source heterogeneous data; Step S5: Inputting the multi-source heterogeneous data into a pre-trained deep learning model, extracting spatial features through a convolutional neural network, and analyzing temporal associations through a long short-term memory network to generate a fused damage feature vector; Step S6: Based on the damage feature vector, similarity matching is performed in combination with the dynamically updated damage threshold library to output the damage location and damage level.

2. A method for identifying structural damage of a high-pile wharf according to claim 1, characterized in that: Step S1 includes: Step S1-1: Using distributed optical fiber sensors symmetrically arranged along the axis of the pile foundation, the strain distribution data on the pile foundation surface and the ambient temperature change data are collected to generate an original strain data set; Step S1-2: inputting the original strain data set into a Kalman filter module to remove noise interference data to generate a first data set; Step S2 includes: Step S2-1: Scan the pile cap, pile body, and panel of the high-pile wharf at multiple angles using a 3D laser scanner to obtain surface deformation point cloud data; Step S2-2: inputting the surface deformation point cloud data into a geographic information system coordinate conversion module, registering it to a unified spatial reference, and generating a second data set.

3. A method for identifying structural damage of a high-pile wharf according to claim 1, characterized in that: The dimensionality reduction by principal component analysis in step S4-4 includes: Step S4-4-1: inputting the frequency domain eigenvector, the deformation gradient vector and the tomographic imaging map into a covariance matrix calculation module to generate a covariance matrix of multi-source data; Step S4-4-2: Perform eigenvalue decomposition on the covariance matrix, select the top three principal components with the highest contribution rate, and generate a eigenvector after dimensionality reduction.

4. A method for identifying structural damage of a high-pile wharf according to claim 1, characterized in that: In step S4-3, the geometric parameters and location information of the internal defects are analyzed, including: According to the relationship between the acoustic wave propagation time and amplitude attenuation, the three-dimensional spatial coordinates and equivalent aperture of the internal defects are reconstructed.

5. The method for identifying structural damage of a high-pile wharf according to claim 1, characterized in that: The step S5 comprises: Step S5-1: inputting the multi-source heterogeneous data into a convolutional neural network, extracting spatial local features through multi-layer convolution kernels, and generating a damage-sensitive feature map; Step S5-2: Input the damage-sensitive feature map into the long short-term memory network, analyze the temporal variation pattern, and generate a damage feature vector that integrates spatiotemporal correlation.

6. A method for identifying structural damage of a high-pile wharf according to claim 1, characterized in that: In step S6, the generation of the dynamically updated damage threshold library includes: Step S6-1: training the initial damage threshold library using sample data in the historical damage database to generate benchmark matching parameters; Step S6-2: Input the first data set, the second data set, and the third data set into the adaptive update module, dynamically adjust the benchmark matching parameters based on the real-time collected structural strain data, surface deformation data, and ultrasonic tomography raw signal data, and generate a dynamically updated damage threshold library.

7. The method for identifying structural damage of a high-pile wharf according to claim 1, characterized in that: The deployment of the distributed optical fiber sensor network in step S1 includes: Two measuring lines are marked symmetrically along the axial direction on the pile foundation surface, one of which is located on the side facing the water and the other on the side facing the slope, to generate measuring line positioning information; The distributed optical fiber is laid along the survey line to locate the information and fixed to the pile surface in three sections using underwater epoxy resin, located at the pile top, pile middle and pile bottom areas respectively, to form an optical fiber sensing network for continuous monitoring of the entire pile length.

8. The method for identifying structural damage of a high-pile wharf according to claim 1, characterized in that: The training process of the deep learning model includes: Finite element simulation is used to simulate the structural strain data, surface deformation data and ultrasonic fault raw signal data of high-pile docks under different damage conditions to generate virtual training samples; Integrating the virtual training samples with historical monitoring data and field loading test data to construct a training sample set covering multiple scenarios; Inputting the training sample set into a generative adversarial network to generate an enhanced training set; The convolutional neural network and the long short-term memory network are trained using the enhanced training set to generate the deep learning model.

9. A high-pile wharf structure damage identification system, characterized in that: The system adopts the method according to any one of claims 1 to 8, and the system includes: The data acquisition module includes a distributed optical fiber sensor network, a three-dimensional laser scanner, and an ultrasonic tomography scanner, which is used to collect structural strain data, temperature data, surface deformation data, and ultrasonic tomography raw signal data of the high-pile wharf in real time to generate a first data set, a second data set, and a third data set respectively; A data processing module is connected to the data acquisition module and is used to perform feature extraction and standardization processing on the first data set, the second data set, and the third data set, including dimensionality reduction through principal component analysis, eliminating dimensional differences using a normalization algorithm, and aligning feature vectors based on geographic information system coordinates to generate multi-source heterogeneous data, specifically including: Step S4-1: Input the structural strain data in the first data set into the Hilbert-Huang transform module, decompose it into multiple intrinsic mode functions, extract the main frequency components and transient energy characteristics, and generate frequency domain feature vectors; Step S4-2: Inputting the surface deformation data in the second data set into an iterative closest point algorithm module, removing noise points and performing registration based on the geographic information system coordinates, extracting the pile cap displacement and the pile body inclination angle, and generating a deformation gradient vector; Step S4-3: inputting the ultrasonic tomography raw signal data in the third data set into a back-projection reconstruction algorithm module, analyzing the geometric parameters and position information of the internal defects, and generating a tomographic imaging map; Step S4-4: Inputting the frequency domain feature vector, the deformation gradient vector, and the tomographic imaging map into a normalization module, eliminating dimensional differences, and performing dimensionality reduction through principal component analysis to generate standardized multi-source heterogeneous data; a model processing module, connected to the data processing module, comprising a pre-trained deep learning model composed of a convolutional neural network and a long short-term memory network, for performing spatial feature extraction and temporal correlation analysis on the multi-source heterogeneous data to generate a fused damage feature vector; The damage identification module is connected to the model processing module and includes a dynamically updated damage threshold library and a matching calculation unit, which is used to perform similarity matching based on the damage feature vector and output the damage location and damage level.

Images