Building structure crack intelligent detection method

Abstract

The invention discloses a building structure crack intelligent detection method, which comprises the steps of S1, constructing a multi-dimensional intelligent sensing array, and obtaining four-dimensional spatio-temporal data including vision, stress, vibration and temperature; s2, performing feature enhancement processing on the multi-modal data; s3, realizing cross-modal crack identification and positioning based on a space-time attention neural network; s4, adopting a sub-pixel edge detection and ultrasonic genetic inversion algorithm; s5, constructing a coupling dynamics prediction model; s6, establishing a dynamic threshold evaluation and multi-dimensional self-calibration mechanism; 0.05 mm micro-crack identification and 3mm depth detection precision are achieved and are improved by 50% compared with a traditional method, the detection accuracy under the complex environment is larger than or equal to 98% through the multi-modal data fusion and GAN enhancement technology, the RMSE is smaller than or equal to 0.03 mm through short-term prediction, the accuracy is larger than or equal to 85% through long-term prediction, double guarantees of a physical mechanism and data driving are established, and the detection accuracy is improved by more than or equal to 98% through the multi-modal data fusion and GAN enhancement technology. And the system has self-calibration, self-adaptive sampling and multi-modal fusion decision-making capabilities, so that the manual intervention cost is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of building crack detection, and in particular to an intelligent detection method for building structure cracks. Background Art

[0002] In the field of construction engineering, cracks are one of the most common structural defects, and their generation and development directly affect the safety, durability, and applicability of the structure. Traditional detection methods have technical bottlenecks. Among them, existing visual inspection is easily interfered by light and stains, ultrasonic inspection relies on the application of artificial coupling agents, and vibration detection is not sensitive enough to microcracks. A single technology is difficult to cope with complex working conditions. There is a lack of collaborative analysis of the dynamic characteristics of the multi-field coupling of stress, strain, vibration, and temperature during the development of cracks, and it is impossible to accurately describe the physical mechanism of crack expansion. Deep learning models rely on large-scale labeled data, have weak generalization capabilities for rare crack morphologies, and have not established an effective fusion mechanism for physical models and data-driven models. Existing time series models do not fully consider the influence of multiple variables such as environmental loads and material aging, and the long-term prediction error rate exceeds 20%.

[0003] With the advancement of smart city construction, building structures are becoming more complex and large-scale. There is an urgent need for a detection technology that integrates multi-source data fusion, physical mechanism modeling, and intelligent algorithm innovation to achieve a technological leap from static detection to dynamic monitoring, and from qualitative analysis to quantitative prediction. Summary of the Invention

[0004] The purpose of this invention is to solve at least one of the technical problems existing in the prior art and provide an intelligent detection method for building structure cracks, which can achieve 0.05mm micro-crack identification and 3mm depth detection accuracy, which is 50% higher than the traditional method. Through multimodal data fusion and GAN enhancement technology, the detection accuracy in complex environments is ≥98%.

[0005] By achieving short-term prediction RMSE ≤ 0.03mm and long-term prediction accuracy ≥ 85%, the system establishes dual guarantees of physical mechanism and data-driven, and possesses self-calibration, adaptive sampling, and multi-modal fusion decision-making capabilities, significantly reducing the cost of manual intervention.

[0006] The present invention also provides the above-mentioned intelligent detection method for cracks in building structures, comprising the following steps:

[0007] S1. Build a multi-dimensional intelligent sensor array to obtain four-dimensional spatiotemporal data including vision, stress, vibration, and temperature;

[0008] S2. Feature enhancement processing of multimodal data, including image spectrum compensation and depth-guided enhancement, temperature compensation calibration of stress and strain data, and time-frequency energy aggregation processing of vibration data;

[0009] S3, based on spatiotemporal attention neural network, cross-modal crack recognition and location are realized, and the crack area is determined by combining the strain energy density factor;

[0010] S4, using sub-pixel edge detection and ultrasonic genetic inversion algorithm to achieve high-precision measurement of crack width, length and depth;

[0011] S5. Construct a coupled dynamic prediction model, combining the spatiotemporal sequence Transformer and the modified Paris formula to achieve short-term and long-term predictions of crack development trends;

[0012] S6. Establish dynamic threshold assessment and multi-dimensional self-calibration mechanism to generate intelligent detection results and early warning decisions.

[0013] According to a method for intelligent detection of cracks in building structures provided by the present invention, the multi-dimensional intelligent sensor array comprises:

[0014] Visual sensing layer: 8K linear array camera with active polarization filtering and structured light projector to generate visual images containing 3D point cloud information;

[0015] Stress sensing layer: Distributed fiber Bragg grating (FBG) sensors and MEMS piezoresistive stress gauges are hybrid networked. FBG sensors are arranged along the main reinforcement of the structure, and stress gauges are densely packed in the node areas.

[0016] Vibration sensing layer: Three-axis micro-electromechanical accelerometer, with optimized placement based on modal analysis to ensure capture integrity of the first five modal shapes ≥ 95%;

[0017] Environmental sensing layer: Integrates temperature and humidity, atmospheric pressure, and light intensity sensors for data calibration and modeling of environmental influencing factors.

[0018] According to the intelligent detection method for building structure cracks provided by the present invention, the spatiotemporal synchronous acquisition protocol adopts a variable density sampling strategy, and the formula is as follows:

[0019] f s =f base ×(1+k×σ strain )

[0020] Where: f base is the basic sampling rate, σ strain is the real-time strain standard deviation, k is the dynamic adjustment coefficient (0.5-2.0), and adaptive encrypted sampling is achieved in high strain gradient areas with a synchronization error of ≤1με.

[0021] According to a method for intelligent detection of cracks in building structures provided by the present invention, the image feature enhancement process adopts a generative adversarial network (GAN) model, which includes a spectral compensation branch and a depth guidance branch:

[0022] The spectral compensation branch separates the reflected light and scattered light components through polarization image decomposition and constructs an illumination invariant feature map;

[0023] The depth-guided branch uses structured light 3D point clouds to generate a pseudo-color map of crack depth and constructs a multi-channel input with the visual image.

[0024] The WGAN-GP loss function is used for adversarial training to generate enhanced samples with different stains and lighting conditions.

[0025] According to the intelligent detection method for cracks in building structures provided by the present invention, the stress-strain data calibration model is a two-variable calibration model, and the formula is as follows:

[0026] ε true =ε measured -αΔT

[0027] λ true =k ε ε true +K T ΔT

[0028] Where: α is the thermal expansion coefficient of the material, k ε =1.2pm / με, K T =10pm / ℃ is the sensor sensitivity coefficient, and the temperature drift curve is fitted in real time by the least square method.

[0029] According to the intelligent detection method for cracks in building structures provided by the present invention, the time-frequency feature extraction of vibration data adopts an improved synchronous compressed wavelet transform (SWT), and energy aggregation is achieved through a time-frequency redistribution algorithm. The formula is as follows:

[0030]

[0031] in For instantaneous frequency estimation, the ability to detect frequency changes at the 0.1Hz level is improved.

[0032] According to a method for intelligent detection of cracks in building structures provided by the present invention, the spatiotemporal attention neural network (STANet) includes:

[0033] Spatial branch: A ResNeXt-50 backbone network embedded with a crack direction perception module is used to generate anisotropic feature weights through gradient direction histogram (GDH);

[0034] Time branch: 3D convolution-LSTM hybrid structure to capture the dynamic correlation of multiple time series data;

[0035] Fusion module: The cross-modal attention mechanism (CMA) calculates the mutual information of visual and stress vibration features. The decision-level fusion formula is as follows:

[0036] M=Softmax(W υ V+W S S+W υ V⊙W S S)

[0037] Among them, V and S are visual and stress vibration feature vectors respectively, and W υ 、W S is the learnable weight matrix.

[0038] According to the present invention, a method for intelligent detection of cracks in building structures is provided. The crack location method is based on the strain energy density factor and defines a position indication function:

[0039]

[0040] The strain energy density was calculated by inversion of the finite element model and the measured stress. The crack area was identified at that time, and the positioning accuracy reached 1 / 5 of the sensor spacing.

[0041] The crack width measurement adopts a sub-pixel edge detection algorithm based on Zernike moments, and the edge normal direction is calculated by fitting the second-order moment distribution. The conversion formula is as follows:

[0042]

[0043] Where d is the pixel spacing, p is the sub-pixel edge spacing, and R is the camera calibration coefficient. The measurement accuracy is 0.1 pixel.

[0044] The fracture depth inversion adopts the ultrasonic propagation model optimized by genetic algorithm, and the objective function is as follows:

[0045] Among them, x i is the horizontal distance of the measuring point, v is the longitudinal wave velocity, and the regularization term is introduced to solve the multi-solution problem. The depth measurement error is ≤3%;

[0046] The short-term prediction adopts the spatiotemporal sequence Transformer model to encode the crack parameters and environmental loads into a spatiotemporal position vector: X = E p +E t +E f

[0047] Among them E p is the spatial position code, E t For timestamp encoding, E f For feature encoding, long-distance dependencies are captured through a multi-head self-attention mechanism, with a prediction step length of 72 hours and a root mean square error (RMSE) ≤ 0.03mm;

[0048] The long-term prediction combines fracture mechanics with machine learning to modify the Paris formula as follows:

[0049]

[0050] Where ΔK is the stress intensity factor amplitude, ΔK th is the threshold value, f(T,RH) is the temperature and humidity influence function, and the environmental factors are fitted by random forest regression. The 5-year prediction accuracy is ≥85%.

[0051] According to the present invention, a method for intelligent detection of cracks in building structures is provided. The intelligent decision-making system establishes a dynamic threshold evaluation model based on a Bayesian network. The input parameters include crack geometry parameters, material performance parameters, and environmental parameters. The dynamic safety threshold is generated through Monte Carlo simulation.

[0052] The multi-dimensional self-calibration mechanism includes:

[0053] Sensor level: built-in standard strain block for regular calibration, interval ≤ 24h;

[0054] Algorithm level: Use unlabeled data for self-supervised learning to update model parameters and automatically train weekly;

[0055] System level: Achieve closed-loop optimization of detection accuracy through comparative analysis of historical data and finite element simulation.

[0056] Compared with the existing technology, the intelligent detection method of building structure cracks in this invention can achieve 0.05mm micro-crack recognition and 3mm depth detection accuracy, which is 50% higher than the traditional method. Through multimodal data fusion and GAN enhancement technology, the detection accuracy in complex environments is ≥98%.

[0057] Compared with the existing technology, the intelligent detection method for building structure cracks of the present invention establishes dual guarantees of physical mechanism and data-driven through short-term prediction RMSE ≤ 0.03mm and long-term prediction accuracy ≥ 85%, and has self-calibration, adaptive sampling, and multimodal fusion decision-making capabilities, greatly reducing the cost of manual intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The present invention will be further described below with reference to the accompanying drawings and embodiments;

[0059] Figure 1 This is an overall flow chart of an intelligent detection method for cracks in building structures according to the present invention. DETAILED DESCRIPTION

[0060] This section will describe in detail the specific embodiments of the present invention. The preferred embodiments of the present invention are shown in the accompanying drawings. The purpose of the accompanying drawings is to supplement the description of the text part of the specification with graphics, so that people can intuitively and vividly understand each technical feature and the overall technical solution of the present invention, but it should not be understood as a limitation on the scope of protection of the present invention.

[0061] Reference Figure 1 An embodiment of the present invention provides an intelligent detection method for cracks in building structures, which includes: multi-dimensional intelligent sensor array deployment and data acquisition, wherein the sensor installation process is divided into a visual sensing layer, which uses an 8K resolution linear array camera, an integrated 940nm structured light projector with a projection angle of 15°, a polarizing filter, a transmittance of ≥90%, and an extinction ratio of ≥1000:1.

[0062] The installation method is to use a magnetic universal bracket, adjust the accuracy of ±0.1° to fix the camera, ensure that the lens optical axis is perpendicular to the surface to be measured, control the distance between 1.5-3.0 meters, and cover the detection area of ​​2m×1.5m.

[0063] The structured light projector is installed at an angle of 35° to the camera. The light plane parameters are calibrated with a calibration plate with an accuracy of ±0.05mm to generate three-dimensional point cloud data with a point spacing of ≤0.5mm.

[0064] The stress sensing layer is made by pasting FBG sensors every 500mm along the main reinforcement of the concrete beam. The grid area length is 10mm, the reflectivity is ≥90%, and epoxy resin glue is used. The curing time is 24h and the working temperature is -40℃~80℃. A 100mm pigtail is reserved as a temperature compensation section.

[0065] MEMS piezoresistive stress gauges (range ±5000με, resolution 1με) are densely arranged in the node areas of the steel structure and fit tightly to the surface of the structure through special adapters.

[0066] The vibration sensing layer uses PCB356A16 triaxial accelerometers with a frequency response of 0.5Hz-10kHz. They are rigidly fixed to vibration-sensitive points on the structure using M5 stainless steel bolts with a torque of 8N·m. Based on ANSYS modal analysis results, the sensor placement was optimized to ensure measurement errors of ≤0.5% for the first five natural frequencies and a mode shape capture integrity of ≥95%.

[0067] The environmental sensing layer integrates the SHT30 temperature and humidity sensor (accuracy of ±0.3°C, ±2%RH) and the MS5611 atmospheric pressure sensor (accuracy of ±0.1hPa), which are deployed synchronously with other sensors with a sampling frequency of 1Hz.

[0068] The time and space synchronization acquisition process uses the IEEE1588v2 time server to synchronize time for all sensors, with a synchronization error of ≤1με.

[0069] Visual camera calibration: Zhang’s calibration method is used to obtain the intrinsic parameter matrix (focal length error ≤ 0.1px, distortion coefficient ≤ 0.01).

[0070] FBG sensor calibration: load standard strain blocks (0με, 500με, 1000με), record the wavelength drift, and establish the strain-wavelength mapping relationship.

[0071] Adaptive sampling strategy: real-time calculation of strain data standard deviation σ strain , when σ strain When the sampling rate is >50, the vibration sensor sampling rate is increased from 10kHz to 20kHz, and the strain sensor sampling rate is increased from 1kHz to 5kHz. The visual sensor uses dynamic sampling of regions of interest (ROIs) to increase the resolution to 4K local scans in areas with high strain gradients (such as suspected cracks).

[0072] Data caching mechanism: Establish a circular buffer (capacity for 24 hours of data) and store it in HDF5 format, including timestamp, sensor ID, raw data and calibration parameters, supporting breakpoint resuming and abnormal data marking.

[0073] Multimodal data feature enhancement processing, in which spectral compensation branch polarization image decomposition is performed by collecting linear polarization (0°, 45°, 90°, 135°) images of the same scene and calculating the reflected light intensity I through the Stokes vector s And scattered light intensity:

[0074]

[0075] Constructing illumination invariant feature map: F li =log(I d +1)-log(I s +1), eliminating reflection interference.

[0076] Depth-guided branch, in which 3D point cloud processing generates a depth map with an accuracy of ±1mm through the structured light stripe image through the Gray code decoding algorithm, which is converted into point cloud data (XYZ coordinates).

[0077] Crack depth pseudo-color map generation: The depth value is normalized to [0, 255], mapped to the lightness channel of the HSV color space, and merged with the RGB image into a 4-channel input.

[0078] GAN training details, including the network structure, generator: ResNetGenerator, input 3 channels (RGB) + 1 channel (depth), and output enhanced image.

[0079] Discriminator: UNetDiscriminator, using spectral normalization (SpectralNorm) for stable training.

[0080] Loss function:

[0081] Where λ = 0, the training period is 500 epochs, and the batch size is 32.

[0082] Stress and strain data calibration, where the temperature compensation model is established under 20℃, 25℃, 30℃, 35℃, and 40℃ environments, collecting unloaded FBG wavelength data and fitting the temperature drift curve: λ T =λ0+K T ΔT+δ T ,(δ T ~N(0,0.05pm)), where λ0 is the central wavelength at 20℃, K T =10pm / ℃ is the temperature sensitivity.

[0083] Measured strain value ε in the strain calibration process measured After temperature compensation, the true strain is obtained: K ε =1·2pm / με, automatically collect ambient temperature data every hour and update compensation parameters in real time.

[0084] Vibration data time-frequency processing, which is achieved by synchronous compression wavelet transform (SWT), by selecting Meyer wavelet as the mother wavelet, decomposition layer 5, covering the frequency range of 0.1Hz-500Hz. Time-frequency redistribution algorithm: calculate the instantaneous frequency Redistribute each wavelet coefficient to the most relevant time-frequency point to generate a time-frequency diagram after energy concentration:

[0085] The 12-dimensional features of the time-frequency graph, such as energy entropy and center frequency, are extracted as input for subsequent analysis.

[0086] Cross-modal crack recognition and localization, in which the spatiotemporal attention neural network (STANet) is trained through an image dataset, collecting 100,000 images of cracks in concrete and steel structures (resolution 2048×1536), annotating the crack width (0.05-5mm) and depth (1-50mm), and using data enhancement (rotation ±15°, Gaussian noise σ=0.05, brightness adjustment ±20%).

[0087] The stress-vibration data set was collected by simulating the load condition of 0% to 80% of the design load on the structural test bench, and 500 sets of spatiotemporal series data were collected: measurement point × time × feature = 20 × 1000 × 6.

[0088] Network training strategy space branch: Crack direction perception module: Calculate the gradient direction histogram (GDH, 8 direction bins) and generate the direction weight matrix W dir , weight the convolution feature map: F dir =F×W dir =Softmax(GDH)

[0089] The time branch extracts spatiotemporal features through a 3D convolutional layer (kernel size 3×3×3), and the LSTM layer (hidden unit 256) captures temporal dependencies, with the output dimension being time steps × number of features.

[0090] The training parameters are passed through the optimizer: AdamW (weight decay 0.01), learning rate 1e-4, and cosine annealing schedule (period 200 epochs).

[0091] Loss function: focal loss (α=0.8, γ=2) + Dice loss to handle the imbalance of positive and negative samples (the proportion of crack pixels ≤ 5%).

[0092] The reasoning process inputs 3-channel visual images (RGB) + 1-channel depth map + 6-channel stress and vibration data (XYZ direction stress, acceleration).

[0093] The crack probability map is output (resolution is consistent with the input), the probability threshold is 0.5, and the final detection box is generated by non-maximum suppression (NMS, IOU=0.3).

[0094] The strain energy density inversion positioning is established through the finite element model: ANSYS is used to build a three-dimensional structural model, using SOLID65 (grid size ≤ 100mm) for concrete elements and SOLID185 for steel structures. Material parameters (elastic modulus, Poisson's ratio) and boundary conditions are imported. Inverse calculation: Using the measured stress data as boundary conditions, the strain energy density field U is solved using the conjugate gradient method, and the position indicator function is calculated:

[0095] Threshold θ I Determined according to the fracture toughness of the material (such as 0.5MPa for C30 concrete), it is adaptively optimized using the Otsu algorithm.

[0096] Region refinement: I(x,y)>θ I A morphological closing operation (kernel size 5×5) is performed on the region to remove isolated noise points. The final positioning accuracy reaches 1 / 5 of the sensor spacing (for example, the positioning accuracy of a 500mm spacing sensor is 100mm).

[0097] Accurately measure crack parameters through sub-pixel width measurement and Zernike moment calculation: Calculate the 0-3 order Zernike moments of the grayscale image of the crack ROI area, with the geometric center as the origin:

[0098] Where V nm is the Zernike basis function, and f(x,y) is the pixel grayscale value.

[0099] Edge location is determined by the second-order moment a 20 、a 02 Calculate the edge normal direction: Sample 10 sub-pixel points along the normal direction and fit the edge curve.

[0100] The width conversion is done by calculating the sub-pixel spacing p between the two edges and converting the actual width based on the camera calibration coefficient R (unit: mm / pixel): d = 1.55 μm (physical pixel size) with a measurement accuracy of 0.1 pixel (corresponding to an actual width of 0.05 mm).

[0101] Ultrasonic depth inversion was performed using genetic algorithm parameters with a population size of 50, an iteration number of 100, a crossover probability of 0.8, a mutation probability of 0.05, and a real number encoding method (depth range 0–100 mm, precision 0.1 mm).

[0102] The objective function is to minimize the sum of the squares of the differences between the measured time and the theoretical time:

[0103] The regularization term λ = 0.01 is used to prevent overfitting, and the longitudinal wave velocity v is calibrated by a crack-free area of ​​the same material (error ≤ 2%).

[0104] Multi-point fusion is achieved by arranging more than three ultrasonic probes (with a spacing of 100-300 mm), taking the median of the inversion results, and the depth measurement error is ≤3%.

[0105] The crack development trend prediction is done through short-term prediction (space-time sequence Transformer) input encoding, where the space-time position vector X consists of three parts: spatial encoding E p :The sensor position coordinates (x, y, z) are generated by Fourier transform, and the time code E t :Timestamp is mapped to sine / cosine function, feature encoding E f : Crack parameters and environmental parameters.

[0106] The attention mechanism calculates the dependencies between features through multi-head self-attention (8 heads):

[0107] Position-wise feedforward network: two fully connected layers (activation function GELU), dimension 64→256→64.

[0108] Training configuration through time window: 72 hours (360 time steps, 12 minutes / step), forecasting the next 24 hours (120 steps)

[0109] Evaluation indicators: root mean square error (RMSE), mean absolute error (MAE). On measured data, RMSE is ≤ 0.03 mm.

[0110] Training configuration through time window: 72 hours (360 time steps, 12 minutes / step), forecasting the next 24 hours (120 steps)

[0111] Evaluation indicators: root mean square error (RMSE), mean absolute error (MAE). On measured data, RMSE is ≤ 0.03 mm.

[0112] Long-term predictions are calculated by fracture mechanics parameters: stress intensity factor amplitude Where Y is the geometric factor (obtained through finite element simulation), σ is the nominal stress, and a is the crack half-length. The environmental impact function is obtained by collecting 5 years of temperature and humidity data and fitting it through random forest regression: f(T,RH)=1+α T (T-20)+α RH (RH-60%)

[0113] where α T =0.01 / ℃,α RH =0.005 / %RH is the environmental sensitivity coefficient.

[0114] The model is trained using input features such as historical crack growth rates, stress spectra, and the number of annual temperature and humidity cycles. The output target is the crack length increment over the next 1-5 years, with a 5-year prediction accuracy of ≥85% (with a verified error of ≤15%).

[0115] Intelligent decision-making and self-calibration system, through dynamic threshold evaluation through Bayesian network structure, where the nodes include: crack width (0.05-5mm), depth (1-50mm), material strength (C25-C50), annual temperature difference (ΔT = 10-40℃), humidity cycle (50-200 times / year).

[0116] The conditional probability table was trained using 100,000 sets of simulation data, and Monte Carlo simulation was used to generate dynamic thresholds (e.g., the crack width threshold for C30 concrete decreases by 0.1 mm for every 10°C increase in annual temperature difference).

[0117] Multi-dimensional self-calibration involves sensor-level calibration, triggering standard strain block loading (0με, 1000με) at 0:00 daily, automatically calculating sensor drift and triggering hardware calibration if it exceeds 5%. Algorithm-level calibration utilizes weekly self-supervised learning from unlabeled data, updating feature extraction layer parameters through comparative learning to maintain the model's adaptability to new crack morphologies. System-level calibration involves monthly finite element simulations, comparing measured crack parameters with simulation results. Model fine-tuning is initiated when the error exceeds 10%, updating the physical constraint parameters.

[0118] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in the technical field without departing from the scope of the present invention.

Claims

1. A method for intelligent detection of cracks in building structures, characterized in that: The following steps are involved: S1. Build a multi-dimensional intelligent sensor array to obtain four-dimensional spatiotemporal data including vision, stress, vibration, and temperature; S2. Feature enhancement processing of multimodal data, including image spectrum compensation and depth-guided enhancement, temperature compensation calibration of stress and strain data, and time-frequency energy aggregation processing of vibration data; S3, based on spatiotemporal attention neural network, cross-modal crack recognition and location are realized, and the crack area is determined by combining the strain energy density factor; S4, using sub-pixel edge detection and ultrasonic genetic inversion algorithm to achieve high-precision measurement of crack width, length and depth; S5. Construct a coupled dynamic prediction model, combining the spatiotemporal sequence Transformer and the modified Paris formula to achieve short-term and long-term predictions of crack development trends; S6. Establish dynamic threshold assessment and multi-dimensional self-calibration mechanism to generate intelligent detection results and early warning decisions.

2. The intelligent detection method for building structure cracks according to claim 1, characterized in that: The multi-dimensional intelligent sensing array includes: Visual sensing layer: 8K linear array camera with active polarization filtering and structured light projector to generate visual images containing 3D point cloud information; Stress sensing layer: Distributed fiber Bragg grating (FBG) sensors and MEMS piezoresistive stress gauges are hybrid networked. FBG sensors are arranged along the main reinforcement of the structure, and stress gauges are densely packed in the node areas. Vibration sensing layer: Three-axis micro-electromechanical accelerometer, with optimized placement based on modal analysis to ensure capture integrity of the first five modal shapes ≥ 95%; Environmental sensing layer: Integrates temperature and humidity, atmospheric pressure, and light intensity sensors for data calibration and modeling of environmental influencing factors.

3. The intelligent detection method for cracks in building structures according to claim 1, characterized in that: The spatiotemporal synchronous acquisition protocol adopts a variable density sampling strategy, and the formula is as follows: f s =f base ×(1+k×σ strain ) Where: f base is the basic sampling rate, σ strain is the real-time strain standard deviation, k is the dynamic adjustment coefficient (0.5-2.0), and adaptive encrypted sampling is achieved in high strain gradient areas with a synchronization error of ≤1μs.

4. The intelligent detection method for cracks in building structures according to claim 1, characterized in that: The image feature enhancement process uses a generative adversarial network (GAN) model, which includes a spectral compensation branch and a depth guidance branch: The spectral compensation branch separates the reflected light and scattered light components through polarization image decomposition and constructs an illumination invariant feature map; The depth-guided branch uses structured light 3D point clouds to generate a pseudo-color map of crack depth and constructs a multi-channel input with the visual image. The WGAN-GP loss function is used for adversarial training to generate enhanced samples with different stains and lighting conditions.

5. The intelligent detection method for cracks in building structures according to claim 1, characterized in that: The stress-strain data calibration model is a two-variable calibration model, and the formula is as follows: e true =e measured -αΔT l true =k ε e true +K T ΔT Where: α is the thermal expansion coefficient of the material, k ε =1.2pm / με, K T =10pm / ℃ is the sensor sensitivity coefficient, and the temperature drift curve is fitted in real time by the least square method.

6. The intelligent detection method for cracks in building structures according to claim 1, characterized in that: The vibration data time-frequency feature extraction adopts the improved synchronous compression wavelet transform (SWT) and realizes energy aggregation through the time-frequency redistribution algorithm. The formula is as follows: in For instantaneous frequency estimation, the ability to detect frequency changes at the 0.1Hz level is improved.

7. The intelligent detection method for cracks in building structures according to claim 1, characterized in that: The spatiotemporal attention neural network (STANet) includes: Spatial branch: A ResNeXt-50 backbone network embedded with a crack direction perception module is used to generate anisotropic feature weights through gradient direction histogram (GDH); Time branch: 3D convolution-LSTM hybrid structure to capture the dynamic correlation of multiple time series data; Fusion module: The cross-modal attention mechanism (CMA) calculates the mutual information of visual and stress vibration features. The decision-level fusion formula is as follows: M=Softmax(W υ V+W S S+W υ V⊙W S S) Among them, V and S are visual and stress vibration feature vectors respectively, and W υ 、W S is the learnable weight matrix.

8. The intelligent detection method for cracks in building structures according to claim 1, characterized in that: The crack location method is based on the strain energy density factor and defines the position indication function: Where U is the strain energy density, which is calculated by inverse analysis of the finite element model and the measured stress. I It is determined to be a crack area, and the positioning accuracy reaches 1 / 5 of the sensor spacing; The crack width measurement adopts a sub-pixel edge detection algorithm based on Zernike moments, and the edge normal direction is calculated by fitting the second-order moment distribution. The conversion formula is as follows: Where d is the pixel spacing, p is the sub-pixel edge spacing, and R is the camera calibration coefficient. The measurement accuracy is 0.1 pixel. The fracture depth inversion adopts the ultrasonic propagation model optimized by genetic algorithm, and the objective function is as follows: Among them, x i is the horizontal distance of the measuring point, v is the longitudinal wave velocity, and the regularization term is introduced to solve the multi-solution problem. The depth measurement error is ≤3%; The short-term prediction adopts the spatiotemporal sequence Transformer model to encode the crack parameters and environmental loads into a spatiotemporal position vector: X = E p +E t +E f Among them E p is the spatial position code, E t For timestamp encoding, E f For feature encoding, long-distance dependencies are captured through a multi-head self-attention mechanism, with a prediction step length of 72 hours and a root mean square error (RMSE) ≤ 0.03mm; The long-term prediction combines fracture mechanics with machine learning to modify the Paris formula as follows: Where ΔK is the stress intensity factor amplitude, ΔK th is the threshold value, f(T,RH) is the temperature and humidity influence function, and the environmental factors are fitted by random forest regression. The 5-year prediction accuracy is ≥85%.

9. The intelligent detection method for cracks in building structures according to claim 1, characterized in that: The intelligent decision-making system establishes a dynamic threshold evaluation model based on a Bayesian network, with input parameters including crack geometry parameters, material performance parameters, and environmental parameters, and generates a dynamic safety threshold through Monte Carlo simulation; The multi-dimensional self-calibration mechanism includes: Sensor level: built-in standard strain block for regular calibration, interval ≤ 24h; Algorithm level: Use unlabeled data for self-supervised learning to update model parameters and automatically train weekly; System level: Achieve closed-loop optimization of detection accuracy through comparative analysis of historical data and finite element simulation.

Images