A Method and System for Predicting Bridge Crack Trends Based on Big Data
By using a big data-based method for predicting bridge crack trends and employing spatiotemporal conjugate tensors and shape boundary conditions to remove temperature response components, the method achieves accurate prediction of the long-term damage evolution trend of bridge cracks. This solves the problem of insufficient accuracy in predicting bridge health status trends in existing technologies and provides a scientific basis for prediction.
Patent Information
- Application Number
- CN202611256607.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-19
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-19
AI Technical Summary
Existing technologies rely excessively on static spatial AI visual feature recognition, which makes it impossible to effectively integrate multimodal spatiotemporal damage data. This makes it difficult to accurately capture the essential damage evolution trend of bridge cracks in the long-term time series dimension, and it is also difficult to eliminate the nonlinear alternating interference of environmental temperature on crack features, resulting in insufficient accuracy in predicting the long-term health status trend of bridges.
By collecting images of apparent defects in the cracked areas of bridges and sensor network data, three-dimensional spatial coordinate transformation and pixel-level edge segmentation are performed to form a spatiotemporal conjugate tensor. White noise signals are added for adaptive screening to remove high-frequency texture errors and temperature response components. Combined with shape boundary conditions and weather forecast temperatures, long-term damage evolution trends are predicted.
It has achieved accurate capture of the evolution trend of bridge cracks and high-precision prediction of long-term health status, eliminated the influence of temperature interference, ensured the physical rationality and reliability of the prediction results, and provided a scientific basis for bridge health assessment and maintenance.
Smart Images

Figure CN122841383B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bridge crack prediction technology, and in particular to a method and system for predicting bridge crack data trends based on big data. Background Technology
[0002] In recent years, with the rapid development of transportation infrastructure construction in my country, long-term safety monitoring and intelligent maintenance of bridge structures have become a focus of attention in the engineering community. As one of the most direct characteristics of structural damage, the accurate identification and prediction of the evolution trend of apparent cracks in bridges are of paramount engineering significance.
[0003] Currently, Chinese invention patent application CN121639703A discloses a method and system for dynamic detection of bridge construction cracks based on AI visual recognition. The method includes: acquiring bridge appearance images; extracting first-scale crack features through a cross-stage partial large-kernel separable kernel attention module and multiple double-convolutional cross-stage local bottleneck modules; performing aggregation-distribution mechanism through a neck network, performing feature alignment and feature fusion processing through low-stage aggregation branches and high-stage aggregation branches; predicting crack categories and regressing bounding box positions based on third-scale crack features, and outputting bridge crack detection results. However, the related technology does not consider the essential evolution trend of bridge cracks in the long-term temporal dimension, fails to effectively isolate and integrate the elastic fluctuation interference caused by the periodic alternation of environmental temperature on crack morphology, and does not involve a deep decoupling mechanism for multimodal spatiotemporal data or extrapolation prediction of long-term evolution trajectory under physical boundary constraints. Summary of the Invention
[0004] The technical problem solved by this invention is that existing technologies rely too heavily on static spatial AI visual feature recognition, which makes it difficult to effectively integrate multimodal spatiotemporal damage data and accurately capture the essential damage evolution trend of bridge cracks in the long-term time series dimension. At the same time, existing technologies are unable to eliminate or effectively utilize the nonlinear alternating interference caused by ambient temperature on crack characteristics. When making long-term extrapolation predictions, they are prone to masking the true damage accumulation trajectory due to short-term elastic expansion and contraction caused by temperature. Furthermore, due to the lack of constraints from physical boundary conditions, it is difficult to guarantee the mathematical and physical rationality of the predicted trajectory of the future evolution state of cracks, resulting in insufficient accuracy in predicting the trend of the long-term health state of bridges.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a method for predicting bridge crack data trends based on big data, comprising the following steps:
[0006] Step S1: Collect the original image of the apparent disease in the cracked area and obtain the long-period sampling time series. Perform three-dimensional spatial coordinate transformation and pixel-level edge segmentation processing, and reconstruct to obtain the spatiotemporal conjugate tensor.
[0007] Step S2: Add white noise signal to the spatiotemporal conjugate tensor and sieve it along the time axis. Group it according to the average fluctuation period, energy and coherence with the ambient temperature time series data to obtain the residual trend tensor and store the mid-frequency sub-tensor group in isolation.
[0008] Step S3: Discretize and iteratively solve the residual trend tensor under shape boundary conditions to obtain the evolution trend index sequence;
[0009] Step S4: Obtain the future weather forecast temperature time series, extrapolate the trajectory expectation value of the evolution trend index sequence within the future time step, calculate the periodic amplitude component based on the mid-frequency sub-tensor group and the future weather forecast temperature time series, and perform linear superposition to output the crack evolution state trajectory.
[0010] Preferably, step S1 specifically includes:
[0011] Step S11: Use the drone camera to collect the original images of the apparent damage in the cracked area, and simultaneously acquire the long-period sampling time series corresponding to the sensor network;
[0012] Step S12: Deblur the original image of the apparent defects to obtain a high-quality image of the bridge crack.
[0013] Step S13: Use the aerial triangulation algorithm to perform full-field collinearity equation adjustment calculation on high-quality bridge crack images, calculate the spatial internal and external parameters of the UAV camera and generate dense point cloud data.
[0014] Acquire laser point cloud data and register it with dense point cloud data to obtain the laser point cloud spatial coordinate transformation matrix.
[0015] Preferably, step S1 further includes:
[0016] Step S14: Use a computer vision network to perform pixel-level edge segmentation on the high-quality bridge crack image, extract the maximum crack width, physical length of the crack skeleton line and pixel-level apparent area of the crack region in the current iteration step, and mark the two-dimensional pixel coordinates.
[0017] Step S15: Using the laser point cloud spatial coordinate transformation matrix, the two-dimensional pixel coordinates are projected onto the spatial geometric surface of the bridge to calculate the spatial three-dimensional projection curvature.
[0018] Step S16: The maximum crack width, the physical length of the crack skeleton line, the pixel-level apparent area of the crack region, and the spatial three-dimensional projection curvature are tensorized and recombined according to a preset arrangement order to obtain the spatiotemporal conjugate tensor.
[0019] Preferably, step S2 specifically includes:
[0020] Step S21: Add M-order white noise signals to the spatiotemporal conjugate tensor to obtain the corresponding auxiliary mask signal matrix;
[0021] Step S22: Use the time-domain adaptive sieving operator to sieve the auxiliary mask signal matrix along the time axis, extract the sub-tensors of each order of intrinsic mode function, and define the remaining data after sieving as residual components until the residual components meet the preset termination condition, and define them as residual component tensors.
[0022] Step S23: Calculate the average fluctuation period and energy of each order of intrinsic mode function subtensor, and divide the order of subtensors with average fluctuation period less than a preset first threshold and energy with random distribution into high-frequency subtensor groups and perform a zeroing operation.
[0023] The orders of intrinsic mode function subtensors with a coherence greater than a preset second threshold and a deviation between the average fluctuation period and the main period of the ambient temperature less than a preset third threshold are classified into mid-frequency subtensor groups and isolated storage operations are performed.
[0024] The remaining low-frequency subtensor sets, excluding the high-frequency and mid-frequency subtensor sets, are merged with the residual component tensors to obtain the residual trend tensor.
[0025] Preferably, step S23 specifically includes:
[0026] Step S231: Calculate the energy density of each order intrinsic mode function subtensor and match it with the center frequency of each order subtensor to determine the frequency distribution characteristics of the energy concentration interval.
[0027] Step S232: Compare the average fluctuation period of each order of intrinsic mode function subtensor with a preset first threshold. If the average fluctuation period is less than the preset first threshold and the energy density is spatially random, then divide the order of intrinsic mode function subtensor into a high-frequency subtensor group and perform a filtering operation.
[0028] Step S233: Obtain ambient temperature time series data synchronized with the sensor network, calculate the ambient temperature main period of the ambient temperature time series data, and calculate the coherence between the subtensors of each order of intrinsic mode function and the ambient temperature time series data.
[0029] When the coherence is greater than the preset second threshold and the deviation between the average fluctuation period and the main period of the ambient temperature is less than the preset third threshold, the subtensor of this order is divided into the mid-frequency subtensor group.
[0030] Step S234: The intermediate frequency sub-tensor group is isolated and stored, and the remaining low frequency sub-tensor group is merged with the residual component tensor to obtain the residual trend tensor.
[0031] Preferably, step S3 specifically includes:
[0032] Step S31: Using the residual trend tensor as the objective approximation term, establish a constrained optimization objective function:
[0033] ;
[0034] in, To smooth the trend function in The values of each historical sampling time step, Let be the residual trend tensor at the t-th historical sampling time step. For smoothing coefficients, This represents the total number of historical sampling time steps. To smooth the trend function in The second derivative of each historical sampling time step with respect to time. To find, among all smooth trend functions that satisfy the shape boundary conditions, the objective function of constraint optimization is... Obtain the minimum smoothing trend function;
[0035] Step S32, apply shape boundary conditions with respect to the time derivative below the constrained optimization objective function:
[0036] ;
[0037] Step S33: Discretize and iteratively solve the constraint optimization objective function and shape boundary conditions to obtain a smooth trend function sequence. Perform first-order difference operation on the smooth trend function sequence to obtain an evolution trend index sequence.
[0038] Preferably, step S33 specifically includes:
[0039] The values of the smoothing trend function at discrete time steps are represented as a smoothing trend function vector, and the constrained optimization objective function is discretized into a matrix form of quadratic and linear terms. The quadratic form matrix and linear term vector corresponding to the smoothing regularization term are constructed using the second-order difference operator.
[0040] The shape boundary conditions are discretized into a linear inequality constraint matrix with non-negative differences between adjacent time steps;
[0041] The interior point method is used to iteratively solve the quadratic form matrix, linear term vector, and linear inequality constraint matrix, and the iterative smoothing trend function vector is updated in each iteration;
[0042] When the absolute difference between the corresponding elements of the iterative smoothing trend function vectors of adjacent iteration steps is less than the preset convergence threshold, the current iterative smoothing trend function vector is determined as the convergent smoothing trend function vector, and the smoothing trend function sequence is obtained by arranging them in the order of time steps.
[0043] The evolutionary trend index sequence is obtained by performing a first-order difference operation on the smoothed trend function sequence.
[0044] Preferably, step S4 specifically includes:
[0045] Step S41: Obtain the future weather forecast temperature time series, which includes the ambient temperature data for the future time step;
[0046] Step S42: Calculate the mean of the first-order difference at the end of the smoothed trend function sequence, use it as the baseline step size increment, and construct the state transition matrix of the first-order Kalman filter operator.
[0047] Step S43: Input the baseline step size increment into the state transition matrix of the first-order Kalman filter operator, and extrapolate the expected value of the trajectory of the smooth trend function sequence within the future time step using a forward time sliding window rolling method.
[0048] Step S44: Extract the intermediate frequency sub-tensor set, calculate the dynamic mapping coefficient between the intermediate frequency sub-tensor set and the future weather forecast temperature time series, and obtain the periodic amplitude component of the future ambient temperature.
[0049] Step S45: Linearly superimpose the trajectory expectation value and the periodic amplitude component to obtain the crack evolution state trajectory.
[0050] Preferably, step S44 specifically includes:
[0051] The intermediate frequency sub-tensor set is extracted, and the instantaneous amplitude sequence and instantaneous phase sequence of the intermediate frequency sub-tensor set are calculated using Hilbert transform;
[0052] Obtain the main period of ambient temperature corresponding to the ambient temperature time series data, and calculate the phase synchronization relationship between the intermediate frequency sub-tensor set and the ambient temperature time series.
[0053] A temperature response mapping model is constructed based on the instantaneous amplitude sequence, instantaneous phase sequence, and phase synchronization relationship;
[0054] Input the future weather forecast temperature time series into the temperature response mapping model to calculate the temperature driving amplitude corresponding to the future time step;
[0055] Based on the temperature-driven amplitude and instantaneous phase evolution law, the future periodic oscillation trajectory is reconstructed to obtain the periodic amplitude component of the future ambient temperature.
[0056] Secondly, a bridge crack data trend prediction system based on big data includes:
[0057] The data acquisition module acquires original images of the apparent damage in the cracked area and obtains long-period sampling time series. It then performs three-dimensional spatial coordinate transformation and pixel-level edge segmentation processing to reconstruct the spatiotemporal conjugate tensor.
[0058] The tensor construction module adds white noise to the spatiotemporal conjugate tensor and sieves it along the time axis. It then groups the tensors according to the average fluctuation period, energy, and coherence with the ambient temperature time series data to obtain the residual trend tensor and isolate and store the mid-frequency sub-tensor group.
[0059] The decomposition and sieving module discretizes and iteratively solves the residual trend tensor under shape boundary conditions to obtain the evolution trend index sequence.
[0060] The prediction fusion module obtains the future weather forecast temperature time series, extrapolates the trajectory expectation value of the evolution trend index sequence within the future time step, and calculates the periodic amplitude component based on the mid-frequency sub-tensor group and the future weather forecast temperature time series. The two are then linearly superimposed to output the crack evolution state trajectory.
[0061] The beneficial effects of this invention are as follows: This invention innovatively adds white noise to the spatiotemporal conjugate tensor and performs adaptive sieving along the time axis. Based on the average fluctuation period, energy, and coherence with the environmental temperature time series, it performs multi-scale fine grouping, which can not only accurately remove high-frequency texture errors and measurement noise, but also successfully achieve complete separation and adaptive decoupling of the residual trend tensor and the mid-frequency temperature response component. This completely solves the problem that the real damage evolution is easily masked by the periodic temperature variation interference in traditional methods. At the same time, when solving the evolution trend, it does not use blind mathematical fitting, but cleverly introduces shape boundary condition constraints on the time derivative (i.e., non-negative difference between adjacent time steps) in the iterative solution process. This perfectly fits the physical objective law of monotonous non-negative accumulation of structural damage. From both mathematical and engineering physics perspectives, it eliminates the false contraction and abnormal oscillation of the trend line caused by the decrease in environmental temperature or single measurement error, which greatly ensures the absolute rationality and reliability of the evolution trend index sequence over a long period. Furthermore, this invention acquires original images of the apparent defects in the cracked area and simultaneously obtains long-term sampling time series from the sensor network. It then performs three-dimensional spatial coordinate transformation and pixel-level edge segmentation, reconstructing the data into a spatiotemporal conjugate tensor. This effectively breaks through the limitations of traditional technologies that rely solely on static AI visual feature recognition, achieving deep coupling and benchmark unification between spatial morphology and long-term temporal dimensions. On the prediction side, this invention employs a forward-looking dual-track fusion strategy. On one hand, it uses the state transition matrix of the first-order Kalman filter operator to robustly extrapolate the expected trajectory value within future time steps using a forward time sliding window, locking in long-term... On the one hand, it utilizes the inertia of damage evolution, and on the other hand, it extracts the instantaneous amplitude and phase of the mid-frequency sub-tensor using Hilbert transform to construct a high-precision temperature response mapping model. Based on the future weather forecast temperature time series, it dynamically reconstructs the future periodic amplitude components. By linearly superimposing the trajectory expectation value and the periodic amplitude components, it outputs the final trajectory of bridge crack evolution. This makes the prediction results include both the long-term monotonic damage accumulation trend and perfectly take into account the future short-term seasonal temperature fluctuations. This provides a core basis with extremely high scientific quantitative credibility for the long-term health status assessment, proactive preventive maintenance, and inspection decisions of bridges. Attached Figure Description
[0062] Figure 1 A flowchart illustrating the steps of a bridge crack data trend prediction method based on big data, provided in one embodiment of the present invention;
[0063] Figure 2 This is a basic schematic diagram of a bridge crack data trend prediction system based on big data, provided as an embodiment of the present invention. Detailed Implementation
[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0065] Example 1, referring to Figure 1 This paper provides a method for predicting bridge crack trends based on big data, including the following steps:
[0066] Step S1: Collect the original image of the apparent disease in the cracked area and obtain the long-period sampling time series. Perform three-dimensional spatial coordinate transformation and pixel-level edge segmentation processing, and reconstruct to obtain the spatiotemporal conjugate tensor.
[0067] Step S2: Add white noise signal to the spatiotemporal conjugate tensor and sieve it along the time axis. Group it according to the average fluctuation period, energy and coherence with the ambient temperature time series data to obtain the residual trend tensor and store the mid-frequency sub-tensor group in isolation.
[0068] Step S3: Discretize and iteratively solve the residual trend tensor under shape boundary conditions to obtain the evolution trend index sequence;
[0069] Step S4: Obtain the future weather forecast temperature time series, extrapolate the trajectory expectation value of the evolution trend index sequence within the future time step, calculate the periodic amplitude component based on the mid-frequency sub-tensor group and the future weather forecast temperature time series, and perform linear superposition to output the crack evolution state trajectory.
[0070] In specific implementation, step S1 includes:
[0071] Step S11: Use the drone camera to collect the original images of the apparent damage in the cracked area, and simultaneously acquire the long-period sampling time series corresponding to the sensor network;
[0072] Step S12: Deblur the original image of the apparent defects to obtain a high-quality image of the bridge crack.
[0073] Step S13: Use the aerial triangulation algorithm to perform full-field collinearity equation adjustment calculation on high-quality bridge crack images, calculate the spatial internal and external parameters of the UAV camera and generate dense point cloud data.
[0074] Acquire laser point cloud data and register it with dense point cloud data to obtain the laser point cloud spatial coordinate transformation matrix.
[0075] Specifically, each frame of the original image of the apparent disease captured by the drone camera records the timestamp of the acquisition time, and the long-cycle sampling time series also records the timestamp of the acquisition time. Both are stored together with the same crack object identifier. To ensure that subsequent processing is aligned at the same time step, the long-cycle sampling time series is resampled according to the drone camera acquisition time. Missing points are filled in using linear interpolation. If high-frequency jitter exists, a moving average filter is used for smoothing. The sliding window length is set to 6 sampling time steps. If the window is too small, the noise reduction will be insufficient; if the window is too large, it will weaken the true fluctuation and affect the period discrimination of subsequent screening and decomposition.
[0076] Deblurring is achieved using a multi-scale encoder-decoder structure with three scales. The encoder employs four levels of downsampling, each containing two 3x3 convolutional layers connected sequentially with batch normalization and linear rectified activation functions. The channel counts are 64, 128, 256, and 512, respectively. The bottleneck layer contains six residual blocks, each with two 3x3 convolutional layers connected using identity shorting. The decoder performs four levels of upsampling, skipping connections with the encoder before passing through two 3x3 convolutional layers to recover details. The output performs pixel-wise weighted fusion of the three scale branches, with a weight sum of 1. The weights represent the contribution ratio of different scales to the restoration result, balancing fine crack edges with large-scale motion blur suppression. The deblurring model is trained using pairs of sharp and blurred images. The blurred image is generated by convolving the sharp image with a motion blur kernel and adding Gaussian noise. The optimal length of the motion blur kernel is 5 to 30 pixels to cover common drone jitter amplitudes. The training loss function is a weighted sum of pixel absolute error loss and structural similarity loss. The structural similarity loss weight is preferably between 0.1 and 0.5. Too small a weight will cause blurring of crack edge details, while too large a weight will cause an overall brightness contrast shift. The optimization method uses an adaptive moment estimation optimizer with an initial learning rate of 5×10. -4 The batch size is set to 10, and the training epochs are 100. A learning rate that is too large can easily produce restoration artifacts, while a learning rate that is too small will result in slow convergence and insufficient recovery of fine textures. The weight parameters are shown in Table 1.
[0077] Table 1 Weight Name Value Weight meaning reason Multi-scale output fusion weights original resolution branch 0.4 Control the contribution ratio of the original resolution branch to the final deblurred output. The original resolution branch is more sensitive to fine crack edges and texture details. A lower weight will cause blurry edges, while an excessively high weight will amplify local noise. Multi-scale output fusion weights half-resolution branch 0.35 Controlling the contribution ratio of the half-resolution branch to the final deblurred output This branch strikes a trade-off between detail and global deblurring; too low a weight weakens mesoscale motion blur suppression, while too high a weight loses fine edges. Multi-scale output fusion weights quarter-resolution branch 0.25 Controlling the contribution ratio of the quarter-resolution branch to the final deblurred output The low-resolution branch is better at suppressing large-scale motion blur and ghosting. Excessive weighting can cause overly smooth edges in fine cracks, while insufficient weighting can leave large-scale blur. Structural similarity loss weights 0.2 The relative contribution of balancing pixel absolute error loss and structural similarity loss If the structural similarity loss weight is too small, it will result in insufficient recovery of crack edges and texture details; if it is too large, it will easily lead to brightness contrast shift and structural artifacts.
[0078] Scale-invariant feature points are extracted from high-quality bridge crack images from multiple perspectives and feature matching is performed. The number of feature points per image is set to 2000. Too few feature points will lead to insufficient geometric constraints, while too many will increase the risk of mismatches and increase computational cost. After matching, outliers are removed using a random sampling consensus method, and the geometric model is estimated. The threshold for intra-pixel points is set to 2. Too small a threshold will easily delete true matches, while too large a threshold will retain outliers, leading to unstable adjustment. Subsequently, a least squares adjustment model is established based on the full-field collinearity equation, and the spatial intrinsic and extrinsic parameters are iteratively solved. The spatial intrinsic parameters include at least the focal length, principal point coordinates, radial distortion coefficient, and tangential distortion coefficient. The spatial extrinsic parameters include the camera's position and attitude angle in a unified coordinate system. The preferred termination condition for adjustment iteration is that the root mean square error of reprojection is less than 1 pixel or the number of iterations reaches 50. Too large a root mean square error of reprojection will lead to unstable dense matching depth. After solving the spatial parameters, dense point cloud data is generated by multi-view dense matching. The density of the dense point cloud is 10,000 points per square meter. Too low a density will reduce the surface details and thus affect the pixel projection accuracy, while too high a density will significantly increase the storage and registration overhead.
[0079] Voxel downsampling was performed on both laser point cloud data and dense point cloud data to unify point density. The voxel side length was set to 20 mm. Points falling within the same voxel were replaced by their geometric center point, thus reducing the number of points while maintaining the overall consistency of the bridge's geometric surface shape. Coarse registration was then performed to obtain the initial pose. Coarse registration was achieved using 3D keypoint feature matching. Specifically, spatially uniform 3D keypoints were selected from both types of point clouds. For each 3D keypoint, a feature vector was constructed in its neighborhood, and nearest neighbor matching was performed. Based on the matched point pairs, the initial rotation and translation components were calculated using random sampling consistency to obtain the initial pose. Fine registration was then performed, using an iterative nearest-neighbor algorithm to minimize the Euclidean distance residual between corresponding points. The nearest-neighbor search radius of the iterative nearest-neighbor algorithm was set to 150 mm, the maximum number of iterations was set to 60, and the convergence condition was that the mean square error change between two adjacent iterations was less than 1 × 10⁻⁶. -6 In each iteration, a corresponding point set is established based on nearest neighbor search, and the rotation and translation components are updated until the convergence condition is met or the maximum number of iterations is reached. After registration, the laser point cloud spatial coordinate transformation matrix is output. This matrix is set as a 4x4 homogeneous transformation matrix, with the upper left 3×3 representing the rotation component, the upper right 3×1 representing the translation component, and the last row being 0, 0, 0, 1. After representing any point in the laser point cloud data as a homogeneous coordinate column vector, multiplying it with the laser point cloud spatial coordinate transformation matrix yields the coordinates of that point in the unified coordinate system of the dense point cloud data, thus providing a consistent spatial reference for subsequently projecting the two-dimensional pixel coordinates onto the bridge's spatial geometric surface.
[0080] In specific implementation, step S1 also includes:
[0081] Step S14: Use a computer vision network to perform pixel-level edge segmentation on the high-quality bridge crack image, extract the maximum crack width, physical length of the crack skeleton line and pixel-level apparent area of the crack region in the current iteration step, and mark the two-dimensional pixel coordinates.
[0082] Step S15: Using the laser point cloud spatial coordinate transformation matrix, the two-dimensional pixel coordinates are projected onto the spatial geometric surface of the bridge to calculate the spatial three-dimensional projection curvature.
[0083] Step S16: The maximum crack width, the physical length of the crack skeleton line, the pixel-level apparent area of the crack region, and the spatial three-dimensional projection curvature are tensorized and recombined according to a preset arrangement order to obtain the spatiotemporal conjugate tensor.
[0084] Specifically, a computer vision network is used to perform pixel-level edge segmentation on high-quality bridge crack images, outputting a binary segmentation result of the crack region. The maximum crack width, the physical length of the crack skeleton line, and the pixel-level apparent area of the crack region are extracted from this binary segmentation result, and the two-dimensional pixel coordinates are labeled. The computer vision network in this embodiment adopts an encoder-decoder structure. The encoder end has 4 levels of downsampling, and the decoder end has 4 levels of upsampling. Each level of the encoder end contains two 3×3 convolutional layers, followed by batch normalization and linear rectified activation functions. The number of channels at the four levels is set to 64, 128, 256, and 512 respectively, with the bottom layer having 1024 channels. Each level of the decoder end first performs a 2x upsampling, then performs a skip connection with the corresponding encoder end features, and then recovers edge details through two 3×3 convolutional layers. The output end obtains a crack probability map through a 1×1 convolution and maps it to 0 to 1 using the Sigmoid function. In this embodiment, the crack probability map is binarized using a threshold of 0.5 to obtain a crack region mask. Morphological opening and closing operations are then performed for denoising. The opening structuring element is set to 3×3, and the closing structuring element is set to 5×5. Connected components with an area less than 100 pixels are removed to suppress isolated noise. During computer vision network training, manually labeled crack masks are used as supervision signals. The training loss function consists of the linear sum of a binary cross-entropy loss and a Dice loss, both with weights set to 1. The learning rate of the adaptive moment estimation optimizer is set to 1×10⁻⁶. -4The batch size was set to 8, the training epochs to 80, and the input image size to a uniform 512×512. Training data augmentation employed brightness and contrast perturbation and affine transformation to adapt to different lighting conditions and shooting angles. The maximum crack width was obtained on the binary mask through distance transformation. The maximum value of the distance transformation was multiplied by 2 to obtain the pixel dimension value of the maximum crack width, which was then converted to a physical dimension value using the pixel-to-physical scale conversion coefficient obtained in step S15. The physical length of the crack skeleton line was obtained by refining the binary mask to obtain the crack skeleton line pixel coordinate sequence, and then the Euclidean distance between adjacent skeleton pixels projected in step S15 was accumulated segment by segment. The pixel-level apparent area of the crack region was obtained by the number of pixels with a value of 1 in the binary mask. This number of pixels was stored together with the two-dimensional pixel coordinates, which included the crack skeleton line pixel coordinate set and the crack edge pixel coordinate set, used for subsequent three-dimensional spatial coordinate transformation and curvature calculation.
[0085] The laser point cloud data is transformed to the unified coordinate system of the dense point cloud data using a laser point cloud spatial coordinate transformation matrix. The bridge's spatial geometric surface is then reconstructed from the fused point cloud. The bridge's spatial geometric surface is constructed using a triangular mesh, which is obtained by triangulating the local neighborhood of the point cloud. Subsequently, using the UAV camera's spatial and intrinsic parameters obtained in step S13, a corresponding spatial ray is generated for each two-dimensional pixel coordinate. The starting point of the spatial ray is the camera's position in the unified coordinate system, and the direction is obtained by back-projecting the pixel point through the spatial intrinsic parameters and combining it with the camera's attitude transformation. The spatial ray is intersected with the triangular mesh of the bridge's spatial geometric surface. The intersection point closest to the camera is taken as the three-dimensional projection point corresponding to the two-dimensional pixel coordinate. If the ray intersection fails, nearest neighbor search is used to find the nearest point in the point cloud near the ray as the three-dimensional projection point. The spatial three-dimensional projection curvature is calculated by analyzing the sequence of three-dimensional projection points of the crack skeleton line. In this embodiment, each skeleton three-dimensional projection point is used as the center, and a neighborhood point set with a radius of 20 mm is searched in the point cloud. When the number of neighborhood points is less than 30, the search radius is increased to 50 mm. A quadratic surface is fitted to the neighborhood point set in the local coordinate system, and the average curvature at the skeleton point is calculated based on the quadratic surface parameters. The arithmetic mean of the absolute values of the average curvature at all skeleton points is taken as the spatial three-dimensional projection curvature at that time step, thus enabling the spatial three-dimensional projection curvature to have a stable characterization ability for local concave and convex disturbances.
[0086] The preset arrangement order is: maximum crack width, physical length of crack skeleton line, pixel-level apparent area of crack region, and spatial three-dimensional projection curvature. A feature vector is generated for each historical sampling time step. The four components of the feature vector correspond to the four quantities mentioned above. The feature vectors of all historical sampling time steps are stacked in chronological order to obtain a spatiotemporal conjugate tensor. The time dimension of the spatiotemporal conjugate tensor corresponds to the time step of the long-period sampling time series, and the channel dimension corresponds to the four types of crack evolution features. This provides a unified input data structure for the subsequent step S2, which involves sieving and decomposing along the time axis, and ensures that time series decomposition, trend solving, and temperature response reconstruction are performed on the same object.
[0087] In specific implementation, step S2 includes:
[0088] Step S21: Add M-order white noise signals to the spatiotemporal conjugate tensor to obtain the corresponding auxiliary mask signal matrix;
[0089] Step S22: Use the time-domain adaptive sieving operator to sieve the auxiliary mask signal matrix along the time axis, extract the sub-tensors of each order of intrinsic mode function, and define the remaining data after sieving as residual components until the residual components meet the preset termination condition, and define them as residual component tensors.
[0090] Step S23: Calculate the average fluctuation period and energy of each order of intrinsic mode function subtensor, and divide the order of subtensors with average fluctuation period less than a preset first threshold and energy with random distribution into high-frequency subtensor groups and perform a zeroing operation.
[0091] The orders of intrinsic mode function subtensors with a coherence greater than a preset second threshold and a deviation between the average fluctuation period and the main period of the ambient temperature less than a preset third threshold are classified into mid-frequency subtensor groups and isolated storage operations are performed.
[0092] The remaining low-frequency subtensor sets, excluding the high-frequency and mid-frequency subtensor sets, are merged with the residual component tensors to obtain the residual trend tensor.
[0093] Specifically, an auxiliary mask signal matrix is obtained by adding M-fold white noise signals to the spatiotemporal conjugate tensor, where M is set to 50. When adding white noise, a white noise sequence with a mean of 0 is superimposed on each sequence along the time axis of the spatiotemporal conjugate tensor. The amplitude of the white noise is scaled according to the standard deviation of the sequence, with a scaling factor set to 0.20. After 50 superpositions, the 50 results are stacked according to the noise index direction to form the auxiliary mask signal matrix, which can be used for subsequent sieving along the time axis.
[0094] A time-domain adaptive sieving operator is used to sieve the auxiliary mask signal matrix along the time axis, extracting intrinsic mode function (IMF) subtensors of each order. The remaining data after sieving is defined as residual components until a preset termination condition is met, at which point the residual component tensor is defined as a residual component tensor. The time-domain adaptive sieving operator is implemented using an envelope mean iteration method. Specifically, for each time series, local maxima and local minima are found. Cubic spline interpolation is used to obtain the upper and lower envelopes, respectively. The mean of these two values is taken and subtracted from the original series to obtain candidate components. This process is repeated until the candidate components satisfy the IMF conditions. The IMF conditions are set as follows: the difference between the number of local extrema and the number of zero-crossing points does not exceed 1, and the mean square value of the upper and lower envelope mean curves is not greater than 0.10 of the mean square value of the candidate component. Each IMF component of order 1 is extracted from the original series, and the remaining part is further sieved. The preset termination condition is set as the number of local extrema of the residual component on the entire time axis is less than 2. Since 50 instances of white noise were added in step S21, in this embodiment, the arithmetic mean of the 50 results for the same order intrinsic mode function components is taken over time steps to obtain the subtensor of the intrinsic mode function of that order. The arithmetic mean of the final residual components is also taken to obtain the residual component tensor.
[0095] In specific implementation, step S23 includes:
[0096] Step S231: Calculate the energy density of each order intrinsic mode function subtensor and match it with the center frequency of each order subtensor to determine the frequency distribution characteristics of the energy concentration interval.
[0097] Step S232: Compare the average fluctuation period of each order of intrinsic mode function subtensor with a preset first threshold. If the average fluctuation period is less than the preset first threshold and the energy density is spatially random, then divide the order of intrinsic mode function subtensor into a high-frequency subtensor group and perform a filtering operation.
[0098] Step S233: Obtain ambient temperature time series data synchronized with the sensor network, calculate the ambient temperature main period of the ambient temperature time series data, and calculate the coherence between the subtensors of each order of intrinsic mode function and the ambient temperature time series data.
[0099] When the coherence is greater than the preset second threshold and the deviation between the average fluctuation period and the main period of the ambient temperature is less than the preset third threshold, the subtensor of this order is divided into the mid-frequency subtensor group.
[0100] Step S234: The intermediate frequency sub-tensor group is isolated and stored, and the remaining low frequency sub-tensor group is merged with the residual component tensor to obtain the residual trend tensor.
[0101] Specifically, the energy density is calculated by squaring and summing the values of a certain order intrinsic mode function subtensor across all historical sampling time steps, all spatial locations, and all feature channels, and then dividing by the product of the number of historical sampling time steps, the number of spatial locations, and the number of feature channels to obtain the unit sample average energy of that order intrinsic mode function subtensor. The center frequency is calculated by performing a discrete Fourier transform on the time-axis sequence of the order intrinsic mode function subtensor to obtain the power spectral density. The center frequency is taken as the weighted average frequency of the power spectral density, ensuring that the center frequency characterizes the main oscillation frequency band of the order intrinsic mode function subtensor. A one-to-one correspondence is established between the energy density and center frequency of each order intrinsic mode function subtensor according to its order, and the frequencies are sorted from high to low according to the center frequency to form a frequency distribution sequence, thus obtaining the frequency distribution characteristics of the energy concentration interval.
[0102] The average fluctuation period of each order of intrinsic mode function subtensor is compared with a preset first threshold, and the high-frequency subtensor group is determined and filtered out based on the spatial random distribution of energy density. The average fluctuation period is calculated by statistically analyzing the time step intervals of adjacent local maxima points along the time axis sequence of the order of intrinsic mode function subtensor and taking the arithmetic mean. The unit of the average fluctuation period is time steps. In this embodiment, the preset first threshold is set to 3 time steps. The determination method for spatial random distribution of energy density is to first statistically analyze the coefficient of variation of the energy density in the spatial position dimension, and then calculate the absolute value of the correlation coefficient of energy density at adjacent spatial positions. In this embodiment, the threshold for the coefficient of variation is set to 1.0, and the threshold for the absolute value of the correlation coefficient is set to 0.10. When the coefficient of variation is greater than 1.0 and the absolute value of the correlation coefficient is less than 0.10, it is determined to be spatially random distribution. When the average fluctuation period is less than 3 time steps and the energy density is spatially random, the intrinsic mode function subtensor of this order is divided into a high-frequency subtensor group and a filtering operation is performed. In this embodiment, the filtering operation is implemented by clearing to zero, that is, the intrinsic mode function subtensor of this order is set to zero in the subsequent reconstruction calculation to suppress the influence of measurement noise and local texture error on the trend term.
[0103] Ambient temperature time-series data synchronized with the sensor network is acquired. The dominant period of the ambient temperature time-series data is calculated, and the coherence between the intrinsic mode function subtensors of each order and the ambient temperature time-series data is calculated to determine the mid-frequency subtensor group. In this embodiment, the dominant period of the ambient temperature time-series data is calculated using the discrete Fourier transform method. Specifically, the amplitude spectrum of the ambient temperature time-series data is calculated, and the period corresponding to the spectral peak with the largest amplitude (excluding zero frequency) is selected as the dominant period. The coherence is calculated using the squared amplitude coherence in this embodiment. Specifically, the time series corresponding to the intrinsic mode function subtensors and the ambient temperature time-series data are subjected to segmented windowed spectral estimation, the cross-power spectrum and the self-power spectrum are calculated, and then the squared amplitude coherence value is calculated at the frequency corresponding to the dominant period of the ambient temperature as the coherence. A second preset threshold is set to 0.75 in this embodiment. The deviation between the average fluctuation period and the dominant period of the ambient temperature is calculated using the relative deviation, which is calculated by dividing the absolute value of the difference between the two by the dominant period of the ambient temperature. A third preset threshold is set to 0.20 in this embodiment. When the coherence is greater than 0.75 and the relative deviation is less than 0.20, the subtensor of the intrinsic mode function of this order is divided into a mid-frequency subtensor group.
[0104] In this embodiment, the intermediate frequency sub-tensor set is isolated and stored using the crack object identifier and the intrinsic mode order as index keys. It is written to a database or file system, and simultaneously records the center frequency, average fluctuation period, energy density, coherence, and ambient temperature principal period of the intrinsic mode function sub-tensor of that order. This information is used in subsequent step S4 to calculate the period amplitude component. The merging method involves taking the intrinsic mode function sub-tensors that are not classified into high-frequency sub-tensor sets and intermediate-frequency sub-tensor sets as the remaining low-frequency sub-tensor sets. The remaining low-frequency sub-tensor sets are numerically added point-by-point along the time step, and then added point-by-point with the residual component tensor to obtain the residual trend tensor. This residual trend tensor retains the low-frequency slowly varying terms and residual components while removing high-frequency noise terms. Simultaneously, the intermediate-frequency fluctuation term, which is strongly correlated with ambient temperature, is retained separately for subsequent temperature-driven oscillation reconstruction.
[0105] In specific implementation, step S3 includes:
[0106] Step S31: Using the residual trend tensor as the objective approximation term, establish a constrained optimization objective function:
[0107] ;
[0108] in, To smooth the trend function in The values of each historical sampling time step, Let be the residual trend tensor at the t-th historical sampling time step. For smoothing coefficients, This represents the total number of historical sampling time steps. To smooth the trend function in The second derivative of each historical sampling time step with respect to time. To find, among all smooth trend functions that satisfy the shape boundary conditions, the objective function of constraint optimization is... Obtain the minimum smoothing trend function;
[0109] Step S32, apply shape boundary conditions with respect to the time derivative below the constrained optimization objective function:
[0110] ;
[0111] Step S33: Discretize and iteratively solve the constraint optimization objective function and shape boundary conditions to obtain a smooth trend function sequence. Perform first-order difference operation on the smooth trend function sequence to obtain an evolution trend index sequence.
[0112] Specifically, the residual trend tensor is used as the objective approximation term to establish a constrained optimization objective function. Since the residual trend tensor is a tensor sequence that varies with the total number of historical sampling time steps, the tensor slices of the residual trend tensor at each historical sampling time step are standardized into a residual trend sequence, which is then used to fit a one-to-one correspondence with the smoothed trend function on the time axis. Let the residual trend tensor at the i-th... Tensor slices for each historical sampling time step are Its spatial index With channel index The element at is The total number of spatial indexes is If the total number of channels is 4, then the residual trend sequence is defined as follows:
[0113] ;
[0114] in, For the first The residual trend scalar value at each historical sampling time step Take 1 to , This represents the total number of historical sampling time steps.
[0115] Using the residual trend tensor as the objective approximation term, a constrained optimization objective function is established. In this embodiment, the smoothing coefficient is set to 10 to achieve a balance between conforming to the residual trend sequence and suppressing local noise fluctuations. In this embodiment, the residual trend tensor represents the long-term trend of bridge cracks after removing high-frequency noise and periodic fluctuations in ambient temperature. Considering that bridge cracks generally remain stable or continuously expand during long-term evolution without human intervention, a non-negative shape boundary condition with respect to the time derivative is applied to the constrained optimization objective function to ensure that the smoothing trend function conforms to the long-term evolution trend of the cracks.
[0116] Using the discrete second-order difference approximation, it can be expressed as:
[0117] ;
[0118] in, To smooth the trend function in The values of each historical sampling time step, To smooth the trend function in The values of each historical sampling time step, The value range is 2 to .
[0119] In specific implementation, step S33 includes:
[0120] The values of the smoothing trend function at discrete time steps are represented as a smoothing trend function vector, and the constrained optimization objective function is discretized into a matrix form of quadratic and linear terms. The quadratic form matrix and linear term vector corresponding to the smoothing regularization term are constructed using the second-order difference operator.
[0121] The shape boundary conditions are discretized into a linear inequality constraint matrix with non-negative differences between adjacent time steps;
[0122] The interior point method is used to iteratively solve the quadratic form matrix, linear term vector, and linear inequality constraint matrix, and the iterative smoothing trend function vector is updated in each iteration;
[0123] When the absolute difference between the corresponding elements of the iterative smoothing trend function vectors of adjacent iteration steps is less than the preset convergence threshold, the current iterative smoothing trend function vector is determined as the convergent smoothing trend function vector, and the smoothing trend function sequence is obtained by arranging them in the order of time steps.
[0124] The evolutionary trend index sequence is obtained by performing a first-order difference operation on the smoothed trend function sequence.
[0125] Specifically, the values of the smoothing trend function at discrete time steps are arranged in time step order to form a smoothing trend function vector, which is denoted as: ;
[0126] in, For transpose, For smoothing trend function vectors;
[0127] The residual trend sequence obtained by scaling the residual trend tensor is arranged into a residual trend vector according to the time step order. The residual trend vector is denoted as: ;
[0128] in, This represents the residual trend vector;
[0129] The second-order difference operator matrix is denoted as... Its dimensions are OK Column, number Only in the first row Column, No. +1 column, number The +2 column takes non-zero values, which are 1, -2, and 1 respectively, thus satisfying the following:
[0130] ;
[0131] in, The result vector is obtained by taking the second difference of the smoothed trend function vector along the time axis of the total historical sampling time steps;
[0132] Discretize the constrained optimization objective function into matrix form:
[0133] ;
[0134] in, It is the element-wise difference vector between the smoothed trend function vector and the residual trend vector at the same time step;
[0135] Expanding it, we obtain the matrix form of the quadratic and linear terms:
[0136] ;
[0137] in, It is a quadratic matrix. For a linear term vector, For the constant term, the quadratic form matrix and the linear term vector of the constraint optimization objective function in this embodiment are taken as:
[0138] ;
[0139] ;
[0140] in, for An identity matrix of order 1. For smoothing coefficients, This is the transpose of the second-order difference operator matrix;
[0141] In this embodiment, the shape boundary conditions are discretized into a linear inequality constraint matrix of non-negative differences between adjacent time steps. A first-order difference operator matrix is constructed and denoted as... Its dimensions are OK Column, number line in Take -1 in the column, at the 1st... Take 1 from column +1 to satisfy:
[0142] ;
[0143] in, This is the result vector obtained after applying the first-order difference operator matrix to the smoothing trend function vector;
[0144] The shape boundary conditions, after discretization, are written as:
[0145] Where 0 represents the dimension. The zero vector, which is a constraint that indicates that the difference between adjacent time steps is not negative.
[0146] The interior-point method is implemented by adding a logarithmic obstacle term to the objective function, allowing the constraints to pass through the obstacle function into an unconstrained optimization form. The obstacle objective function is constructed as follows:
[0147] ;
[0148] in, For obstacle parameters, for The One element, This is the transpose of the smoothing trend function vector. The objective function is the obstacle; to ensure that the logarithmic terms are defined, the logarithmic term is maintained during the iteration process. All elements are positive. In this embodiment, an initial barrier parameter is used. In this embodiment, the initial barrier parameter is 1, and it decays proportionally after each round of Newton updates, with a decay coefficient of 0.1. The maximum number of iterations is 100. The initial iterative smoothing trend function vector is... And through cumulative maximum value correction All elements are positive, and the correction value is 1×10. -6 Multiply by the time step number to ensure that the initial value meets the feasibility requirements of the interior point method.
[0149] In this embodiment, the preset convergence threshold is 1×10. -6 and with As a basis for convergence, among which... For the first The iteration vector of the nth iteration vector One element, The preset convergence threshold is used;
[0150] The evolutionary trend index sequence is obtained by performing a first-order difference operation on the smoothed trend function sequence. The evolutionary trend index sequence is obtained at historical sampling time steps. The value of is denoted as: ;
[0151] in, For the evolutionary trend index sequence in the 1st The values taken at each time step are used to characterize the magnitude of change in the smoothing trend function between adjacent historical sampling time steps, and serve as an indicator for analyzing the long-term evolution rate of cracks. Take 2 to .
[0152] In specific implementation, step S4 includes:
[0153] Step S41: Obtain the future weather forecast temperature time series, which includes the ambient temperature data for the future time step;
[0154] Step S42: Calculate the mean of the first-order difference at the end of the smoothed trend function sequence, use it as the baseline step size increment, and construct the state transition matrix of the first-order Kalman filter operator.
[0155] Step S43: Input the baseline step size increment into the state transition matrix of the first-order Kalman filter operator, and extrapolate the expected value of the trajectory of the smooth trend function sequence within the future time step using a forward time sliding window rolling method.
[0156] Step S44: Extract the intermediate frequency sub-tensor set, calculate the dynamic mapping coefficient between the intermediate frequency sub-tensor set and the future weather forecast temperature time series, and obtain the periodic amplitude component of the future ambient temperature.
[0157] Step S45: Linearly superimpose the trajectory expectation value and the periodic amplitude component to obtain the crack evolution state trajectory.
[0158] Specifically, the smoothed trend function sequence is arranged in order of the total number of historical sampling time steps, and the last first-order difference is the change in the smoothed trend function value of adjacent historical time steps. The mean of the last first-order difference is calculated by taking the arithmetic mean of the adjacent differences of the last 5 consecutive time steps of the smoothed trend function sequence, and using this mean as the baseline step size increment. The baseline step size increment is used to characterize the average rate of change of the smoothed trend function sequence at the end of the history, so that the extrapolation process can still be calculated according to a uniform progression law even without future observations. The state of the first-order Kalman filter operator uses two components: the first component is the smoothed trend function value, and the second component is the rate of change of the smoothed trend function. A constant velocity assumption is used to construct the state transition matrix, such that the smoothed trend function value of the next time step is equal to the smoothed trend function value of the current time step plus the rate of change of the current time step, and the rate of change of the next time step is equal to the rate of change of the current time step. The corresponding state transition matrix is a 2x2 matrix, with the first row containing 1s and 1s, and the second row containing 0s and 1s, so that the state update relationship can be directly used for time-step extrapolation.
[0159] Starting with the historical final time step, the first component of the state of the first-order Kalman filter operator is initialized to the smoothed trend function value corresponding to the historical final time step, and the second component is initialized to the baseline step size increment. Subsequently, state transition prediction updates are performed sequentially according to future time steps. After each future time step prediction is completed, the predicted state is used as the input for the next future time step for extrapolation, thus forming a sequence of trajectory expectation values within the future time steps. The trajectory expectation value represents the predicted value of the smoothed trend function corresponding to each future time step. This trajectory expectation value sequence maintains consistency with the future weather forecast temperature time series in terms of time steps, ensuring that step S45 can be superimposed at the same time step.
[0160] For each future time step, the expected value of the trajectory at that time step is directly added to the periodic amplitude component to obtain the fracture evolution state value at that time step. These values are then arranged in the order of future time steps to obtain the fracture evolution state trajectory. Specifically, the linear superposition involves the expected value representing the low-frequency trend term extrapolated from the smoothed trend function sequence, and the periodic amplitude component representing the temperature-driven periodic term obtained from the mid-frequency subtensor set and the future weather forecast temperature series. Both have consistent physical dimensions, and their superposition simultaneously includes the gradually accumulating trend of the fracture and the temperature-related periodic fluctuations, thus outputting the fracture evolution state trajectory within the future time steps.
[0161] The fracture evolution trajectory is composed of the fracture evolution state values at each future time step, arranged sequentially. These values are obtained by linearly superimposing the trajectory expectation value and the periodic amplitude component at the same future time step. In other words, the fracture evolution trajectory contains two parts of information: the first is the long-term, gradually changing trend extrapolated from the evolution trend index sequence; the second is the temperature-driven periodic fluctuation correction obtained from the mid-frequency subtensor set and future weather forecast temperature time series. The sequence obtained by adding these two parts at each future time step is the fracture evolution trajectory.
[0162] Each element of the crack evolution trajectory corresponds to a prediction result for a future time step, and its value reflects the strength of the crack evolution at that time step. The crack evolution trajectory is used to indicate the future evolution trend of the crack over a period of time, such as continuous growth, slowed growth, or regular fluctuations accompanied by temperature cycles, thus providing a basis for subsequent threshold alarms, maintenance decisions, or inspection plans.
[0163] In specific implementation, step S44 includes:
[0164] The intermediate frequency sub-tensor set is extracted, and the instantaneous amplitude sequence and instantaneous phase sequence of the intermediate frequency sub-tensor set are calculated using Hilbert transform;
[0165] Obtain the main period of ambient temperature corresponding to the ambient temperature time series data, and calculate the phase synchronization relationship between the intermediate frequency sub-tensor set and the ambient temperature time series.
[0166] A temperature response mapping model is constructed based on the instantaneous amplitude sequence, instantaneous phase sequence, and phase synchronization relationship;
[0167] Input the future weather forecast temperature time series into the temperature response mapping model to calculate the temperature driving amplitude corresponding to the future time step;
[0168] Based on the temperature-driven amplitude and instantaneous phase evolution law, the future periodic oscillation trajectory is reconstructed to obtain the periodic amplitude component of the future ambient temperature.
[0169] Specifically, the subtensor of each intrinsic mode function is transformed into a one-dimensional intermediate frequency (IF) time series at each historical time step. The transformation is performed by taking the arithmetic mean of the spatial and channel dimensions within that historical time step to obtain the IF time series of that order. This IF time series is aligned with the ambient temperature time series data at the same historical time step. The instantaneous amplitude and instantaneous phase sequences are calculated using a Hilbert transform for each IF time series. Let this IF time series be... The result of the Hilbert transform is Then the instantaneous amplitude sequence With instantaneous phase sequence The calculation formula is:
[0170] ;
[0171] ;
[0172] in, The four-quadrant arctangent operation is used to ensure the correct phase direction. The instantaneous amplitude sequence is used to characterize the amplitude strength of the intermediate frequency fluctuation of this order at each historical time step, and the instantaneous phase sequence is used to characterize the phase position of the intermediate frequency fluctuation of this order at each historical time step.
[0173] The dominant period of ambient temperature is calculated using the method of corresponding periods to the main peaks of the spectrum. Frequency domain analysis is performed on the ambient temperature time series data, and the period corresponding to the spectral peak with the largest amplitude (excluding the DC component) is selected as the dominant period of ambient temperature. Subsequently, a Hilbert transform is performed on the ambient temperature time series data to obtain the instantaneous phase sequence of the ambient temperature time series, and the phase synchronization relationship is calculated with the instantaneous phase sequence of the intermediate frequency subtensor set. The calculation formula is as follows:
[0174] ;
[0175] in, For phase synchronization, The value is an imaginary unit. The phase synchronization relationship is used to characterize the degree of synchronization between this order of intermediate frequency fluctuations and the ambient temperature time series; the larger the value, the stronger the synchronization.
[0176] Using the temperature deviation of ambient temperature time-series data relative to the historical mean as the driving input and the instantaneous amplitude sequence as the response output, a linear mapping is used to obtain dynamic mapping coefficients, and the phase synchronization relationship is used as the basis for the reliability of this order of mapping. The core relationship of the temperature response mapping model is:
[0177] ;
[0178] in, Instantaneous amplitude, The first dynamic mapping coefficient, The second dynamic mapping coefficient, For ambient temperature time series data at historical sampling time steps The value of , This represents the average value of the ambient temperature time series data over historical sampling time steps;
[0179] Input the future weather forecast temperature time series into the temperature response mapping model to calculate the temperature driving amplitude corresponding to the future time step. Assume the future weather forecast temperature time series is at the future time step... The value is Then the temperature-driven amplitude The calculation formula is:
[0180] ;
[0181] Using the instantaneous phase at the end of the historical period as the initial value, and advancing to future time steps according to the phase increment corresponding to the principal period of the ambient temperature. Let the principal period of the ambient temperature be... Then the phase increment is: ;
[0182] in, For phase increment, Pi;
[0183] Let the historical end phase be The future phase is: ;
[0184] in, For future phases;
[0185] And reconstruct the future periodic oscillation trajectory using temperature-driven amplitude and future phase. The calculation formula is as follows:
[0186] ;
[0187] When the intermediate frequency subtensor set contains subtensors of multiple orders of intrinsic mode functions, this embodiment uses weighted superposition of the future periodic oscillation trajectories of each order to obtain the periodic amplitude component. Let the first... The future cyclical oscillation trajectory corresponding to the order is The energy density of this order is Energy weight is ,but:
[0188] ;
[0189] ;
[0190] in, Let be the set of orders contained in the mid-frequency subtensor set. This represents the periodic amplitude component of the future ambient temperature corresponding to a future time step. set of order The Middle Energy density corresponding to the subtensor of the first-order intrinsic mode function.
[0191] The purpose of the temperature response mapping model is to establish the driving relationship between environmental temperature time series data and the instantaneous amplitude sequence corresponding to the mid-frequency sub-tensor set. This allows the model to output the temperature-driven amplitude at the future time step after inputting the future weather forecast temperature time series, which can then be used to reconstruct the future periodic oscillation trajectory and obtain the periodic amplitude component. One temperature response mapping model is constructed for each intrinsic mode function sub-tensor in the mid-frequency sub-tensor set. The temperature response mapping model is implemented using a regression structure, specifically a single fully connected layer with no hidden layers, an input dimension of 1, an output dimension of 1, and dynamic mapping coefficients as model parameters. These coefficients contain two learnable parameters, representing the baseline amplitude when the temperature deviation is zero and the linear influence of temperature deviation changes on amplitude changes. The model input is the temperature deviation sequence of the environmental temperature time series data relative to the historical temperature mean, and the model output is the predicted value of the instantaneous amplitude sequence. The temperature response mapping model is trained using weighted least squares training. The training samples consist of temperature deviations at historical time steps and corresponding instantaneous amplitude sequences at those historical time steps. Sample weights are determined by phase synchronization relationships, ensuring that samples corresponding to mid-frequency sub-tensor groups with stronger phase synchronization relationships contribute more to the fitting process. This guarantees that the dynamic mapping coefficients primarily learn temperature-driven components rather than temperature-independent random fluctuations. During training, a L2 regularization term is added to the dynamic mapping coefficients to suppress overfitting; the regularization strength is set to 0.001 in this embodiment. The solution is implemented using a one-time closed-form solution. Specifically, the weighted training samples are written as matrix equations, and the dynamic mapping coefficients are solved using singular value decomposition, allowing a stable solution to be obtained without gradient iteration during training. To ensure model reliability, this embodiment only constructs temperature response mapping models for intrinsic mode function sub-tensors with phase synchronization relationships not less than the stable interval corresponding to the coherence determination used to divide the mid-frequency sub-tensor groups in step S23. Intrinsic mode function sub-tensors with low phase synchronization relationships do not have temperature response mapping models constructed and do not participate in the reconstruction of periodic amplitude components. After training the temperature response mapping model, the future weather forecast temperature time series is input into the model. The input method involves calculating the temperature deviation for each future time step of the future weather forecast temperature time series and using it as the model input. The model output is the temperature driving amplitude corresponding to that future time step. The temperature driving amplitude, together with the instantaneous phase evolution law of the subtensor of the intrinsic mode function of that order, is used to reconstruct the future periodic oscillation trajectory. The reconstruction results of each order are then synthesized into a periodic amplitude component according to the predetermined weighted superposition rule in step S44. This component, along with the trajectory expectation value, is then linearly superimposed with the component in step S45 at the same future time step, outputting the crack evolution state trajectory.
[0192] Example 2, refer to Figure 2 It provides a bridge crack data trend prediction system based on big data, including:
[0193] The data acquisition module acquires original images of the apparent damage in the cracked area and obtains long-period sampling time series. It then performs three-dimensional spatial coordinate transformation and pixel-level edge segmentation processing to reconstruct the spatiotemporal conjugate tensor.
[0194] The tensor construction module adds white noise to the spatiotemporal conjugate tensor and sieves it along the time axis. It then groups the tensors according to the average fluctuation period, energy, and coherence with the ambient temperature time series data to obtain the residual trend tensor and isolate and store the mid-frequency sub-tensor group.
[0195] The decomposition and sieving module discretizes and iteratively solves the residual trend tensor under shape boundary conditions to obtain the evolution trend index sequence.
[0196] The prediction fusion module obtains the future weather forecast temperature time series, extrapolates the trajectory expectation value of the evolution trend index sequence within the future time step, and calculates the periodic amplitude component based on the mid-frequency sub-tensor group and the future weather forecast temperature time series. The two are then linearly superimposed to output the crack evolution state trajectory.
[0197] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.
Claims
1. A bridge crack data trend prediction method based on big data, characterized by, Includes the following steps: Step S1: Collect the original image of the apparent disease in the cracked area and obtain the long-period sampling time series. Perform three-dimensional spatial coordinate transformation and pixel-level edge segmentation processing, and reconstruct to obtain the spatiotemporal conjugate tensor. Step S2 involves adding white noise to the spatiotemporal conjugate tensor and sieving it along the time axis. The tensor is then grouped according to its average fluctuation period, energy, and coherence with the ambient temperature time series data to obtain the residual trend tensor. The intermediate frequency sub-tensor groups are then stored in isolation. Specifically, this includes: Step S21: Add M-order white noise signals to the spatiotemporal conjugate tensor to obtain the corresponding auxiliary mask signal matrix; Step S22: Use the time-domain adaptive sieving operator to sieve the auxiliary mask signal matrix along the time axis, extract the sub-tensors of each order of intrinsic mode function, and define the remaining data after sieving as residual components until the residual components meet the preset termination condition, and define them as residual component tensors. Step S23: Calculate the average fluctuation period and energy of each order of intrinsic mode function subtensor, and divide the order of subtensors with average fluctuation period less than a preset first threshold and energy with random distribution into high-frequency subtensor groups and perform a zeroing operation. The orders of intrinsic mode function subtensors with a coherence greater than a preset second threshold and a deviation between the average fluctuation period and the main period of the ambient temperature less than a preset third threshold are classified into mid-frequency subtensor groups and isolated storage operations are performed. The remaining low-frequency subtensor sets, excluding the high-frequency and mid-frequency subtensor sets, are merged with the residual component tensors to obtain the residual trend tensor. Specifically, step S23 includes: Step S231: Calculate the energy density of each order intrinsic mode function subtensor and match it with the center frequency of each order subtensor to determine the frequency distribution characteristics of the energy concentration interval. Step S232: Compare the average fluctuation period of each order of intrinsic mode function subtensor with a preset first threshold. If the average fluctuation period is less than the preset first threshold and the energy density is spatially random, then divide the order of intrinsic mode function subtensor into a high-frequency subtensor group and perform a filtering operation. Step S233: Obtain ambient temperature time series data synchronized with the sensor network, calculate the ambient temperature main period of the ambient temperature time series data, and calculate the coherence between the subtensors of each order of intrinsic mode function and the ambient temperature time series data. When the coherence is greater than the preset second threshold and the deviation between the average fluctuation period and the main period of the ambient temperature is less than the preset third threshold, the subtensor of this order is divided into the mid-frequency subtensor group. Step S234: Isolate and store the mid-frequency sub-tensor group, and merge the remaining low-frequency sub-tensor group with the residual component tensor to obtain the residual trend tensor; Step S3: Discretize and iteratively solve the residual trend tensor under shape boundary conditions to obtain the evolution trend index sequence; Step S4: Obtain the future weather forecast temperature time series, extrapolate the trajectory expectation value of the evolution trend index sequence within the future time step, calculate the periodic amplitude component based on the mid-frequency sub-tensor group and the future weather forecast temperature time series, and perform linear superposition to output the crack evolution state trajectory.
2. The big data based bridge crack data trend prediction method of claim 1, wherein, Step S1 specifically includes: Step S11: Use the drone camera to collect the original images of the apparent damage in the cracked area, and simultaneously acquire the long-period sampling time series corresponding to the sensor network; Step S12: Deblur the original image of the apparent defects to obtain a high-quality image of the bridge crack. Step S13: Use the aerial triangulation algorithm to perform full-field collinearity equation adjustment calculation on high-quality bridge crack images, calculate the spatial internal and external parameters of the UAV camera and generate dense point cloud data. Acquire laser point cloud data and register it with dense point cloud data to obtain the laser point cloud spatial coordinate transformation matrix.
3. The big data based bridge crack data trend prediction method of claim 2, wherein, Step S1 also includes: Step S14: Use a computer vision network to perform pixel-level edge segmentation on the high-quality bridge crack image, extract the maximum crack width, physical length of the crack skeleton line and pixel-level apparent area of the crack region in the current iteration step, and mark the two-dimensional pixel coordinates. Step S15: Using the laser point cloud spatial coordinate transformation matrix, the two-dimensional pixel coordinates are projected onto the spatial geometric surface of the bridge to calculate the spatial three-dimensional projection curvature. Step S16: The maximum crack width, the physical length of the crack skeleton line, the pixel-level apparent area of the crack region, and the spatial three-dimensional projection curvature are tensorized and recombined according to a preset arrangement order to obtain the spatiotemporal conjugate tensor.
4. The big data based bridge crack data trend prediction method of claim 3, wherein, Step S3 specifically includes: Step S31: Using the residual trend tensor as the objective approximation term, establish a constrained optimization objective function: ; in, To smooth the trend function in The values of each historical sampling time step, Let be the residual trend tensor at the t-th historical sampling time step. For smoothing coefficients, This represents the total number of historical sampling time steps. To smooth the trend function in The second derivative of each historical sampling time step with respect to time. To find, among all smooth trend functions that satisfy the shape boundary conditions, the objective function of constraint optimization is... Obtain the minimum smoothing trend function; Step S32, apply shape boundary conditions with respect to the time derivative below the constrained optimization objective function: ; Step S33: Discretize and iteratively solve the constraint optimization objective function and shape boundary conditions to obtain a smooth trend function sequence. Perform first-order difference operation on the smooth trend function sequence to obtain an evolution trend index sequence.
5. The method for predicting bridge crack data trends based on big data as described in claim 4, characterized in that, Step S33 specifically includes: The values of the smoothing trend function at discrete time steps are represented as a smoothing trend function vector, and the constrained optimization objective function is discretized into a matrix form of quadratic and linear terms. The quadratic form matrix and linear term vector corresponding to the smoothing regularization term are constructed using the second-order difference operator. The shape boundary conditions are discretized into a linear inequality constraint matrix with non-negative differences between adjacent time steps; The interior point method is used to iteratively solve the quadratic form matrix, linear term vector, and linear inequality constraint matrix, and the iterative smoothing trend function vector is updated in each iteration; When the absolute difference between the corresponding elements of the iterative smoothing trend function vectors of adjacent iteration steps is less than the preset convergence threshold, the current iterative smoothing trend function vector is determined as the convergent smoothing trend function vector, and the smoothing trend function sequence is obtained by arranging them in the order of time steps. The evolutionary trend index sequence is obtained by performing a first-order difference operation on the smoothed trend function sequence.
6. The method for predicting bridge crack data trends based on big data as described in claim 5, characterized in that, Step S4 specifically includes: Step S41: Obtain the future weather forecast temperature time series, which includes the ambient temperature data for the future time step; Step S42: Calculate the mean of the first-order difference at the end of the smoothed trend function sequence, use it as the baseline step size increment, and construct the state transition matrix of the first-order Kalman filter operator. Step S43: Input the baseline step size increment into the state transition matrix of the first-order Kalman filter operator, and extrapolate the expected value of the trajectory of the smooth trend function sequence within the future time step using a forward time sliding window rolling method. Step S44: Extract the intermediate frequency sub-tensor set, calculate the dynamic mapping coefficient between the intermediate frequency sub-tensor set and the future weather forecast temperature time series, and obtain the periodic amplitude component of the future ambient temperature. Step S45: Linearly superimpose the trajectory expectation value and the periodic amplitude component to obtain the crack evolution state trajectory.
7. The method for predicting bridge crack data trends based on big data as described in claim 6, characterized in that, Step S44 specifically includes: The intermediate frequency sub-tensor set is extracted, and the instantaneous amplitude sequence and instantaneous phase sequence of the intermediate frequency sub-tensor set are calculated using Hilbert transform; Obtain the main period of ambient temperature corresponding to the ambient temperature time series data, and calculate the phase synchronization relationship between the intermediate frequency sub-tensor set and the ambient temperature time series. A temperature response mapping model is constructed based on the instantaneous amplitude sequence, instantaneous phase sequence, and phase synchronization relationship; Input the future weather forecast temperature time series into the temperature response mapping model to calculate the temperature driving amplitude corresponding to the future time step; Based on the temperature-driven amplitude and instantaneous phase evolution law, the future periodic oscillation trajectory is reconstructed to obtain the periodic amplitude component of the future ambient temperature.
8. A bridge crack data trend prediction system based on big data, which is applied to the bridge crack data trend prediction method based on big data as described in any one of claims 1-7, characterized in that, include: The data acquisition module acquires original images of the apparent damage in the cracked area and obtains long-period sampling time series. It then performs three-dimensional spatial coordinate transformation and pixel-level edge segmentation processing to reconstruct the spatiotemporal conjugate tensor. The tensor construction module adds white noise to the spatiotemporal conjugate tensor and sieves it along the time axis. It then groups the tensors according to the average fluctuation period, energy, and coherence with the ambient temperature time series data to obtain the residual trend tensor and isolate and store the mid-frequency sub-tensor group. The decomposition and sieving module discretizes and iteratively solves the residual trend tensor under shape boundary conditions to obtain the evolution trend index sequence. The prediction fusion module obtains the future weather forecast temperature time series, extrapolates the trajectory expectation value of the evolution trend index sequence within the future time step, and calculates the periodic amplitude component based on the mid-frequency sub-tensor group and the future weather forecast temperature time series. The two are then linearly superimposed to output the crack evolution state trajectory.
Citation Information
Patent Citations
Bridge construction crack dynamic detection method and system based on AI visual identification
CN121639703A
Error compensation method for short-term wind power trend prediction
CN113505909A
Modeling analysis method for ablation characteristics of cable buffer layer
CN121980934A
