Quality inspection method for quay armor blocks

By integrating multi-source data fusion and a closed-loop feedback mechanism, combined with 3D laser scanning and frequency-tunable flaw detection devices, full-dimensional inspection of the wharf revetment blocks was achieved, solving the problems of low efficiency and insufficient accuracy in existing technologies and providing scientific maintenance decision support.

WO2026232678A1PCT designated stage Publication Date: 2026-11-12CHINA HARBOUR ENGINEERING
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2025/104312
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2025-06-27
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

In existing technologies, the detection efficiency of wharf revetment blocks is low and the accuracy is insufficient, making it difficult to adapt to complex marine environments. The detection results are mostly presented in two-dimensional reports, which cannot intuitively reflect the three-dimensional spatial relationship of defects. This leads to maintenance plans relying on experience-based judgment, which carries the risk of over-repair or under-repair.

Method used

Employing a multi-source data fusion and closed-loop feedback mechanism, point cloud data is acquired through a mobile 3D laser scanning device, surface features are extracted using a deep convolutional neural network, internal inspection is performed using a frequency-tunable dual-mode flaw detection device, a 3D defect distribution map is established by combining multi-sensor data fusion, compressive strength testing and infrared spectroscopy analysis are conducted, a dynamic evaluation model is established, and a 3D visualization inspection report is generated.

Benefits of technology

It enables full-dimensional detection of the facing blocks, significantly improving the reliability of structural safety assessment and the scientific nature of maintenance decisions, reducing the risk of missed detections, improving detection efficiency and accuracy, and providing multi-dimensional data support and intelligent decision support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2025104312_12112026_PF_FP_ABST
    Figure CN2025104312_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A quality inspection method for quay armor blocks, relating to the technical field of concrete structure inspection. The method comprises: constructing a surface model by means of a mobile three-dimensional laser scanning device, and identifying a high-risk flaw region; in the high-risk region, using a frequency-adjustable dual-mode flaw detection device to perform high-density scanning and impact-echo testing, and generating a three-dimensional flaw distribution map; drilling core samples in the flaw region for compressive strength testing, and analyzing material properties on the basis of non-contact infrared spectroscopy; and establishing a dynamic assessment model, calculating a safety coefficient on the basis of surface flaws, internal flaws, and material variability data, and generating a three-dimensional visual report. The present invention solves the problems of low efficiency of conventional inspection methods, insufficient accuracy in locating internal flaws, and incomplete assessment of material properties. By means of multi-source data fusion and a closed-loop feedback mechanism, full-dimensional inspection from the surface to the interior and from the macroscopic level to the microscopic level is achieved, thereby significantly improving the reliability of structural safety assessment and the scientificity of maintenance decisions.
Need to check novelty before this filing date? Find Prior Art

Description

Methods for quality inspection of wharf revetment blocks Technical Field

[0001] This invention relates to the field of concrete structure testing technology, and in particular to a method for testing the quality of wharf revetment blocks. Background Technology

[0002] In port and coastal engineering, wharf revetment blocks, as the core protective structure against wave impact, are subjected to complex marine environmental loads over long periods, making them prone to damage such as surface cracking, internal voids, and material aging. These defects can significantly reduce structural durability and even trigger localized collapses. Therefore, regular quality inspections of revetment blocks are crucial for ensuring the safe operation of ports.

[0003] In existing technologies, the inspection of revetment blocks mainly relies on the following methods: Manual visual inspection: Technicians observe surface cracks by climbing or approaching from a vessel. This method is limited by the experience of the inspectors, cannot quantify crack depth, and carries high risks associated with working at heights or near water. For example, a port experienced a dam collapse due to underwater cracks being missed during manual inspection. Single-point ultrasonic testing: A handheld flaw detector is used to locally scan suspected areas. Due to the high surface roughness of the revetment blocks (Ra > 12.5 μm), ultrasonic waves are easily scattered, resulting in an internal void detection rate of less than 60%, and the inability to obtain three-dimensional defect distribution information. Destructive core sampling: Randomly drilled core samples for laboratory strength testing. This method has a low sampling rate (typically < 0.1%), compromises structural integrity, and cannot reflect the spatial heterogeneity of material properties. One study showed that traditional core sampling methods may miss more than 80% of micro-defects in the interface transition zone.

[0004] Existing technologies suffer from the following key problems: a contradiction between detection efficiency and coverage: high-precision detection (such as laser scanning) requires long-term static operation, making it difficult to adapt to the detection window limitations caused by tidal changes; lack of multi-scale defect collaborative analysis: surface morphology, internal defects, and material performance data are isolated, lacking a fusion evaluation model; insufficient suppression of environmental interference: temperature and humidity fluctuations and mechanical vibrations in the marine environment can easily lead to data distortion, for example, the signal-to-noise ratio of ultrasonic echo signals is often below 20dB. Furthermore, existing detection results are mostly presented in two-dimensional reports, failing to intuitively reflect the three-dimensional spatial relationship of defects, leading to maintenance plan formulation relying on experience-based judgment, and posing a risk of over-repair or under-repair. For example, in one project, a 30% waste of reinforcement material resulted from misjudging the crack direction using two-dimensional drawings.

[0005] The aforementioned problems indicate an urgent need for a high-efficiency, accurate, and adaptable full-lifecycle inspection technology for revetment blocks to meet the requirements of intelligent operation and maintenance in modern ports. Summary of the Invention

[0006] This invention overcomes the problems of low efficiency, insufficient accuracy in locating internal defects, and one-sided material performance evaluation in traditional detection methods. It provides a method for detecting the quality of wharf revetment blocks. Through multi-source data fusion and closed-loop feedback mechanism, it achieves full-dimensional detection from surface to interior and from macro to micro, which significantly improves the reliability of structural safety assessment and the scientific nature of maintenance decisions.

[0007] To achieve the above objectives, the present invention adopts the following solution:

[0008] The method for quality inspection of wharf revetment blocks includes the following steps:

[0009] S1: A mobile 3D laser scanning device is used to scan the protective block to obtain laser point cloud data. A 3D surface model is constructed through a real-time point cloud registration algorithm. Surface geometric features, including surface crack depth and peeling area, are extracted based on a deep convolutional neural network. Areas where the surface crack depth exceeds a preset depth threshold or the peeling area is greater than a preset area threshold are marked as high-risk areas, and a heat map of high-risk areas is generated.

[0010] S2: Based on the thermal map of high-risk areas, a frequency-adjustable dual-mode flaw detection device is used to detect voids and cracks inside the protective block: an ultrasonic probe is used to perform intensified scanning in high-risk areas; multiple impact points are arranged in the projection area of ​​the surface crack extension direction according to the stress diffusion path, and an array-type impact echo device is used to collect vibration signals at the impact points;

[0011] A three-dimensional defect probability distribution map is established by fusing ultrasonic data and vibration signal data through a multi-sensor data fusion algorithm, and areas where the diameter of internal cavities or the length of cracks are greater than the corresponding upper limit threshold are marked as severe defect areas.

[0012] S3: Drill multiple core samples within a predetermined distance around the severely defective area and conduct compressive strength tests and acoustic emission damage detection; conduct infrared spectral composition surveys in non-severely defective areas to detect concrete carbonation and chloride ion erosion index; in the defect-matrix interface transition zone, set up micro-area hardness test points according to the material property gradient and conduct hardness tests.

[0013] S4: Establish a dynamic evaluation model that includes surface defect distribution entropy, internal defect volume ratio, and material property dispersion coefficient; calculate the comprehensive safety factor S using a multi-objective optimization algorithm, and compare it with the upper limit of the safety threshold S. u and lower limit S d The comparison is performed to generate a 3D visualization inspection report.

[0014] Preferably, in step S1, the robotic arm of the mobile 3D laser scanning device integrates a dual-mode lidar and an inertial measurement unit, and performs the following operations:

[0015] Based on continuous three-point cloud slice data, a differential geometry algorithm is used to calculate the local curvature change rate of the surface of the protective block in real time. The curvature change rate is quantified by the ratio of the Gaussian curvature difference between adjacent measurement points to the arc length distance. When the local minimum curvature radius is less than 2 meters, the surface roughness index exceeds the standard defined threshold of 0.5 mm, and environmental vibration in the range of 0.1-5 Hz is detected, the high frame rate lidar mode is activated, and the scanning frame rate is increased to no less than 200 Hz.

[0016] An adaptive Kalman filter is used to eliminate low-frequency vibration noise caused by waves. By fusing the state equations of the robot arm's pitch angle, angular velocity, and linear acceleration with the observation data of the inertial measurement unit, the process noise and observation noise covariance matrix is ​​updated in real time to suppress wave interference.

[0017] In high frame rate LiDAR mode, the movement speed of the robotic arm end effector is dynamically adjusted according to the local minimum radius of curvature. When the minimum radius of curvature is less than 1 meter, a low-speed scan of 0.2 m / s is used, and when the radius of curvature is in the range of 1-2 meters, a medium-speed scan of 0.5 m / s is used. At the same time, the laser line density is increased to 800 lines per meter, and the point cloud overlap rate of adjacent scanning paths is ensured to be no less than 40%.

[0018] Preferably, in step S1, the real-time point cloud registration algorithm specifically includes the following steps:

[0019] Coarse registration is performed: an improved iterative nearest point algorithm is adopted, with the weighted sum of geometric distance and curvature difference as the optimization objective. The curvature difference is quantified by the mixed difference between the Gaussian curvature and the average curvature of the corresponding point, and the Gaussian curvature weight factor is set to 0.5. When screening matching point pairs, only point pairs with a principal curvature radius ratio greater than 0.7 and a cosine value of the angle between the normal vectors greater than 0.9 are retained, and mismatches with inconsistent surface geometric features are excluded.

[0020] Fine registration is performed: the Gauss-Newton iterative algorithm is accelerated based on field-programmable gate arrays, the Jacobian matrix containing rotation and translation components is calculated in parallel, and the optimal transformation parameters are solved by damped least squares method. The damping factor is set to 0.0001 to prevent matrix singularity. Combined with 3D feature descriptor matching, the local surface is divided into 32 orientation intervals and the height distribution histogram is statistically analyzed. The chi-square distance is used to calculate feature similarity, and regions with matching scores below the threshold are marked as low confidence regions.

[0021] Dynamic evaluation and replanning are performed: the registration error of the local area is calculated in real time. When the root mean square error of the area exceeds 0.8 mm, it is determined to be a registration failure area; the scanning path replanning is triggered, and a ±15 degree deflection angle is added to the original scanning path to increase the point cloud line density to 1200 lines per meter and the overlap rate to 60%. The low confidence area is also re-scanned.

[0022] Preferably, in step S2, the tunable dual-mode flaw detection device includes an ultrasonic probe and an array-type impact echo device, wherein the ultrasonic probe adopts the following layered waveguide structure and dynamic control strategy:

[0023] The waveguide layer of the ultrasonic probe consists of a piezoelectric composite substrate and five gradient impedance matching layers. The impedance value of each layer gradually changes exponentially from the substrate to the coupling agent interface. The total thickness of the matching layer is 2 mm to suppress sound energy scattering caused by surface roughness. When the arithmetic mean deviation of surface roughness is detected to exceed 12.5 micrometers, a wideband sweep frequency detection mode is activated. The frequency is continuously swept at a rate of 50 kHz per second in the transmission frequency range of 1.5 MHz to 2.5 MHz, and the transmission pulse width is dynamically adjusted according to the sweep frequency range.

[0024] The effective focusing depth is calculated based on the longitudinal wave velocity of the material and the current transmission frequency. When the focusing depth fluctuates beyond ±10% of the nominal value, the calibration mode is triggered: a 1 MHz reference pulse is emitted, and the actual wave velocity of the material is corrected in reverse through the echo time difference to eliminate the sound velocity error caused by concrete aging.

[0025] A nonlinear relationship model between contact pressure and surface roughness was established. The probe height was adjusted in real time using a piezoelectric actuator to stabilize the coupling agent pressure in the range of 0.2 to 0.5 MPa. A proportional-integral-derivative controller was used to achieve dynamic balance, with the proportional coefficient set to 0.8, the integral time set to 0.1 seconds, the derivative time set to 0.05 seconds, and the pressure sensor sampling frequency set to 1 kHz.

[0026] Preferably, in step S2, the array-type impact echo device locates internal defects using the following method:

[0027] The impact energy is controlled to five levels by adjusting the impact speed of the array-type impact echo device: 0.5J, 1J, 2J, 3.5J, and 5J; the impact frequency is dynamically selected according to the Vickers hardness of the surface: 10 Hz high-frequency impact is used when the hardness is below 300 MPa, 5 Hz medium-frequency impact is used when the hardness is in the range of 300-500 MPa, and 2 Hz low-frequency impact is used when the hardness exceeds 500 MPa.

[0028] The surface wave velocity is calculated based on the material's elastic modulus, density, and Poisson's ratio. The surface wave wavelength is calculated based on the central analysis frequency and the surface wave velocity. Piezoelectric sensors are arranged at intervals one-eighth of the surface wave wavelength. Several piezoelectric sensors are arranged in a spiral-grid hybrid layout to form a sensor array covering three-dimensional space.

[0029] The vibration signal acquired by the echo device is fused with short-time Fourier transform and wavelet packet decomposition. The short-time Fourier transform uses a Gaussian-modulated cosine window function, with the window width dynamically adjusted according to the impact frequency. The wavelet packet decomposition uses a 10th-order Daubechies wavelet basis, based on the sampling frequency f. sample With center frequency f central Set the number of decomposition layers to

[0030] A dual-channel analysis method with high and low frequencies is adopted to calculate the signal energy integral value of the low-frequency channel (0.5-10 kHz) and the kurtosis index of the high-frequency channel (10-50 kHz).

[0031] Based on the data from the sensor array, a time-reversal algorithm is used to reconstruct the three-dimensional wave field, with the observation time window length set to 50 milliseconds. When the energy integral value of the low-frequency channel signal exceeds 0.8 times the reference value of healthy concrete, the kurtosis index value of the high-frequency channel exceeds the threshold, and the diameter of the three-dimensional wave field focusing area is greater than 10 centimeters, the excess area is judged as a valid internal defect.

[0032] Preferably, step S3, core sampling and testing, specifically includes the following steps:

[0033] Based on the fusion data of laser displacement sensor and inertial measurement unit, the drill bit axis offset angle is calculated in real time: the lateral offset is measured by laser sensor, and mechanical vibration error is compensated by gyroscope angular velocity data, with the compensation coefficient set to 0.05 seconds / radian; the stroke of four hydraulic cylinders is adjusted by fuzzy proportional-integral-derivative controller, with the goal of minimizing the weighted sum of offset angle and axial force change rate, and the maximum deflection angle is limited to ±15 degrees, triggering an alarm when the limit is exceeded;

[0034] The dimensionless working condition index is calculated in real time and defined as the product of axial drilling force and drill radius divided by the product of feed rate and hydraulic pressure. When the index exceeds the set threshold, three levels of protection are implemented: the retraction speed is adjusted to twice the current feed rate plus 50 mm / s; the drill speed is reduced to 300 rpm; and coolant is sprayed at a flow rate of 0.2 times the axial force.

[0035] The following reinforcement methods were used in the core sample compressive strength test:

[0036] The loading rate is controlled in three stages: a slow loading rate of 0.5 MPa / s is used when the estimated intensity is below 30%, the rate is increased to 1.0 MPa / s in the 30%-70% range, and the rate is reduced to 0.2 MPa / s after exceeding 70%.

[0037] Acoustic emission signals are collected synchronously. The number of events with amplitudes exceeding 45 dB to 95 dB is counted to calculate the damage evolution b-value. When the rate of change of the b-value exceeds the set critical threshold, it is determined to be a precursor to critical damage.

[0038] A linear positive correlation model between the axial force-feed speed ratio and compressive strength, and a negative correlation model between the dimensionless working condition index time integral and strength were established to comprehensively calculate the actual strength of the core sample.

[0039] Preferably, step S3, which involves conducting an infrared spectral composition survey of non-severely defective areas, specifically includes the following steps:

[0040] For the 2.5–5 μm short-wave infrared channel, a quantum cascade laser with a center frequency of 120 terahertz was used as the light source to scan the characteristic peaks of silicate crystals at a resolution of 4 cm⁻¹. For the 8–12 μm long-wave infrared channel, the characteristic peaks of carbonate and chloride ions were scanned with a step size of 0.5 μm, and the single integration time was set to 50 ms. The fusion weights were dynamically allocated based on the dual-band signal-to-noise ratio, and the weight coefficients were set as the normalized results of the squared values ​​of the signal-to-noise ratios of each band.

[0041] By calibrating the influence of ambient temperature using a blackbody radiation source at a specified temperature, a wavelength drift model including the steady-state thermal expansion coefficient and transient response coefficient is established; the original spectrum is nonlinearly corrected, and Planck's blackbody radiation formula is used to eliminate ambient thermal radiation interference.

[0042] An improved convolutional recurrent neural network was used for analysis. The network input layer received a 256×256-dimensional dual-band spectral stitching matrix. The convolutional module used a dilated convolutional kernel with an expansion rate of 2 to extract multi-scale features, and the recurrent module used a bidirectional long short-term memory network to capture spectral sequence dependencies. The hidden layer dimension was set to 128. During model training, a weighted cross-entropy loss function was used, with weights inversely distributed according to the material aging level. ±2% spectral perturbation was added to generate adversarial examples to improve robustness.

[0043] When calculating the aging index, the output probability of the neural network is fused and assigned a weight of 60%, the carbonation index is calculated and assigned a weight of 30%, and the chloride ion corrosion index is calculated and assigned a weight of 10%. The weighted sum is used to calculate the final aging index. When the aging index exceeds 0.6, it is judged as severe aging, and a maintenance warning is generated.

[0044] Preferably, in step S4, the method for the dynamic evaluation model to conduct a comprehensive structural safety assessment is as follows:

[0045] An improved weighted Shannon entropy model was used to calculate the disorder of surface defect distribution, with the coupling effect of crack depth and area as the weighting factor. A fractal dimension correction term was introduced. When the fractal dimension exceeds 1.2, the total entropy value is increased by 30% of the fractal dimension deviation to enhance the sensitivity to complex crack networks.

[0046] To calculate the effective stress concentration factor, the reference volume is set to 1000 cubic centimeters, and the void volume is calculated by summing the equivalent sphere radii. When the crack length exceeds 30 centimeters, an additional correction term is added to the stress concentration factor, and the correction magnitude is proportional to the length exceeding the value. When continuous surface cracks are detected, the internal defect weighting factor is increased by 40%.

[0047] The uniformity of material property distribution is evaluated based on the spatial autocorrelation Moran's I index. When calculating the index, an exponential decay function is used to define the spatial weight matrix, and the decay coefficient is set to 0.5 m. When the Moran's I index exceeds 0.3, it is determined that the material properties are clustered. The dispersion coefficient is amplified according to the exponential difference, and the amplification factor is set to a maximum of 0.5.

[0048] The safety factor is dynamically synthesized: the surface entropy value, the corrected internal defect coefficient, and the dispersion coefficient are calculated using the inverse weighting formula with a weight ratio of 0.5:0.3:0.2 to obtain the comprehensive safety factor; when the safety factor is lower than 1.2, a red warning is generated and the high-risk area is marked in the three-dimensional model.

[0049] Preferably, in step S4, the IoT module is configured to optimize the evaluation process. The operation of the IoT module includes the following steps:

[0050] The detection data stream is decomposed into wavelet packets and dynamically quantized and compressed. The Daubechies 9 / 7 wavelet basis function is used to decompose the data into multiple sub-bands, and the quantization step size is dynamically adjusted according to the signal-to-noise ratio of each sub-band. Under normal conditions, a compression ratio of 10:1 is used, and when a high-risk defect is detected, it is switched to a 5:1 high-fidelity compression mode to prioritize the preservation of high-frequency detail features.

[0051] A multi-physics compensation matrix for temperature, humidity, and pressure is constructed to quantify the influence of environmental factors on the measured values. The sensitivity of strain measurements to temperature is set to 0.3 microstrain / degree Celsius, and the sensitivity of stress measurements to humidity is set to 0.05 MPa / percent relative humidity. Environmental data is collected in real time and substituted into the compensation matrix to correct the strain and stress detection results.

[0052] A time series prediction model is constructed based on a long short-term memory network. The input feature vector includes surface defect entropy, internal defect volume ratio, material dispersion coefficient, and environmental parameters. The hidden layer dimension of the network is set to 64, and the time step covers 24 consecutive hours of data. The output is the risk probability value for the next 30 days. The risk level is divided into three levels: low risk is below 0.3, medium risk is between 0.3 and 0.6, and high risk is above 0.6. Key parameters are identified through the Fisher information matrix, and an elasticity coefficient of 0.8 is set to limit the parameter update range, while retaining the characteristic distribution of historical data.

[0053] Preferably, the IoT module includes a visualization terminal interface unit, and the 3D visualization inspection report uses an augmented reality interface, including the following methods and steps:

[0054] An environmental point cloud map is constructed based on synchronous localization and mapping technology to achieve six-degree-of-freedom spatial alignment between the detection data coordinate system and the real scene, and the real-time update error of the pose matrix is ​​controlled to be less than 0.3 mm. The surface morphology layer is superimposed and displayed with a resolution of 0.5 mm / pixel, and the defect distribution layer is colored with red-orange gradient. The coloring intensity increases linearly with the crack depth, and the transparency drops to 0.2 when the depth exceeds 3 mm.

[0055] Based on the recognition of hand gestures using a 3D convolutional neural network, the system takes 16 consecutive depth image sequences as input, fits the spatial trajectory of the hand gestures, and outputs operation commands. It also performs voxel cutting on any specified plane to generate material property profiles, extracts the mechanical parameter distribution of the profiles using a Gaussian kernel interpolation algorithm to generate profile stress distribution maps, and sets the interpolation bandwidth to a fixed 5 mm to suppress noise interference.

[0056] The volume of the defect area to be repaired is calculated, and the Heaviside function is used to determine the void boundary with an integration step size of 1 mm. With the goal of maximizing stiffness and minimizing material usage, the volume of the reinforcement material does not exceed 30% of the original structure, and the optimization density threshold is set to 0.3. The topology optimization is performed to design the reinforcement structure. The Mises stress cloud diagram of the reinforced structure is displayed in real time by the visualization terminal interface unit. The reinforcement scheme is modified by gesture interaction, and the mechanical simulation results are dynamically updated with an optimization iteration cycle of less than 2 seconds.

[0057] The present invention has at least the following beneficial effects: (1) By integrating three-dimensional laser scanning technology and intelligent recognition algorithm, the automated and rapid detection of surface defects of the protective block is realized; compared with the traditional manual inspection method, this method can fully cover the structural surface, effectively identify micro-cracks and local peeling areas, and significantly improve the accuracy of defect location; at the same time, the multi-angle scanning and real-time data fusion mechanism overcomes the limitations of single-view detection, ensuring no omissions in the detection of complex geometric feature areas; through standardized data processing, the error of human judgment is greatly reduced, providing reliable basic data support for subsequent graded evaluation; (2) The adaptive scanning strategy designed for the complex curved surface features of the protective block, through By dynamically adjusting the scanning parameters, the point cloud acquisition quality in high curvature areas is effectively improved; by combining anti-interference algorithms and closed-loop control mechanisms, measurement deviations caused by factors such as wave vibration and temperature and humidity changes in the marine environment are significantly suppressed; the optimized point cloud registration process shortens the overall operation time while ensuring millimeter-level accuracy, making large-scale structural inspection feasible in engineering; this method is particularly suitable for the inspection of irregularly shaped protective blocks with complex surface textures and varied geometric shapes; (3) Multi-physics field coupling detection technology is adopted to give full play to the complementary advantages of ultrasonic flaw detection and impact echo methods; high-frequency ultrasonic scanning provides fine imaging of near-surface defects, while impact echo technology enhances the detection of deep defects. Capabilities; Through signal fusion and three-dimensional reconstruction algorithms, the spatial distribution characteristics of defects are fully revealed, avoiding the field of view limitations of traditional single-modal detection; This scheme shows good adaptability in the identification of typical defects such as internal voids and crack networks in concrete, significantly reducing the risk of missed detection and providing multi-dimensional data basis for structural integrity assessment; (4) Based on the layered detection strategy of material performance gradient, the synergistic assessment of macroscopic mechanical properties and microscopic composition changes is realized; Non-contact spectral analysis technology breaks through the spatial limitations of traditional sampling and detection, and can quickly survey the carbonization depth and erosion status of a large area; Combined with local high-precision hardness testing, the performance mutation characteristics of the defect interface transition zone are accurately captured; This method in While reducing structural damage, it significantly improves the spatial resolution of material performance data, providing a refined data foundation for durability degradation analysis; (5) The constructed multi-source data dynamic evaluation model objectively reflects the overall and regional characteristics of structural safety status through quantitative index fusion and weight adaptive mechanism; The augmented reality visualization system transforms abstract detection data into an interactive model in three-dimensional space, intuitively presenting the distribution pattern and development trend of defects; Combined with the repair scheme generated by the predictive maintenance algorithm, it effectively guides the precise allocation of maintenance resources; This technology system greatly shortens the response cycle from detection to decision-making, improves the pertinence and economy of maintenance work, and provides intelligent decision support for engineering management. Attached Figure Description

[0058] Figure 1 is a flowchart illustrating the principle of the wharf revetment block quality inspection method provided by the present invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0060] As shown in Figure 1, the method for quality inspection of wharf revetment blocks includes the following steps:

[0061] S1: A mobile 3D laser scanning device is used to scan the protective block to obtain laser point cloud data. A 3D surface model is constructed through a real-time point cloud registration algorithm. Surface geometric features, including surface crack depth and peeling area, are extracted based on a deep convolutional neural network. Areas where the surface crack depth exceeds a preset depth threshold or the peeling area is greater than a preset area threshold are marked as high-risk areas, and a heat map of high-risk areas is generated.

[0062] S2: Based on the thermal map of high-risk areas, a frequency-adjustable dual-mode flaw detection device is used to detect voids and cracks inside the protective block: an ultrasonic probe is used to perform intensified scanning in high-risk areas; multiple impact points are arranged in the projection area of ​​the surface crack extension direction according to the stress diffusion path, and an array-type impact echo device is used to collect vibration signals at the impact points;

[0063] A three-dimensional defect probability distribution map is established by fusing ultrasonic data and vibration signal data through a multi-sensor data fusion algorithm, and areas where the diameter of internal cavities or the length of cracks are greater than the corresponding upper limit threshold are marked as severe defect areas.

[0064] S3: Drill multiple core samples within a predetermined distance around the severely defective area and conduct compressive strength tests and acoustic emission damage detection; conduct infrared spectral composition surveys in non-severely defective areas to detect concrete carbonation and chloride ion erosion index; in the defect-matrix interface transition zone, set up micro-area hardness test points according to the material property gradient and conduct hardness tests.

[0065] S4: Establish a dynamic evaluation model that includes surface defect distribution entropy, internal defect volume ratio, and material property dispersion coefficient; calculate the comprehensive safety factor S through a multi-objective optimization algorithm, and compare it with the upper limit Su and lower limit Sd of the safety threshold to generate a three-dimensional visualization inspection report.

[0066] The scanning accuracy of the 3D laser scanning device can be set from 0.5 mm to 2 mm, the crack depth threshold can be set to 3 mm or 5 mm, and the spalling area threshold can be set to 50 square centimeters or 80 square centimeters. It employs a FARO Focus series laser scanner equipped with a UR10e model six-degree-of-freedom robotic arm, with a scanning probe mounted at the end of the arm. The scanning probe is installed within a range of 0.3 meters to 1.5 meters from the surface of the protective block. The robotic arm moves and scans along a preset path, stitching together point cloud data in real time. Defect areas are automatically labeled using a deep convolutional neural network (such as ResNet-50), which can greatly improve the accuracy of surface crack identification and effectively shorten the time per scan.

[0067] The ultrasonic probe frequency can be selected from 1.5 MHz, 2 MHz, or 2.5 MHz, and the scanning grid spacing can be set to 5 cm × 5 cm or 10 cm × 10 cm. The impact energy for impact echo detection can be set to 0.5 joules, 2 joules, or 5 joules, and the vibration signal acquisition time can be set to 40 milliseconds, 50 milliseconds, or 60 milliseconds. Olympus Epoch series ultrasonic flaw detectors and Proceq Pundit PL-200 impact echo analyzers can be used. The ultrasonic probe can be installed at the center of the surface of the inspection area, and the impact hammers can be arranged at intervals of 10 cm to 30 cm along the crack extension direction. After high-frequency ultrasonic scanning, switching to impact echo mode and fusing the two types of data to generate a three-dimensional defect map can effectively reduce errors in locating cavities and measuring crack length within the block.

[0068] The core sample drilling diameter can be selected as 50 mm or 75 mm, and the compressive strength loading rate can be set to 0.3 MPa / s, 0.5 MPa / s, or 1.0 MPa / s. The hardness test point spacing can be set to 2 mm, 5 mm, or 10 mm. A Hilti DD 120 core drill, paired with an Instron 5982 universal testing machine, is used for strength testing. The core drill can be installed 10 cm to 20 cm from the defect edge, and the infrared spectral probe can cover the entire surface of the non-defect area. After drilling the core sample, the pressure is applied in stages until failure, and the peak frequency of the acoustic emission signal is recorded simultaneously, which can reduce the repeatability error of the compressive strength test and control the chloride ion content detection accuracy to 0.01%.

[0069] The safety factor threshold can be set to 1.2, 1.5, or 1.8, and the weight adjustment range can be selected as 20%, 40%, or 60%. The evaluation algorithm can be run on a Dell Precision workstation, with calculations accelerated by an NVIDIA Quadro GPU. The data processing server can be deployed in the dock control center and connected to the inspection equipment via a 5G network. After importing surface defect distribution, internal defect volume, and material property data, the system automatically calculates the safety level and generates a dual-format report (PDF / 3D model). It provides a very short early warning response time for high-risk areas, and the data analysis results effectively improve the accuracy of maintenance plan generation.

[0070] In another technical solution, in step S1, the robotic arm of the mobile 3D laser scanning device integrates a dual-mode lidar and an inertial measurement unit, and performs the following operations:

[0071] Based on continuous three-point cloud slice data, a differential geometry algorithm is used to calculate the local curvature change rate of the surface of the protective block in real time. The curvature change rate is quantified by the ratio of the Gaussian curvature difference between adjacent measurement points to the arc length distance. When the local minimum curvature radius is less than 2 meters, the surface roughness index exceeds the standard defined threshold of 0.5 mm, and environmental vibration in the range of 0.1-5 Hz is detected, the high frame rate lidar mode is activated, and the scanning frame rate is increased to no less than 200 Hz.

[0072] An adaptive Kalman filter is used to eliminate low-frequency vibration noise caused by waves. By fusing the state equations of the robot arm's pitch angle, angular velocity, and linear acceleration with the observation data of the inertial measurement unit, the process noise and observation noise covariance matrix is ​​updated in real time to suppress wave interference.

[0073] In high frame rate LiDAR mode, the movement speed of the robotic arm end effector is dynamically adjusted according to the local minimum radius of curvature. When the minimum radius of curvature is less than 1 meter, a low-speed scan of 0.2 m / s is used, and when the radius of curvature is in the range of 1-2 meters, a medium-speed scan of 0.5 m / s is used. At the same time, the laser line density is increased to 800 lines per meter, and the point cloud overlap rate of adjacent scanning paths is ensured to be no less than 40%.

[0074] The robotic arm can have 6 or 7 axes of joints, and its end effector load capacity can be set to 5 kg or 10 kg. The radius of curvature threshold can be set to 2 m, 1.5 m, or 1 m. A Universal Robots UR10e six-DOF robotic arm can be used, and the arm body material can be aluminum alloy or carbon fiber composite. The robotic arm base can be mounted on the top center of the mobile scanning platform and rigidly connected to the scanning device. The robotic arm moves along the normal direction of the surface of the protective block through a preset program, and the scanning speed can be set to 0.2 m / s or 0.5 m / s. In the efficient scanning mode, it can maintain a high coverage of complex curved surfaces and reduce the robotic arm positioning error to 0.1 mm.

[0075] The curvature change rate calculation time interval can be set to 0.1 seconds or 0.2 seconds, and the surface roughness threshold can be set to 0.3 mm, 0.5 mm, or 0.8 mm. The LiDAR can be a FARO Focus S 350 model, and the scanning frame rate can be switched to 100 Hz, 200 Hz, or 300 Hz. The LiDAR can be installed 30 cm in front of the flange at the end of the robotic arm, and the scanning angle can be adjusted to ±45 degrees. In practice, the curvature difference between three adjacent point cloud slices is calculated in real time. When the curvature radius is less than the threshold and the roughness exceeds the standard, the scanning frame rate is automatically increased to above 200 Hz, which can increase the point cloud density in high curvature areas to 1200 points / square meter, improving data acquisition efficiency by 40%.

[0076] The wave vibration filtering bandwidth can be set to 0.1-5 Hz or 0.5-3 Hz, and the scan line density can be set to 500 lines / m, 800 lines / m, or 1000 lines / m. The inertial measurement unit (IMU) can be an Xsens MTi-670 model, and the accelerometer range can be set to ±8g or ±16g. The IMU can be installed 10 cm behind the lidar and connected to the lidar data interface via an RS485 bus. A Kalman filter is used to fuse the robotic arm's attitude data and the IMU vibration signal, correcting the scan path coordinate offset in real time. By filtering and denoising the data, the point cloud distortion rate caused by wave interference is reduced to below 5%, and the registration error in overlapping areas is less than 0.8 mm.

[0077] In step S1, the real-time point cloud registration algorithm specifically includes the following steps:

[0078] Coarse registration is performed: an improved iterative nearest point algorithm is adopted, with the weighted sum of geometric distance and curvature difference as the optimization objective. The curvature difference is quantified by the mixed difference between the Gaussian curvature and the average curvature of the corresponding point, and the Gaussian curvature weight factor is set to 0.5. When screening matching point pairs, only point pairs with a principal curvature radius ratio greater than 0.7 and a cosine value of the angle between the normal vectors greater than 0.9 are retained, and mismatches with inconsistent surface geometric features are excluded.

[0079] The curvature consistency weighting factor can be set to 0.3, 0.5, or 0.7; the principal curvature radius ratio threshold can be set to 0.6, 0.7, or 0.8; and the cosine value threshold of the normal vector angle can be set to 0.85, 0.9, or 0.95. An Intel Arria 10 series FPGA chip can be used to implement the coarse registration algorithm, and an existing point cloud processing library such as PCL (Point Cloud Library) can be selected for the matching point pair filtering module. The FPGA module can be installed in the PCIe slot of the scanning device control cabinet and connected to the main control computer via Gigabit Ethernet. The original point cloud is sliced ​​at 5 mm intervals, the curvature difference between adjacent slices is calculated, and mismatched points that do not meet the curvature ratio and normal vector conditions are removed. Under a low mismatch rate, coarse registration has a very fast matching speed.

[0080] Fine-tuning is performed: A Gauss-Newton iterative algorithm accelerated by a field-programmable gate array (FPGA) is used to compute the Jacobian matrix containing rotational and translational components in parallel. The optimal transformation parameters are then solved using damped least squares, with a damping factor set to 0.0001 to prevent matrix singularities. Combined with 3D feature descriptor matching, the local surface is divided into 32 azimuth intervals, and a height distribution histogram is generated. Chi-square distance is used to calculate feature similarity, and regions with matching scores below a threshold are marked as low-confidence regions. The Gauss-Newton iterative damping factor can be set to 0.0001, 0.001, or 0.01. The feature descriptor azimuth intervals can be divided into 16, 32, or 64 partitions, and the histogram resolution can be set to 0.3 mm, 0.5 mm, or 1.0 mm. A Xilinx Kintex UltraScale FPGA development board can be used to accelerate matrix operations, and the feature matching unit can be implemented using the Open3D open-source framework. The FPGA acceleration card can be integrated into the data processing workstation and communicates with the CPU via a PCIe 3.0 interface. Based on the coarse registration results, the Jacobian matrix is ​​calculated in parallel and the optimal transformation parameters are solved iteratively. At the same time, the SHOT feature descriptor is matched, and the fine registration accuracy can reach 0.5 mm, while taking very little processing time per frame.

[0081] Dynamic evaluation and replanning are performed: the registration error of the local area is calculated in real time. When the root mean square error of the area exceeds 0.8 mm, it is determined to be a registration failure area; the scanning path replanning is triggered, and a ±15 degree deflection angle is added to the original scanning path to increase the point cloud line density to 1200 lines per meter and the overlap rate to 60%. The low confidence area is also re-scanned.

[0082] The local registration error threshold can be set to 0.5 mm, 0.8 mm, or 1.0 mm; the scanning path deflection angle can be set to ±10 degrees, ±15 degrees, or ±20 degrees; and the supplementary scanning line density can be set to 1000 lines / meter, 1200 lines / meter, or 1500 lines / meter. A Velodyne VLP-16 LiDAR can be used for supplementary scanning, and the path planning algorithm can use the ROS (Robot Operating System) navigation package. The supplementary scanning device can be installed 50 cm to the side of the original scanning device and synchronized with the main control system via a CAN bus. The registration error of each area is calculated in real time. When the error exceeds the threshold, a spiral progressive supplementary scanning path is generated, automatically adjusting the scanning angle and line density. This effectively reduces the time spent on repeated scanning of low-confidence areas and ensures extremely high integrity of the overall point cloud.

[0083] In another technical solution, in step S2, the frequency-tunable dual-mode flaw detection device includes an ultrasonic probe and an array-type impact echo device, wherein the ultrasonic probe adopts the following layered waveguide structure and dynamic control strategy:

[0084] The waveguide layer of the ultrasonic probe consists of a piezoelectric composite substrate and five gradient impedance matching layers. The impedance value of each layer gradually changes exponentially from the substrate to the couplant interface. The total thickness of the matching layers is 2 mm to suppress sound energy scattering caused by surface roughness. When the arithmetic mean deviation of the surface roughness exceeds 12.5 micrometers, a wideband sweep detection mode is activated, continuously sweeping the frequency at a rate of 50 kHz per second within the transmission frequency range of 1.5 MHz to 2.5 MHz, and dynamically adjusting the transmission pulse width according to the sweep range. The total number of waveguide layers can be set to 3, 5, or 7 layers, and the sweep range can be set to 1 to 2.5 MHz or 1.5 to 3 MHz. The ultrasonic probe can be the Olympus V358-SU model, and the piezoelectric composite material can be PZT-5H or PMN-PT. The probe waveguide layer can be installed between the transducer wafer and the protective film, with a total thickness controlled within the range of 2 to 3 mm. When the surface roughness exceeds 12.5 micrometers, the 1.5 to 2.5 MHz sweep frequency mode is automatically activated, and ultrasonic waves with a pulse width of 3 to 5 microseconds are emitted at each frequency point. The signal-to-noise ratio of the rough surface echo can be increased to more than 30 dB, and the defect detection rate exceeds 90%.

[0085] The effective focusing depth is calculated based on the material's longitudinal wave velocity and the current transmission frequency. When the focusing depth fluctuation exceeds ±10% of the nominal value, a calibration mode is triggered: a 1 MHz reference pulse is emitted, and the actual material wave velocity is corrected in reverse by the echo time difference, eliminating the sound velocity error caused by concrete aging. The wave velocity correction coefficient can be set to 2.3, 2.5, or 2.8, and the roughness correction index can be selected as 0.25, 0.3, or 0.35. The focusing calibration pulse frequency can be set to 1 MHz or 0.5 MHz, and the transmission pulse width can be set to 2 microseconds or 3 microseconds. The calibration signal transmission module can be integrated inside the probe handle and connected to the main control unit via a coaxial cable. A calibration pulse is automatically emitted every 10 minutes, correcting the wave velocity calculation value based on the echo time difference and updating the material parameters in the depth formula. This results in a very small focusing depth control error, reducing the detection deviation caused by concrete aging to within 3%.

[0086] A nonlinear model of the relationship between contact pressure and surface roughness is established. The probe height is adjusted in real time using a piezoelectric actuator to stabilize the coupling agent pressure within the range of 0.2 to 0.5 MPa. A proportional-integral-derivative (PID) controller is used to achieve dynamic balance, with a proportional gain of 0.8, an integral time of 0.1 seconds, and a derivative time of 0.05 seconds. The pressure sensor sampling frequency is 1 kHz. The contact pressure range can be set to 0.15 to 0.5 MPa, and the PID control parameters can be set to a proportional gain of 0.6, an integral time of 0.2 seconds, and a derivative time of 0.1 seconds. The piezoelectric actuator can be the PI P-623.1CD model, and the pressure sensor can be the Honeywell FMA series. The actuator can be mounted on the top of the probe housing, parallel to the pressure sensor at a 5 cm interval. Closed-loop feedback is used to adjust the probe's downward pressure distance in real time, maintaining the coupling agent film thickness at 0.1 to 0.3 mm and controlling the pressure fluctuation within ±0.05 MPa to reduce signal loss due to poor contact.

[0087] In step S2, the array-type impact echo device locates internal defects using the following method:

[0088] The impact energy is controlled in five levels—0.5J, 1J, 2J, 3.5J, and 5J—by adjusting the impact speed of the array-type impact echo device. The impact frequency is dynamically selected based on the Vickers hardness of the surface: 10Hz high-frequency impact is used for hardness below 300 MPa, 5Hz medium-frequency impact for hardness between 300-500 MPa, and 2Hz low-frequency impact for hardness exceeding 500 MPa. The impact hammer weight can be selected as 30g, 50g, or 80g, and the impact speed can be divided into three levels: 4.47 m / s, 6.32 m / s, and 8.94 m / s, corresponding to 0.5 to 5 Joules of energy. The impact device can be a Proceq Pundit PL-300 model, and the hammer head material can be tungsten carbide or zirconia ceramic. The impact hammer can be installed at a height of 2 to 5 cm above the detection surface and is driven vertically by an electromagnetic coil. The impact level is automatically selected based on the surface hardness value. For hardness below 300 MPa, a 10 Hz high-frequency mode is used, and for hardness above 300 MPa, a 2 Hz low-frequency mode is used to effectively suppress the signal overload rate and ensure a high effective data acquisition rate.

[0089] The surface wave velocity is calculated based on the material's elastic modulus, density, and Poisson's ratio. The surface wave wavelength is calculated based on the central analysis frequency and the surface wave velocity. Piezoelectric sensors are arranged at intervals one-eighth of the surface wave wavelength. Several piezoelectric sensors are arranged in a spiral-grid hybrid layout to form a sensor array covering three-dimensional space.

[0090] The vibration signal acquired by the echo device is fused with short-time Fourier transform and wavelet packet decomposition. The short-time Fourier transform uses a Gaussian-modulated cosine window function, with the window width dynamically adjusted according to the impact frequency. The wavelet packet decomposition uses a 10th-order Daubechies wavelet basis, based on the sampling frequency f. sample With center frequency f central Set the number of decomposition layers to

[0091] A dual-channel analysis method with high and low frequencies is adopted to calculate the signal energy integral value of the low-frequency channel (0.5-10 kHz) and the kurtosis index of the high-frequency channel (10-50 kHz).

[0092] Based on the data from the sensor array, a time-reversal algorithm is used to reconstruct the three-dimensional wave field, with the observation time window length set to 50 milliseconds. When the energy integral value of the low-frequency channel signal exceeds 0.8 times the reference value of healthy concrete, the kurtosis index value of the high-frequency channel exceeds the threshold, and the diameter of the three-dimensional wave field focusing area is greater than 10 centimeters, the excess area is judged as a valid internal defect.

[0093] Surface wave wavelength calculation formula: Surface wave velocity Where v is Poisson's ratio and the center frequency f centralThe kHz frequency can be selected, the elastic modulus E can be taken as 30 to 50 GPa, and the material density ρ can be set as 2300 to 2500 kg / m³. 3 The piezoelectric sensor can be a PCB Piezotronics 352C03 model, and the array substrate material can be FR-4 epoxy board or aluminum alloy. The sensors can be arranged at 10 to 15 cm intervals along the crack direction, symmetrically offset at an angle of 20 to 30 degrees on both sides. Surface waves are excited by a 25 kHz center frequency, and vibration signals are acquired at 1 / 8 wavelength intervals. The accuracy of crack extension direction detection exceeds 90%, and the blind zone is reduced to within 5 cm.

[0094] The short-time Fourier transform window width can be set to 5 milliseconds, 10 milliseconds, or 20 milliseconds, and the wavelet packet decomposition level can be selected as 3, 5, or 7 layers. The signal processor can be a NIPXIe-8840 controller, equipped with a 14-bit high-speed acquisition card. The data analysis module can be deployed in an industrial control computer and directly connected to the acquisition card via a PCIe interface. By fusing the energy integral value in the 0.5 to 10 kHz low-frequency band with the kurtosis index in the 10 to 50 kHz high-frequency band, a defect probability distribution heatmap is generated. The detection error of the internal cavity diameter is controlled to less than 1 cm, and the detection time can be shortened from 45 minutes to 20 minutes.

[0095] In another technical solution, step S3, core sampling and testing, specifically includes the following steps:

[0096] Based on the fusion data of laser displacement sensor and inertial measurement unit, the drill bit axis offset angle is calculated in real time: the lateral offset is measured by laser sensor, and mechanical vibration error is compensated by gyroscope angular velocity data, with the compensation coefficient set to 0.05 seconds / radian; the stroke of four hydraulic cylinders is adjusted by fuzzy proportional-integral-derivative controller, with the goal of minimizing the weighted sum of offset angle and axial force change rate, and the maximum deflection angle is limited to ±15 degrees, triggering an alarm when the limit is exceeded;

[0097] The dimensionless working condition index is calculated in real time and defined as the product of axial drilling force and drill radius divided by the product of feed rate and hydraulic pressure. When the index exceeds the set threshold, three levels of protection are implemented: the retraction speed is adjusted to twice the current feed rate plus 50 mm / s; the drill speed is reduced to 300 rpm; and coolant is sprayed at a flow rate of 0.2 times the axial force.

[0098] The following reinforcement methods were used in the core sample compressive strength test:

[0099] The loading rate is controlled in three stages: a slow loading rate of 0.5 MPa / s is used when the estimated intensity is below 30%, the rate is increased to 1.0 MPa / s in the 30%-70% range, and the rate is reduced to 0.2 MPa / s after exceeding 70%.

[0100] Acoustic emission signals are collected synchronously. The number of events with amplitudes exceeding 45 dB to 95 dB is counted to calculate the damage evolution b-value. When the rate of change of the b-value exceeds the set critical threshold, it is determined to be a precursor to critical damage.

[0101] A linear positive correlation model between the axial force-feed speed ratio and compressive strength, and a negative correlation model between the dimensionless working condition index time integral and strength were established to comprehensively calculate the actual strength of the core sample.

[0102] The hydraulic cylinder stroke adjustment accuracy of the core sampling equipment can be set to 0.1 mm or 0.2 mm, and the drilling torque to feed rate ratio threshold can be set to 120 N·s / mm or 150 N·s / mm. The hydraulic servo valve can be a Bosch Rexroth 4WRPE series, and the servo motor can be a Yaskawa SGM7G model. The hydraulic cylinders can be installed symmetrically on both sides of the drill column and connected to the spindle via rigid couplings. A laser displacement sensor monitors the drill bit offset in real time. When the offset angle exceeds a preset threshold, the hydraulic system automatically adjusts the pressure difference of the four sets of cylinders to improve drilling verticality accuracy and reduce the risk of core sample breakage.

[0103] The loading rate for compressive strength testing can be set to three levels: 0.3 MPa / s, 0.5 MPa / s, or 0.8 MPa / s. The acoustic emission signal acquisition frequency can be set to 100 kHz or 200 kHz. An Instron 3367 universal testing machine can be used, and a PAC Micro80 acoustic emission sensor can be selected. The core sample fixture can be installed between the upper and lower pressure plates of the testing machine, with the sensor attached to the center of the core sample's side. The pressure is applied in stages until the specimen fails, and the cumulative number and amplitude distribution of acoustic emission events are recorded simultaneously to accurately identify the material's critical damage point and prevent sudden fracture.

[0104] The linear correlation coefficient of the data fusion model can be set to 0.1 to 0.2, and the integration time range can be set to 30 seconds or 60 seconds. The data acquisition card can be the NI USB-6366, and the control software can be developed based on the LabVIEW platform. The processing module can be integrated into an industrial computer and communicate with the drilling rig controller via Ethernet. It calculates the correlation between drilling parameters and compressive strength in real time, generates material property mapping curves, establishes a quantitative relationship between the drilling process and material strength, and supports in-situ quality assessment.

[0105] In another embodiment, step S3, which involves conducting an infrared spectral composition survey of non-severely defective areas, specifically includes the following steps:

[0106] For the 2.5–5 μm short-wave infrared channel, a quantum cascade laser with a center frequency of 120 terahertz was used as the light source to scan the characteristic peaks of silicate crystals at a resolution of 4 cm⁻¹. For the 8–12 μm long-wave infrared channel, the characteristic peaks of carbonate and chloride ions were scanned with a step size of 0.5 μm, and the single integration time was set to 50 ms. The fusion weights were dynamically allocated based on the dual-band signal-to-noise ratio, and the weight coefficients were set as the normalized results of the squared values ​​of the signal-to-noise ratios of each band.

[0107] By calibrating the influence of ambient temperature using a blackbody radiation source at a specified temperature, a wavelength drift model including the steady-state thermal expansion coefficient and transient response coefficient is established; the original spectrum is nonlinearly corrected, and Planck's blackbody radiation formula is used to eliminate ambient thermal radiation interference.

[0108] An improved convolutional recurrent neural network was used for analysis. The network input layer received a 256×256-dimensional dual-band spectral stitching matrix. The convolutional module used a dilated convolutional kernel with an expansion rate of 2 to extract multi-scale features, and the recurrent module used a bidirectional long short-term memory network to capture spectral sequence dependencies. The hidden layer dimension was set to 128. During model training, a weighted cross-entropy loss function was used, with weights inversely distributed according to the material aging level. ±2% spectral perturbation was added to generate adversarial examples to improve robustness.

[0109] When calculating the aging index, the output probability of the neural network is fused and assigned a weight of 60%, the carbonation index is calculated and assigned a weight of 30%, and the chloride ion corrosion index is calculated and assigned a weight of 10%. The weighted sum is used to calculate the final aging index. When the aging index exceeds 0.6, it is judged as severe aging, and a maintenance warning is generated.

[0110] The spectral resolution of the shortwave infrared channel can be set to 4 cm. -1 Or 8 centimeters -1 The long-wavelength infrared integration time can be set to 40 milliseconds or 60 milliseconds. The quantum cascade laser can be from Block Engineering's LaserTune series, and the infrared detector can be the FLIR A67 model. The laser emitter can be mounted 20 cm above the scanning support, and the detector should be 50 cm away from the sample surface. Dual-band alternating scanning separates different wavelength signals using an optical beam splitter, simultaneously capturing the characteristic absorption peaks of hydration and erosion products.

[0111] The temperature of the blackbody radiation source for temperature drift compensation can be set to 30°C or 40°C, and the coefficient of thermal expansion of the wavelength correction model can be set to 0.015 nm / °C or 0.025 nm / °C. A Landcal R1500TI blackbody can be used as the reference blackbody, and an Omega CNi16 series temperature control module can be used. The blackbody source can be placed 10 cm beside the sample under test and coupled to the spectrometer inlet via optical fiber. The blackbody reference spectrum is acquired every 5 minutes to dynamically correct wavelength shifts caused by ambient temperature and suppress spectral distortion caused by temperature fluctuations.

[0112] The number of hidden layer nodes in the convolutional recurrent neural network can be set to 64 or 128, and the number of samples in the training dataset can be set to 1500 or 2000. An NVIDIA Jetson AGX Xavier processor can be used, and TensorFlow Lite can be used as the framework. The computing unit can be embedded inside the spectrometer host and connected to the main control board via a PCIe interface. The dual-band spectral data is stitched together into a two-dimensional matrix and input into the network, outputting material aging probability values ​​to automatically identify early carbonization and chloride ion corrosion signs.

[0113] In another technical solution, the method for performing a comprehensive structural safety assessment using the dynamic evaluation model in step S4 is as follows:

[0114] An improved weighted Shannon entropy model was used to calculate the disorder of surface defect distribution, with the coupling effect of crack depth and area as the weighting factor. A fractal dimension correction term was introduced. When the fractal dimension exceeds 1.2, the total entropy value is increased by 30% of the fractal dimension deviation to enhance the sensitivity to complex crack networks.

[0115] To calculate the effective stress concentration factor, the reference volume is set to 1000 cubic centimeters, and the void volume is calculated by summing the equivalent sphere radii. When the crack length exceeds 30 centimeters, an additional correction term is added to the stress concentration factor, and the correction magnitude is proportional to the length exceeding the value. When continuous surface cracks are detected, the internal defect weighting factor is increased by 40%.

[0116] The uniformity of material property distribution is evaluated based on the spatial autocorrelation Moran's I index. When calculating the index, an exponential decay function is used to define the spatial weight matrix, and the decay coefficient is set to 0.5 m. When the Moran's I index exceeds 0.3, it is determined that the material properties are clustered. The dispersion coefficient is amplified according to the exponential difference, and the amplification factor is set to a maximum of 0.5.

[0117] The safety factor is dynamically synthesized: the surface entropy value, the corrected internal defect coefficient, and the dispersion coefficient are calculated using the inverse weighting formula with a weight ratio of 0.5:0.3:0.2 to obtain the comprehensive safety factor; when the safety factor is lower than 1.2, a red warning is generated and the high-risk area is marked in the three-dimensional model.

[0118] The fractal dimension correction coefficient for the surface defect distribution entropy can be set to 0.2 or 0.3, and the reference volume in the effective stress concentration factor calculation formula can be set to 800 cubic centimeters or 1200 cubic centimeters. An HP Z8G4 workstation can be used as the computing server, and the optimization algorithm can be implemented based on the MATLAB Global Optimization Toolbox. The data processing terminal can be deployed in the on-site control room and connected to the scanning equipment via fiber optic cable. After importing the 3D defect model, the entropy value and stress concentration factor of each region are automatically calculated, quantifying the impact of defects on the overall structural stability.

[0119] The weight matrix attenuation coefficient for spatial autocorrelation analysis can be set to 0.3 meters or 0.5 meters, and the dispersion coefficient correction range can be set to 0.4 or 0.6. ArcGIS Pro can be used as the geographic information system software, and the statistical analysis module can be developed based on the Python SciPy library. The spatial analysis unit can run on a cloud server and interact with the local system via an API interface. The workflow involves spatial interpolation of the hardness test point data, calculation of Moran's I index, and correction of the dispersion coefficient. The technical effect is the identification of areas of anomalous material properties.

[0120] The weighting ratio in the safety factor synthesis formula can be set to 5:3:2 or 6:2:2. When set to 5:3:2, the corresponding inverse weighting formula for calculating the safety factor S is: Where Ht is the entropy comprehensive index, α is the defect volume weighting factor, and K′ t C′ is the effective stress concentration factor. v The decision threshold can be set to 1.1 or 1.3, representing the coefficient of variation. Visualization software can include ParaView or CloudCompare, and the report generation module can be based on JasperReports. The display terminal can be configured with a touchscreen interface and installed on the monitoring console of the dock dispatch center. The comprehensive evaluation results automatically generate a color-coded 3D model and text report, providing tiered early warnings and maintenance priority suggestions.

[0121] In another technical solution, in step S4, the IoT module is configured to optimize the evaluation process. The operation of the IoT module includes the following steps:

[0122] The detection data stream is decomposed into wavelet packets and dynamically quantized and compressed. The Daubechies 9 / 7 wavelet basis function is used to decompose the data into multiple sub-bands, and the quantization step size is dynamically adjusted according to the signal-to-noise ratio of each sub-band. Under normal conditions, a compression ratio of 10:1 is used, and when a high-risk defect is detected, it is switched to a 5:1 high-fidelity compression mode to prioritize the preservation of high-frequency detail features.

[0123] A multi-physics compensation matrix for temperature, humidity, and pressure is constructed to quantify the influence of environmental factors on the measured values. The sensitivity of strain measurements to temperature is set to 0.3 microstrain / degree Celsius, and the sensitivity of stress measurements to humidity is set to 0.05 MPa / percent relative humidity. Environmental data is collected in real time and substituted into the compensation matrix to correct the strain and stress detection results.

[0124] A time series prediction model is constructed based on a long short-term memory network. The input feature vector includes surface defect entropy, internal defect volume ratio, material dispersion coefficient, and environmental parameters. The hidden layer dimension of the network is set to 64, and the time step covers 24 consecutive hours of data. The output is the risk probability value for the next 30 days. The risk level is divided into three levels: low risk is below 0.3, medium risk is between 0.3 and 0.6, and high risk is above 0.6. Key parameters are identified through the Fisher information matrix, and an elasticity coefficient of 0.8 is set to limit the parameter update range, while retaining the characteristic distribution of historical data.

[0125] The data acquisition frequency of the IoT module can be set to once per minute or once every 5 minutes, the ambient temperature compensation range can be set to -10 to 50 degrees Celsius, and the humidity compensation range can be set to 20% to 95% RH. Huawei Atlas 500 smart stations can be used as edge computing devices, and Quectel EC25 series 4G modules can be used for wireless transmission. Edge nodes can be installed in the dock's power distribution box and connected to the inspection equipment via an RS485 bus. Surface defect coordinates, material hardness data, and environmental parameters are encapsulated in JSON format, encrypted with AES, and uploaded to the Alibaba Cloud IoT platform. Through edge-cloud collaborative processing, remote monitoring and long-term storage of inspection data are achieved, reducing local storage pressure and improving real-time early warning capabilities for abnormal states.

[0126] The time step of the LSTM network can be set to 24 or 48 hours, the number of hidden layer neurons can be set to 64 or 128, and the risk level threshold can be set to 0.3 or 0.5. A Dell PowerEdge R750xa with an NVIDIA A100 GPU accelerator card can be used as the training server. The predictive model can be deployed on a cloud-based Kubernetes cluster, providing query services via a REST API. Risk probability calculations for the next 30 days are performed every 6 hours, and the results are overlaid on the 3D model as a heatmap. Through time-series data analysis, structural performance degradation trends can be predicted, reducing the incidence of sudden failures and improving the operability of preventative maintenance plans.

[0127] The elasticity coefficient of the online learning algorithm can be set to 0.5 or 1.0, and the historical data retention period can be set to 1 year or 2 years. TensorFlow Federated can be used as the incremental learning framework, and Inspur AS5500G5 can be used as the storage server. The data warehouse can be deployed in an OSS storage bucket of a cloud service provider, interconnected with the training cluster via a dedicated line. Samples with a confidence level higher than 90% in newly added data are automatically selected weekly, and model weights are updated while key parameters are retained for protection. Through a continuous learning mechanism, the model achieves adaptive optimization to new environmental conditions, improving the compatibility of detection schemes across different ports.

[0128] The IoT module includes a visualization terminal interface unit. The 3D visualization inspection report uses an augmented reality interface and includes the following methods and steps:

[0129] An environmental point cloud map is constructed based on synchronous localization and mapping technology to achieve six-degree-of-freedom spatial alignment between the detection data coordinate system and the real scene, and the real-time update error of the pose matrix is ​​controlled to be less than 0.3 mm. The surface morphology layer is superimposed and displayed with a resolution of 0.5 mm / pixel, and the defect distribution layer is colored with red-orange gradient. The coloring intensity increases linearly with the crack depth, and the transparency drops to 0.2 when the depth exceeds 3 mm.

[0130] Based on the recognition of hand gestures using a 3D convolutional neural network, the system takes 16 consecutive depth image sequences as input, fits the spatial trajectory of the hand gestures, and outputs operation commands. It also performs voxel cutting on any specified plane to generate material property profiles, extracts the mechanical parameter distribution of the profiles using a Gaussian kernel interpolation algorithm to generate profile stress distribution maps, and sets the interpolation bandwidth to a fixed 5 mm to suppress noise interference.

[0131] The volume of the defect area to be repaired is calculated, and the Heaviside function is used to determine the void boundary with an integration step size of 1 mm. With the goal of maximizing stiffness and minimizing material usage, the volume of the reinforcement material does not exceed 30% of the original structure, and the optimization density threshold is set to 0.3. The topology optimization is performed to design the reinforcement structure. The Mises stress cloud diagram of the reinforced structure is displayed in real time by the visualization terminal interface unit. The reinforcement scheme is modified by gesture interaction, and the mechanical simulation results are dynamically updated with an optimization iteration cycle of less than 2 seconds.

[0132] The layered transparency of the augmented reality display can be set to 0.5 or 0.7, the profile interpolation bandwidth can be set to 3 mm or 8 mm, and the gesture recognition latency can be controlled within 50 ms or 100 ms. The AR glasses can be Microsoft HoloLens 2, and the gesture recognition sensor can be an Intel RealSense D455 depth camera. The display terminal can be installed on the front of a safety helmet and communicate with the data processing terminal via Wi-Fi 6. The 3D defect model is spatially aligned with the real-world video stream, and the operator switches display layers using pinch gestures. Through virtual-real fusion visualization, a three-dimensional presentation of hidden defects is achieved, improving on-site decision-making efficiency.

[0133] The volumetric calculation accuracy of the intelligent repair scheme can be set to 1% or 3%, and the number of iterations for strengthening structure optimization can be set to 100 or 200. Autodesk Revit can be used as the CAD software, and ANSYS Mechanical as the finite element analysis engine. The design module can run on a Dell Precision 7760 mobile workstation, interacting with the AR system via the STEP format. After selecting the defect area, a 3D mesh model of the repair body is automatically generated, variable density topology optimization is performed, and a BOM (Bill of Materials) is exported. Through this automated design process, repair schemes are generated rapidly, reducing manual modeling time and improving the mechanical rationality of the strengthening structure.

[0134] The refresh rate of the stress cloud map in the virtual-real simulation can be set to 25 frames / second or 60 frames / second, and the interactive response latency can be controlled within 0.5 seconds or 1 second. The mechanical solver can be COMSOL Multiphysics, and the rendering engine can be Unity 3D. The simulation server can be deployed in a local server room and connected to a 4K display via an HDMI 2.1 interface. Operators can adjust the reinforcement scheme parameters through gestures, updating the stress distribution in real time and displaying the maximum deformation. Through interactive simulation verification, the design scheme can be visualized and optimized, improving the economic efficiency of the maintenance scheme.

[0135] It should be noted that although the steps are described in a specific order above, this does not mean that they must be performed in that order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required functionality is achieved. The number of devices and processing scale described herein are for simplification of the invention; applications, modifications, and variations of this invention will be readily apparent to those skilled in the art.

[0136] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for quality inspection of wharf revetment blocks, characterized in that, Includes the following steps: S1: A mobile 3D laser scanning device is used to scan the protective block to obtain laser point cloud data. A 3D surface model is constructed through a real-time point cloud registration algorithm. Surface geometric features, including surface crack depth and peeling area, are extracted based on a deep convolutional neural network. Areas where the surface crack depth exceeds a preset depth threshold or the peeling area is greater than a preset area threshold are marked as high-risk areas, and a heat map of high-risk areas is generated. S2: Based on the thermal map of high-risk areas, a frequency-adjustable dual-mode flaw detection device is used to detect voids and cracks inside the protective block: an ultrasonic probe is used to perform intensified scanning in high-risk areas; multiple impact points are arranged in the projection area of ​​the surface crack extension direction according to the stress diffusion path, and an array-type impact echo device is used to collect vibration signals at the impact points; A three-dimensional defect probability distribution map is established by fusing ultrasonic data and vibration signal data through a multi-sensor data fusion algorithm, and areas where the diameter of internal cavities or the length of cracks are greater than the corresponding upper limit threshold are marked as severe defect areas. S3: Drill multiple core samples within a predetermined distance around the severely defective area and conduct compressive strength tests and acoustic emission damage detection; conduct infrared spectral composition surveys of non-severely defective areas and detect concrete carbonation degree and chloride ion erosion index; In the defect-matrix interface transition zone, micro-area hardness test points are arranged according to the material property gradient and hardness tests are performed. S4: Establish a dynamic evaluation model that includes surface defect distribution entropy, internal defect volume ratio, and material property dispersion coefficient; calculate the comprehensive safety factor S using a multi-objective optimization algorithm, and compare it with the upper limit of the safety threshold S. u and lower limit S d The comparison is performed to generate a 3D visualization inspection report.

2. The method for quality inspection of wharf revetment blocks according to claim 1, characterized in that, In step S1, the robotic arm of the mobile 3D laser scanning device integrates a dual-mode lidar and an inertial measurement unit, and performs the following operations: Based on continuous three-point cloud slice data, a differential geometry algorithm is used to calculate the local curvature change rate of the surface of the protective block in real time. The curvature change rate is quantified by the ratio of the Gaussian curvature difference between adjacent measurement points to the arc length distance. When the local minimum curvature radius is less than 2 meters, the surface roughness index exceeds the standard defined threshold of 0.5 mm, and environmental vibration in the range of 0.1-5 Hz is detected, the high frame rate lidar mode is activated, and the scanning frame rate is increased to no less than 200 Hz. An adaptive Kalman filter is used to eliminate low-frequency vibration noise caused by waves. By fusing the state equations of the robot arm's pitch angle, angular velocity, and linear acceleration with the observation data of the inertial measurement unit, the process noise and observation noise covariance matrix is ​​updated in real time to suppress wave interference. In high frame rate LiDAR mode, the movement speed of the robotic arm end effector is dynamically adjusted according to the local minimum radius of curvature. When the minimum radius of curvature is less than 1 meter, a low-speed scan of 0.2 m / s is used, and when the radius of curvature is in the range of 1-2 meters, a medium-speed scan of 0.5 m / s is used. At the same time, the laser line density is increased to 800 lines per meter, and the point cloud overlap rate of adjacent scanning paths is ensured to be no less than 40%.

3. The method for quality inspection of wharf revetment blocks according to claim 2, characterized in that, In step S1, the real-time point cloud registration algorithm specifically includes the following steps: Coarse registration is performed: an improved iterative nearest point algorithm is adopted, with the weighted sum of geometric distance and curvature difference as the optimization objective. The curvature difference is quantified by the mixed difference between the Gaussian curvature and the average curvature of the corresponding point, and the Gaussian curvature weight factor is set to 0.

5. When screening matching point pairs, only point pairs with a principal curvature radius ratio greater than 0.7 and a cosine value of the angle between the normal vectors greater than 0.9 are retained, and mismatches with inconsistent surface geometric features are excluded. Fine registration is performed: the Gauss-Newton iterative algorithm is accelerated based on field-programmable gate arrays, the Jacobian matrix containing rotation and translation components is calculated in parallel, and the optimal transformation parameters are solved by damped least squares method. The damping factor is set to 0.0001 to prevent matrix singularity. Combined with 3D feature descriptor matching, the local surface is divided into 32 orientation intervals and the height distribution histogram is statistically analyzed. The chi-square distance is used to calculate feature similarity, and regions with matching scores below the threshold are marked as low confidence regions. Dynamic evaluation and replanning are performed: the registration error of the local area is calculated in real time. When the root mean square error of the area exceeds 0.8 mm, it is determined to be a registration failure area; the scanning path replanning is triggered, and a ±15 degree deflection angle is added to the original scanning path to increase the point cloud line density to 1200 lines per meter and the overlap rate to 60%. The low confidence area is also re-scanned.

4. The method for quality inspection of wharf revetment blocks according to claim 1, characterized in that, In step S2, the frequency-tunable dual-mode flaw detection device includes an ultrasonic probe and an array-type impact echo device, wherein the ultrasonic probe adopts the following layered waveguide structure and dynamic control strategy: The waveguide layer of the ultrasonic probe consists of a piezoelectric composite substrate and five gradient impedance matching layers. The impedance value of each layer gradually changes exponentially from the substrate to the coupling agent interface. The total thickness of the matching layer is 2 mm to suppress sound energy scattering caused by surface roughness. When the arithmetic mean deviation of surface roughness is detected to exceed 12.5 micrometers, a wideband sweep frequency detection mode is activated. The frequency is continuously swept at a rate of 50 kHz per second in the transmission frequency range of 1.5 MHz to 2.5 MHz, and the transmission pulse width is dynamically adjusted according to the sweep frequency range. The effective focusing depth is calculated based on the longitudinal wave velocity of the material and the current transmission frequency. When the focusing depth fluctuates beyond ±10% of the nominal value, the calibration mode is triggered: a 1 MHz reference pulse is emitted, and the actual wave velocity of the material is corrected in reverse through the echo time difference to eliminate the sound velocity error caused by concrete aging. A nonlinear relationship model between contact pressure and surface roughness was established, and the probe height was adjusted in real time using a piezoelectric actuator to stabilize the coupling agent pressure in the range of 0.2 to 0.5 MPa. Dynamic balance is achieved using a proportional-integral-derivative controller with a proportional coefficient of 0.8, an integral time of 0.1 seconds, a derivative time of 0.05 seconds, and a pressure sensor sampling frequency of 1 kHz.

5. The method for quality inspection of wharf revetment blocks according to claim 4, characterized in that, In step S2, the array-type impact echo device locates internal defects using the following method: The impact energy is controlled to five levels by adjusting the impact speed of the array-type impact echo device: 0.5J, 1J, 2J, 3.5J, and 5J; the impact frequency is dynamically selected according to the Vickers hardness of the surface: 10 Hz high-frequency impact is used when the hardness is below 300 MPa, 5 Hz medium-frequency impact is used when the hardness is in the range of 300-500 MPa, and 2 Hz low-frequency impact is used when the hardness exceeds 500 MPa. The surface wave velocity is calculated based on the material's elastic modulus, density, and Poisson's ratio. The surface wave wavelength is calculated based on the central analysis frequency and the surface wave velocity. Piezoelectric sensors are arranged at intervals one-eighth of the surface wave wavelength. Several piezoelectric sensors are arranged in a spiral-grid hybrid layout to form a sensor array covering three-dimensional space. The vibration signal acquired by the echo device is fused with short-time Fourier transform and wavelet packet decomposition. The short-time Fourier transform uses a Gaussian-modulated cosine window function, with the window width dynamically adjusted according to the impact frequency. The wavelet packet decomposition uses a 10th-order Daubechies wavelet basis, based on the sampling frequency f. sample With center frequency f central Set the number of decomposition layers to A dual-channel analysis method with high and low frequencies is adopted to calculate the signal energy integral value of the low-frequency channel (0.5-10 kHz) and the kurtosis index of the high-frequency channel (10-50 kHz). Based on the data from the sensor array, a time-reversal algorithm is used to reconstruct the three-dimensional wave field, with the observation time window length set to 50 milliseconds. When the energy integral value of the low-frequency channel signal exceeds 0.8 times the reference value of healthy concrete, the kurtosis index value of the high-frequency channel exceeds the threshold, and the diameter of the three-dimensional wave field focusing area is greater than 10 centimeters, the excess area is judged as a valid internal defect.

6. The method for quality inspection of wharf revetment blocks according to claim 1, characterized in that, Step S3, specifically, includes the following steps: Based on the fusion data of laser displacement sensor and inertial measurement unit, the drill bit axis offset angle is calculated in real time: the lateral offset is measured by laser sensor, and mechanical vibration error is compensated by gyroscope angular velocity data, with the compensation coefficient set to 0.05 seconds / radian; the stroke of four hydraulic cylinders is adjusted by fuzzy proportional-integral-derivative controller, with the goal of minimizing the weighted sum of offset angle and axial force change rate, and the maximum deflection angle is limited to ±15 degrees, triggering an alarm when the limit is exceeded; The dimensionless working condition index is calculated in real time and defined as the product of axial drilling force and drill radius divided by the product of feed rate and hydraulic pressure. When the index exceeds the set threshold, three levels of protection are implemented: the retraction speed is adjusted to twice the current feed rate plus 50 mm / s; the drill speed is reduced to 300 rpm; and coolant is sprayed at a flow rate of 0.2 times the axial force. The following reinforcement methods were used in the core sample compressive strength test: The loading rate is controlled in three stages: a slow loading rate of 0.5 MPa / s is used when the estimated intensity is below 30%, the rate is increased to 1.0 MPa / s in the 30%-70% range, and the rate is reduced to 0.2 MPa / s after exceeding 70%. Acoustic emission signals are collected synchronously. The number of events with amplitudes exceeding 45 dB to 95 dB is counted to calculate the damage evolution b-value. When the rate of change of the b-value exceeds the set critical threshold, it is determined to be a precursor to critical damage. A linear positive correlation model between the axial force-feed speed ratio and compressive strength, and a negative correlation model between the dimensionless working condition index time integral and strength were established to comprehensively calculate the actual strength of the core sample.

7. The method for quality inspection of wharf revetment blocks according to claim 1, characterized in that, The infrared spectral composition survey of non-severely defective areas in step S3 specifically includes the following steps: For the 2.5–5 μm short-wave infrared channel, a quantum cascade laser with a center frequency of 120 terahertz was used as the light source to scan the characteristic peaks of silicate crystals at a resolution of 4 cm⁻¹. For the 8–12 μm long-wave infrared channel, the characteristic peaks of carbonate and chloride ions were scanned with a step size of 0.5 μm, and the single integration time was set to 50 ms. The fusion weights were dynamically allocated based on the dual-band signal-to-noise ratio, and the weight coefficients were set as the normalized results of the squared values ​​of the signal-to-noise ratios of each band. By calibrating the influence of ambient temperature using a blackbody radiation source at a specified temperature, a wavelength drift model including the steady-state thermal expansion coefficient and transient response coefficient is established; the original spectrum is nonlinearly corrected, and Planck's blackbody radiation formula is used to eliminate ambient thermal radiation interference. An improved convolutional recurrent neural network was used for analysis. The network input layer received a 256×256-dimensional dual-band spectral stitching matrix. The convolutional module used a dilated convolutional kernel with an expansion rate of 2 to extract multi-scale features, and the recurrent module used a bidirectional long short-term memory network to capture spectral sequence dependencies. The hidden layer dimension was set to 128. During model training, a weighted cross-entropy loss function was used, with weights inversely distributed according to the material aging level. ±2% spectral perturbation was added to generate adversarial examples to improve robustness. When calculating the aging index, the output probability of the neural network is fused and assigned a weight of 60%, the carbonation index is calculated and assigned a weight of 30%, and the chloride ion corrosion index is calculated and assigned a weight of 10%. The weighted sum is used to calculate the final aging index. When the aging index exceeds 0.6, it is judged as severe aging, and a maintenance warning is generated.

8. The method for quality inspection of wharf revetment blocks according to claim 1, characterized in that, In step S4, the method for the dynamic evaluation model to conduct a comprehensive structural safety assessment is as follows: An improved weighted Shannon entropy model was used to calculate the disorder of surface defect distribution, with the coupling effect of crack depth and area as the weighting factor. A fractal dimension correction term was introduced. When the fractal dimension exceeds 1.2, the total entropy value is increased by 30% of the fractal dimension deviation to enhance the sensitivity to complex crack networks. Calculate the effective stress concentration factor by setting the reference volume to 1000 cubic centimeters and calculating the void volume by summing the equivalent sphere radii. When the crack length exceeds 30 cm, the stress concentration factor is increased by an additional correction term, and the correction magnitude is proportional to the length exceeding the value; when continuous surface cracks are detected, the internal defect weighting factor is increased by 40%; The uniformity of material property distribution is evaluated based on the spatial autocorrelation Moran's I index. When calculating the index, an exponential decay function is used to define the spatial weight matrix, and the decay coefficient is set to 0.5 m. When the Moran's I index exceeds 0.3, it is determined that the material properties are clustered. The dispersion coefficient is amplified according to the exponential difference, and the amplification factor is set to a maximum of 0.

5. The safety factor is dynamically synthesized: the surface entropy value, the corrected internal defect coefficient, and the dispersion coefficient are calculated using the inverse weighting formula with a weight ratio of 0.5:0.3:0.2 to obtain the comprehensive safety factor; when the safety factor is lower than 1.2, a red warning is generated and the high-risk area is marked in the three-dimensional model.

9. The method for quality inspection of wharf revetment blocks according to claim 1, characterized in that, In step S4, the IoT module is configured to optimize the evaluation process. The operation of the IoT module includes the following steps: The detection data stream is decomposed into wavelet packets and dynamically quantized and compressed. The Daubechies 9 / 7 wavelet basis function is used to decompose the data into multiple sub-bands, and the quantization step size is dynamically adjusted according to the signal-to-noise ratio of each sub-band. Under normal conditions, a compression ratio of 10:1 is used, and when a high-risk defect is detected, it is switched to a 5:1 high-fidelity compression mode to prioritize the preservation of high-frequency detail features. A multi-physics compensation matrix for temperature, humidity, and pressure is constructed to quantify the influence of environmental factors on the measured values. The sensitivity of strain measurements to temperature is set to 0.3 microstrain / degree Celsius, and the sensitivity of stress measurements to humidity is set to 0.05 MPa / percent relative humidity. Environmental data is collected in real time and substituted into the compensation matrix to correct the strain and stress detection results. A time series prediction model is constructed based on a long short-term memory network. The input feature vector includes surface defect entropy, internal defect volume ratio, material dispersion coefficient, and environmental parameters. The hidden layer dimension of the network is set to 64, and the time step covers 24 consecutive hours of data. The output is the risk probability value for the next 30 days. The risk level is divided into three levels: low risk is below 0.3, medium risk is between 0.3 and 0.6, and high risk is above 0.

6. Key parameters are identified through the Fisher information matrix, and an elasticity coefficient of 0.8 is set to limit the parameter update range, while retaining the characteristic distribution of historical data.

10. The method for quality inspection of wharf revetment blocks according to claim 9, characterized in that, The IoT module includes a visualization terminal interface unit. The 3D visualization inspection report uses an augmented reality interface and includes the following methods and steps: An environmental point cloud map is constructed based on synchronous localization and mapping technology to achieve six-degree-of-freedom spatial alignment between the detection data coordinate system and the real scene, and the real-time update error of the pose matrix is ​​controlled to be less than 0.3 mm. The surface morphology layer is superimposed and displayed with a resolution of 0.5 mm / pixel, and the defect distribution layer is colored with red-orange gradient. The coloring intensity increases linearly with the crack depth, and the transparency drops to 0.2 when the depth exceeds 3 mm. Based on the recognition of hand gestures using a 3D convolutional neural network, the system takes 16 consecutive depth image sequences as input, fits the spatial trajectory of the hand gestures, and outputs operation commands. It also performs voxel cutting on any specified plane to generate material property profiles, extracts the mechanical parameter distribution of the profiles using a Gaussian kernel interpolation algorithm to generate profile stress distribution maps, and sets the interpolation bandwidth to a fixed 5 mm to suppress noise interference. The volume of the defect area to be repaired is calculated, the Heaviside function is used to determine the void boundary, and the integration step size is set to 1 mm. With the goal of maximizing stiffness and minimizing material usage, the volume of the reinforcement material does not exceed 30% of the original structure, the optimization density threshold is set to 0.3, and the topology optimization is performed to design the reinforcement structure. The visualization terminal interface unit displays the Mises stress cloud map of the reinforced structure in real time. The reinforcement scheme is modified by gesture interaction, and the mechanical simulation results are dynamically updated, with an optimization iteration cycle of less than 2 seconds.