Expansive soil slope crack detection method and system based on big data processing

Abstract

The invention relates to the technical field of expansive soil slope crack intelligent detection, and discloses an expansive soil slope crack detection method and system based on big data processing, and the expansive soil slope crack detection method based on big data processing establishes a unified space-time coordinate system through a multi-source data fusion technology. A deep convolutional neural network and an attention mechanism are adopted to perform intelligent feature extraction and multi-modal fusion, and an improved U-Net network is utilized to realize fracture accurate segmentation and geometric parameter extraction. Time sequence modeling is performed on fracture evolution in combination with a multi-head attention mechanism, system characteristics are calculated and judged based on a Lyapunov index of a chaos theory, and accurate prediction of a future development trend is realized. And establishing a multi-dimensional risk assessment system, and carrying out graded early warning and intelligent decision making according to the comprehensive risk index. The method is suitable for intelligent monitoring of the expansive soil slope in the polar region permafrost region, and an innovative technical solution is provided for infrastructure safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for cracks in expansive soil slopes, and more specifically, to a method and system for detecting cracks in expansive soil slopes based on big data processing. Background Technology

[0002] Expansive soil is a special type of cohesive soil with significant swelling and shrinkage characteristics, widely distributed in semi-arid and semi-humid regions across all continents. This soil undergoes significant volume changes with variations in water content; its ability to swell upon contact with water and shrink upon loss of water makes it one of the major sources of geological hazards in engineering construction. Particularly in polar and frigid regions, expansive soil is further affected by freeze-thaw cycles, creating a more complex dual mechanism of freeze-thaw action and swelling / shrinkage. Under these complex environmental conditions, expansive soil slopes are prone to developing fissures of various sizes. The development of these fissures not only directly threatens slope stability but can also trigger serious geological disasters such as landslides and collapses, posing a significant threat to the safe operation of infrastructure such as railways, highways, and buildings.

[0003] Traditional methods for detecting cracks in expansive soil slopes mainly rely on manual inspections and simple instrument measurements. While manual inspections are intuitive and reliable, they suffer from drawbacks such as high labor intensity, low efficiency, strong subjectivity, and high costs. Especially in harsh polar environments, manual operations face significant safety risks and technical challenges. In recent years, with the development of remote sensing, sensor, and computer vision technologies, some automated detection methods have begun to be applied to crack monitoring. However, existing solutions all have significant limitations: single-sensor-based methods are easily affected by environmental factors, making it difficult to guarantee accuracy and reliability; image recognition-based methods experience performance degradation under complex backgrounds and varying lighting conditions; and while laser scanning technology offers high accuracy, the equipment is expensive and it is difficult to achieve large-scale continuous monitoring.

[0004] More importantly, existing technologies generally lack the ability to predict the development trend of cracks, and can only passively detect cracks that have already formed, failing to provide forward-looking technical support for engineering decisions. Under the coupled effects of freeze-thaw cycles and expansion / contraction, the evolution of cracks in expansive soil slopes exhibits strong nonlinear and chaotic characteristics, rendering traditional linear prediction models completely inapplicable. Furthermore, most existing systems lack intelligent adaptive capabilities, failing to automatically adjust detection strategies according to different geological conditions and environmental changes, resulting in unstable performance in complex and variable real-world application environments. Therefore, there is an urgent need to develop a novel expansive soil slope crack detection system integrating multi-source data fusion, artificial intelligence algorithms, and big data processing technologies to meet the high-precision, wide-range, real-time, and intelligent requirements of modern infrastructure safety monitoring. Summary of the Invention

[0005] This invention provides a method and system for detecting cracks in expansive soil slopes based on big data processing, which solves the technical problems of low detection accuracy and difficulty in detecting cracks in expansive soil slopes in frozen soil areas in related technologies.

[0006] This invention provides a method for detecting cracks in expansive soil slopes based on big data processing, comprising: Multi-source data acquisition and standardization processing are carried out to obtain high-resolution remote sensing images, UAV multispectral images, lidar point clouds, sensor network data and meteorological and environmental data. Spatiotemporal registration and quality inspection of multi-source data are performed to obtain standardized multi-source data with a unified spatiotemporal coordinate system. Based on standardized multi-source data, a multi-scale convolutional neural network is used to extract image features, point cloud geometric features, and sensor temporal features. An attention mechanism is used to calculate adaptive fusion weights to perform weighted fusion of features from different modalities. The improved U-Net network is used to segment the crack region by weighted fusion features, the crack centerline is obtained by skeleton extraction algorithm, and the crack geometric parameters are calculated. Based on the calculated fracture parameters, a bidirectional LSTM network is constructed and combined with a multi-head attention mechanism. Chaos theory is introduced to model and predict the future evolution of fractures. Based on the prediction of future fracture evolution, a multi-dimensional risk assessment system is established, and the early warning level is determined according to the comprehensive risk index to generate decision-making recommendations. Based on the generated decision suggestions, an online learning mechanism and an experience replay buffer are used to continuously optimize the model parameters. The detection threshold is dynamically adjusted through Bayesian optimization to improve the system's adaptability under different environmental conditions.

[0007] In a preferred embodiment, the spatiotemporal registration of the multi-source data adopts a seven-parameter coordinate transformation model, which uses three translation parameters, three rotation parameters and one scale factor to unify the coordinates of different data sources, and converts the three-dimensional coordinates of the source coordinate system into the corresponding coordinates in the WGS84 coordinate system.

[0008] In a preferred embodiment, the adaptive fusion weights are calculated as follows: the context features of each data source are calculated, and the weight coefficients are obtained by normalization using the Softmax function. The final fusion features are the weighted sum of each modality feature and its corresponding weight, ensuring dynamic adjustment of the contribution of each data source under different environmental conditions.

[0009] In a preferred embodiment, the calculation of the fracture geometry parameters includes: length calculation, performing path integration along the fracture skeleton line and accumulating the Euclidean distance between adjacent skeleton points; width calculation, calculating the maximum vertical distance at each point on the skeleton line and taking the arithmetic mean of the widths of all points; orientation angle calculation, using principal component analysis to calculate the principal orientation of the fracture skeleton points; and density calculation: calculating the total fracture length per unit area.

[0010] In a preferred embodiment, the chaos theory modeling is achieved by calculating the Lyapunov exponent, which is obtained by statistically calculating the number of sampling points, the sampling time interval, and the logarithm of the derivative of the dynamic system function. When the exponent is greater than zero, it indicates that the system has chaotic characteristics, and a nonlinear prediction model is adopted.

[0011] In a preferred embodiment, the comprehensive risk index is calculated as follows: the geometric risk factor, the development trend risk factor, the environmental risk factor, and the historical risk factor are calculated separately, and the comprehensive risk index is obtained by weighted summation. Based on the magnitude of the risk index, it is divided into four warning levels: low risk, medium risk, relatively high risk, and extremely high risk.

[0012] In a preferred embodiment, the online learning mechanism uses the Adam optimizer to update model parameters, sets the learning rate to 0.001, establishes an experience replay buffer to store historical detection experience, and trains the model by randomly sampling batch data to improve the generalization ability of the algorithm.

[0013] In a preferred embodiment, a big data processing-based expansive soil slope crack detection system is used to perform the steps in the above-described big data processing-based expansive soil slope crack detection method, including: Data acquisition module: used to acquire high-resolution remote sensing images, UAV multispectral images, lidar point clouds, sensor network data, and meteorological data; Data preprocessing module: used for spatiotemporal registration, quality inspection, standardization, and noise filtering of multi-source data; Feature extraction module: Extracts multimodal features based on deep convolutional neural networks and performs adaptive fusion through an attention mechanism; Crack detection module: It uses a semantic segmentation network to identify crack regions and extracts geometric parameters through image processing algorithms; Predictive analysis module: It uses a time-series neural network to model the crack evolution process and combines chaos theory to predict trends; Early warning and decision-making module: Establish a multi-dimensional risk assessment system to generate tiered early warnings and decision-making recommendations; Self-optimization module: Continuously improves system performance through online learning and parameter optimization.

[0014] In a preferred embodiment, the feature extraction module employs a multi-scale pyramid structure convolutional neural network, comprising an encoder and a decoder. The encoder uses residual connections to enhance feature learning capabilities, while the decoder restores spatial resolution through upsampling and skip connections, and integrates channel attention and spatial attention mechanisms.

[0015] In a preferred embodiment, the predictive analysis module includes a bidirectional LSTM network and a multi-head attention mechanism, which can capture the long-term dependencies and short-term fluctuation patterns of fracture evolution, support predictions at multiple time scales, including hourly, daily and weekly predictions, and provide uncertainty quantification of the prediction results.

[0016] The beneficial effects of this invention are as follows: By employing multi-source heterogeneous data fusion technology, the system integrates the advantages of various data sources, including remote sensing imagery, lidar, and sensor networks, achieving millimeter-level precision in crack detection. Compared to traditional single-sensor methods, detection accuracy is improved, and a technological breakthrough has been achieved, moving from local point measurement to large-area surface monitoring. Employing an adaptive weighted fusion mechanism, the system dynamically adjusts the contribution weights based on the quality and reliability of different data sources, ensuring optimal detection results under various environmental conditions.

[0017] By introducing chaos theory into the modeling of fracture evolution in expansive soil and accurately identifying the chaotic characteristics of the system through Lyapunov exponent calculation, a scientific basis for nonlinear prediction is provided. A time-series modeling method combining a bidirectional LSTM network with a multi-head attention mechanism can effectively capture long-term dependencies and short-term fluctuation patterns in the fracture evolution process, achieving accurate prediction of the development trend over the next 7-30 days, thus improving prediction accuracy. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for detecting cracks in expansive soil slopes based on big data processing, according to the present invention. Figure 2 This is a block diagram of a crack detection system for expansive soil slopes based on big data processing, according to the present invention. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0020] At least one embodiment of the present invention discloses a method for detecting cracks in expansive soil slopes based on big data processing, such as... Figure 1 As shown, it includes the following steps: Step 10: Multi-source data acquisition and standardization processing. Acquire high-resolution remote sensing images, UAV multispectral images, lidar point clouds, sensor network data, and meteorological and environmental data. Perform spatiotemporal registration and quality inspection on the multi-source data to obtain standardized multi-source data with a unified spatiotemporal coordinate system. Specifically, the following steps are included: Step 11: Establish a unified spatiotemporal coordinate system; The system first establishes a unified spatiotemporal reference based on the WGS84 coordinate system. For spatial data from different sources, a seven-parameter coordinate transformation model is used for coordinate system unification. This transformation model converts the three-dimensional coordinates of the source coordinate system into the corresponding coordinates in the WGS84 coordinate system through three translation parameters, three rotation parameters, and one scale factor, ensuring that spatial location information from different data sources can be processed and analyzed within a unified coordinate framework.

[0021] Step 12: High-resolution remote sensing image data processing; The system receives high-resolution remote sensing images from various satellite platforms, including optical images and synthetic aperture radar (SAR) images. Atmospheric correction is performed on the optical images, employing a dark pixel method to remove the effects of atmospheric scattering. The specific steps are as follows: First, dark pixel regions in the image are identified, as these regions theoretically have near-zero surface reflectance; then, atmospheric path radiatives are estimated based on the observed values ​​of the dark pixels; finally, the true surface reflectance is calculated using an atmospheric radiative transfer model, eliminating the impact of atmospheric scattering on image quality.

[0022] Coherence analysis is performed on synthetic aperture radar (SAR) data, and the data quality is evaluated by calculating the coherence coefficient between complex SAR images at different times. The coherence coefficient is calculated as the ratio of the cross-correlation of two complex images to the geometric mean of their respective powers, and the value ranges from 0 to 1. The closer the value is to 1, the better the coherence and the higher the data quality.

[0023] Step 13: UAV multispectral data acquisition and processing; The system is equipped with a low-temperature resistant UAV platform and a multispectral camera for high-precision near-ground data acquisition. The UAV flies autonomously along a preset route, which is planned using a serpentine scanning mode to ensure an image overlap of over 80%.

[0024] Geometric and radiometric corrections are performed on the acquired multispectral images. Geometric correction employs a rational function model, establishing a rational polynomial function relationship between image coordinates and ground coordinates to convert image pixel coordinates into accurate locations in the geographic coordinate system, thus eliminating geometric distortions caused by factors such as sensor attitude and terrain undulations.

[0025] The Normalized Difference Vegetation Index (NDVI) is used to identify vegetated areas. NDVI is calculated by dividing the difference between the reflectance in the near-infrared band and the red band by the sum of the two. It effectively distinguishes vegetated areas from bare soil areas by utilizing the spectral characteristics of vegetation, which has high reflectance in the near-infrared band and low reflectance in the red band.

[0026] Step 14: Ground-based lidar point cloud data processing; The system deploys a fixed LiDAR scanner to periodically acquire 3D point cloud data of the slope surface. The raw point cloud is then denoised using statistical filtering methods to remove outliers. For each point in the point cloud, calculate its average distance to its nearest neighbor. By statistically analyzing the distribution of these average distances across all points, calculate the mean and standard deviation. If the average distance of a point exceeds a threshold equal to the mean plus a factor of several times the standard deviation (the threshold factor is typically 2-3), then that point is identified as noise and deleted. This statistical filtering method effectively removes outliers caused by measurement errors or environmental interference.

[0027] Ground segmentation is performed on the denoised point cloud, and a progressive triangulation algorithm is used to identify ground points. This algorithm first selects the point with the lowest elevation as the seed point, and then gradually constructs a triangulation network. By setting distance and angle thresholds, it determines whether a new point belongs to the ground, and finally divides the point cloud into two categories: ground points and non-ground points.

[0028] Step 15: Distributed sensor network data acquisition; The system deploys a wireless sensor network in the monitoring area, including temperature sensors, humidity sensors, soil moisture sensors, strain gauges, and tiltmeters. The sensors employ a low-power design and are powered by solar and wind power systems to achieve energy self-sufficiency.

[0029] Sensor data employs a timestamp synchronization mechanism to ensure time consistency across different sensors. The synchronization algorithm considers the fixed time offset between each sensor node and the reference clock, as well as the clock drift rate. Through a linear correction model, it unifies the local time of each node to the system reference time, eliminating time inconsistencies caused by hardware differences and environmental factors.

[0030] Sensor data quality is assessed by calculating a data integrity index. This integrity index is calculated as the ratio of the number of valid data points to the total number of data points to be collected, ranging from 0 to 1. A value closer to 1 indicates better data integrity. This index quantitatively evaluates the operational status and data acquisition effectiveness of the sensor network.

[0031] Step 20: Based on standardized multi-source data, a multi-scale convolutional neural network is used to extract image features, point cloud geometric features, and sensor temporal features. An attention mechanism is used to calculate adaptive fusion weights, and features from different modalities are weighted and fused. Specifically, the following steps are included: Step 21: Construct a multi-scale feature extraction network; The system employs an improved U-Net architecture as the basic feature extraction network, specifically optimized for the crack detection task. The network encoder utilizes a residual connection structure to enhance feature learning capabilities. The encoder employs a residual connection structure, where the output of each layer is processed through convolution, batch normalization, and ReLU activation functions before being residually connected to the input feature map. This design effectively mitigates the vanishing gradient problem in deep networks and enhances feature learning capabilities.

[0032] To capture crack features at different scales simultaneously, the network employs a dilated convolutional pyramid structure:

[0033] in, This represents the output features of the dilated convolution pyramid. Indicates the input feature map, This represents a 1×1 convolution kernel. This indicates a 3×3 convolution kernel with a void ratio of 0.5%. Hollow convolution, Indicates the void ratio, such as 6, 12, 18, This indicates a feature splicing operation.

[0034] Step 22: Design an attention mechanism enhancement module; By introducing channel attention and spatial attention mechanisms, important feature regions are adaptively focused on.

[0035] The channel attention mechanism extracts global information from the feature map through global average pooling and global max pooling operations, then processes them separately through a multilayer perceptron network. The outputs of the two branches are summed and then passed through a sigmoid activation function to obtain the channel attention weights. This mechanism can adaptively adjust the importance of features in different channels.

[0036] The spatial attention mechanism obtains two spatial feature maps by performing global average pooling and global max pooling along the channel dimension. These maps are then concatenated and spatial attention weights are generated through a 7×7 convolution operation and a sigmoid activation function. This mechanism can highlight important spatial location information and suppress interference from irrelevant regions.

[0037] The final attention-enhanced features are obtained by element-wise multiplying the original feature map with channel attention weights and spatial attention weights. This dual attention mechanism can simultaneously enhance important features in both the channel and spatial dimensions, improving the network's ability to perceive key information.

[0038] Step 23: Multimodal data fusion strategy; The system adopts a hierarchical fusion strategy, performing data fusion at the feature level and the decision level respectively.

[0039] At the feature level, adaptive weights are used to fuse features from different modalities. The fused features are obtained by weighted summation of the features from each modality. The weights are calculated by normalizing the learnable parameters using the softmax function to ensure that the sum of all weights is 1. This adaptive weighting mechanism can dynamically adjust the fusion ratio according to the importance of features from different modalities.

[0040] At the decision level, a Bayesian fusion strategy is used to integrate detection results from different modalities:

[0041] in, Indicates a given multimodal input The probability of crack existence under certain conditions. This represents multimodal input data. Indicates the first Input data for each modality, This indicates the condition where the crack exists. The likelihood probability of a modality. This represents the prior probability of the existence of a crack. Represents all possible categories, Indicates category The prior probability, Indicates the number of modes.

[0042] Step 24: Edge enhancement and detail preservation; For the detection of subtle cracks, a specialized edge-enhancing convolutional kernel was designed. This kernel employs a 3×3 Laplacian operator structure with a central weight of 8 and weights of -1 at the surrounding 8 positions. This effectively highlights edge and detail information in the image, making it particularly suitable for detecting subtle crack features.

[0043] To preserve detail, dense skip connections are used in the network decoder. The output of each layer of the decoder is obtained by upsampling the features of the previous layer and then weighted, fused, and concatenated with the features of the corresponding layer in the encoder. This dense connection method can fully utilize the multi-scale feature information of each layer of the encoder, effectively preserving the detailed features of the image.

[0044] Step 30: Use the improved U-Net network to segment the crack region using the weighted fused features, obtain the crack centerline through the skeleton extraction algorithm, and calculate the crack geometric parameters; Specifically, the following steps are included: Step 31: Precise fracture segmentation; The system employs an improved Dice loss function to train the segmentation network, enhancing its sensitivity to small gaps. The Dice loss measures segmentation quality by calculating the overlap between the predicted results and the ground truth labels, effectively addressing the issue of imbalanced foreground and background pixel counts, making it particularly suitable for segmenting small targets.

[0045] To address the class imbalance problem, the Focal loss function is introduced. This loss function dynamically adjusts the weights of samples with different levels of difficulty, reducing the loss contribution of easily classified samples and increasing the attention given to difficult-to-classify samples, effectively solving the problem of severe imbalance in the number of positive and negative samples.

[0046] The comprehensive loss function combines Dice loss, Focal loss, and a smoothing constraint term, balancing the contributions of different loss terms through a weighted summation. Dice loss focuses on segmentation accuracy, Focal loss handles class imbalance, and the smoothing constraint term maintains the spatial continuity of the prediction results. The three work together to optimize network performance.

[0047] Step 32: Extraction and refinement of the fracture skeleton; The skeleton is extracted from the segmented binary mask using an improved Zhang-Suen thinning algorithm. First mark the boundary points, for each pixel. If a pixel has a background point in its 8-neighborhood, it is marked as a boundary point. Then, iterative refinement is performed. In each iteration, the connectivity number is calculated, and the number of transitions from background to foreground in the 8-neighborhood of the pixel is counted to determine whether the point is a connected point. Calculate the number of non-zero neighbors and count the number of foreground pixels in the 8-neighborhood. If the thinning condition is met, delete the point. The number of non-zero neighbors is between 2 and 6, the connectivity number is equal to 1, and the connectivity condition in a specific direction is met to maintain the continuity of the skeleton. After thinning, a slit skeleton with a single pixel width is obtained.

[0048] In one embodiment of the present invention, the refinement conditions are as follows: the point is a boundary point and has 2-6 foreground pixels in its 8-neighborhood; the connectivity of the point is 1, indicating that deleting the point will not cause the skeleton to break; in the first iteration, if at least one of the neighboring points in the north, east, and south directions of the point is a foreground point, and the neighboring point in the west direction is also a foreground point, then the point is deleted; in the second iteration, if at least one of the neighboring points in the north, west, and south directions of the point is a foreground point, and the neighboring point in the east direction is also a foreground point, then the point is deleted; repeat the above two iterations until there are no points that can be deleted.

[0049] Step 33: Precise calculation of geometric parameters; Based on the skeleton extraction results, the detailed geometric parameters of the fracture are calculated: The crack length is calculated by performing path integration along the skeleton line. The total crack length is obtained by calculating and summing the three-dimensional Euclidean distances between adjacent skeleton points. This method can accurately measure the actual length of cracks in complex three-dimensional space, taking into account the curvature and undulations of the cracks in space.

[0050] The crack width is calculated by determining the vertical crack width at each point along the skeleton line. This is done by radiating rays from each skeleton point along a direction perpendicular to the tangent of the skeleton line, finding the intersection points of these rays and the crack boundaries, and calculating the distance between these intersection points to obtain the crack width at each point. The average width is obtained by taking the arithmetic mean of the width values ​​at all skeleton points.

[0051] The fracture orientation angle is calculated using principal component analysis. First, the covariance matrix of the skeleton point coordinates is calculated, including the x-direction variance, y-direction variance, and xy covariance. Then, the principal orientation is determined through eigenvalue decomposition of the covariance matrix. The principal orientation angle is calculated using the arctangent function of the covariance components, representing the main extension direction of the fracture.

[0052] Fracture density calculation: The degree of fracture development is quantified by calculating the total fracture length per unit area. Fracture density equals the sum of all fracture lengths within the detection area divided by the total area of ​​the area. This indicator can objectively reflect the density of fracture development on the slope.

[0053] Fractal connectivity analysis: Graph theory is used to analyze the connectivity of the fracture network. The connectivity coefficient is calculated by statistically analyzing the proportion of fractures connected to other fractures to the total number of fractures. The value ranges from 0 to 1. The larger the value, the better the connectivity of the fracture network and the more significant the impact on slope stability.

[0054] Optionally, in some embodiments, a fracture network modeling method based on graph neural networks is employed. Traditional methods treat fractures as independent linear features; this embodiment proposes modeling the fracture network as a graph structure and using graph neural networks for analysis. The fracture network graph representation uses a graph structure modeling, where nodes represent fracture intersections and edges represent fracture segments. Graph convolution operations update node features by aggregating neighbor node information, and feature propagation is performed using normalized adjacency and degree matrices, effectively capturing the topological features of the fracture network.

[0055] Step 40: Based on the calculated fracture parameters, construct a bidirectional LSTM network combined with a multi-head attention mechanism, and introduce chaos theory modeling to predict the future evolution of fractures. Specifically, the following steps are included: Step 41: Construct a long short-term memory network; The system employs a bidirectional LSTM network to capture the temporal characteristics of the fracture evolution. The network structure includes a forward LSTM and a backward LSTM. The forward LSTM processes data from the beginning to the end of the time series, capturing the influence of historical information on the current state; the backward LSTM processes data from the end to the beginning, capturing the influence of future information on the current state. The final hidden state is obtained by concatenating the outputs of the forward and backward LSTMs, enabling the simultaneous utilization of past and future temporal information.

[0056] Step 42: Multi-head attention mechanism; Introducing a multi-head attention mechanism enhances the model's ability to perceive critical moments:

[0057] in:

[0058]

[0059] in, Represents the query matrix. Represents the key matrix, Represents a value matrix, They represent the first The query, key, and value weight matrix of each attention head. Indicates the output projection matrix. The feature dimension of the key vector. For the number of attention heads, This indicates a feature concatenation operation. This represents the softmax activation function.

[0060] Step 43: Integrated modeling of environmental factors; Environmental factors such as meteorological conditions and geological parameters are used as additional inputs to jointly model fracture parameters over time. Environmental features are encoded using a multilayer perceptron network to encode multidimensional environmental parameters such as temperature, humidity, rainfall, and wind speed, resulting in a compact environmental feature representation. Joint features are obtained by concatenating LSTM hidden states and environmental feature vectors, achieving collaborative modeling of fracture evolution and environmental conditions.

[0061] Step 44: Nonlinear modeling guided by chaos theory; Considering the strong nonlinear characteristics of the freeze-thaw-expansion-contraction coupling system, chaos theory is introduced for modeling.

[0062] First, the phase space is reconstructed using the delayed embedding theorem:

[0063] in, Indicates time The phase space reconstruction vector, Indicates time The observed values ​​of the fracture parameters The delay time parameter is used to determine the time interval between adjacent embedding points in the time series. The embedding dimension represents the dimension of the reconstructed phase space.

[0064] Calculate the Lyapunov exponent to determine the chaotic characteristics of a system:

[0065] in, Represents the maximum Lyapunov exponent, used to quantify the degree of chaos in a system. This represents the total length of the time series. Indicates the sampling time interval. Indicates the first The state value at each moment. Represents nonlinear dynamic functions At point The derivative at point, This represents the natural logarithm function.

[0066] if This indicates that the system has chaotic characteristics and requires the use of a nonlinear prediction model.

[0067] Optionally, in some embodiments, an anomaly detection method based on a variational autoencoder is employed. This method learns a latent representation of normal crack patterns and detects the development of anomalous cracks through reconstruction errors. The variational autoencoder maps input data to a probability distribution in the latent space through an encoder, and the decoder reconstructs the original data from the latent variables. Both the encoder and decoder are modeled using Gaussian distributions, and the model parameters are optimized through a variational lower bound, which includes the reconstruction loss and the KL divergence regularization term, enabling the learning of latent representations of the data and the detection of anomalous patterns.

[0068] Step 45: Multi-step prediction and uncertainty quantification; The system supports forecasts at multiple time scales, including hourly, daily, and weekly forecasts. Recursive prediction strategy:

[0069] in, Indicates time The predicted value, Indicates the prediction step size. For parameters Predictive models, Represents a series of historical predicted values. This is the length of the history window.

[0070] The basic implementation steps of the prediction model are as follows: Input data preparation, p: The target value at time t+k to be predicted; k: The time step for prediction, such as predicting 1 day, 3 days, etc.; p: The length of the historical data window, used to determine how many historical data points to use. Model building: : Prediction model function, containing trainable parameters θ; input is a historical sequence of length p. The output is the predicted value at future time t+k. Recursive prediction: Each prediction uses historical data from the most recent p time points; the prediction results serve as input for the next prediction; the values ​​for the next k time points are predicted step by step; this recursive prediction strategy achieves continuous prediction of multiple future time points by continuously utilizing new prediction results.

[0071] The Monte Carlo Dropout method is used to quantify prediction uncertainty.

[0072]

[0073] in, This represents the predicted mean. Indicates the prediction variance. For the number of samples taken in Monte Carlo, Indicates the first The model after random dropout. Indicates the first Parameters after random dropout This represents the input feature vector.

[0074] Prediction confidence interval:

[0075] in, This indicates a confidence interval; here it is a 95% confidence interval. It is the 97.5th percentile of the standard normal distribution. This represents the standard deviation of the forecast.

[0076] Step 50: Based on the results of future fracture evolution prediction, establish a multi-dimensional risk assessment system, determine the early warning level according to the comprehensive risk index, and generate decision-making recommendations. Specifically, the following steps are included: Step 51: Construction of a multi-dimensional risk indicator system; The system establishes a comprehensive assessment framework encompassing four dimensions: geometric risk, development risk, environmental risk, and historical risk. Geometric risk indicators:

[0077] in, This represents the geometric risk index value. This indicates the current maximum crack length. This indicates the current maximum crack width. Indicates the current fracture density. Represents the fracture connectivity coefficient. , , These represent the critical thresholds for length, width, and density, respectively. , , , These are the weighting coefficients corresponding to length, width, density, and fracture connectivity coefficient, respectively. .

[0078] Development risk indicators:

[0079] in, Indicates the value of the development risk indicator. This represents the rate of change of the fracture length over time (development rate). The second derivative (development acceleration) represents the fracture development. This indicates the predicted future increase in crack length. This represents the critical growth rate threshold. This represents the critical acceleration threshold for development. This represents the threshold for critical growth. , , These are the weighting coefficients corresponding to the critical growth rate, acceleration, and growth amount, respectively. .

[0080] Environmental risk indicators:

[0081] in, Indicates the value of environmental risk indicators. , , These are the weighting coefficients for temperature, freeze-thaw cycles, and soil moisture content, respectively. , Indicates the range of temperature change. Indicates the danger threshold of temperature changes. This indicates the number of freeze-thaw cycles within the statistical period. This indicates the current soil moisture content. This represents the critical value for soil moisture content.

[0082] Historical risk indicators:

[0083] in, This represents the historical risk indicator value. Weighting coefficients representing the probability of historical failures. The weighting coefficients representing the time factor. The sum of is 1. This represents the probability of failure based on historical data statistics. Indicates the current time. Indicates the time when the last failure event occurred. This refers to a statistically safe time interval.

[0084] Step 52: Dynamic threshold adaptive adjustment; The system dynamically adjusts the warning threshold based on historical data and environmental conditions. The adaptive threshold is calculated by a weighted combination of the base threshold and environmental adjustment factors, and can automatically adjust the warning sensitivity according to real-time environmental conditions.

[0085] Environmental moderating factors include: a temperature moderating factor, which uses a hyperbolic tangent function to normalize temperature deviations and reflect the impact of temperature changes on crack development; a humidity moderating factor, which maps humidity values ​​to the 0-1 range through linear normalization to quantify the impact of humidity on soil swelling and shrinkage; and a seasonal moderating factor, which uses a sine function to simulate seasonal variation patterns and capture annual periodic environmental impact patterns. Step 53: Multi-level early warning decision-making mechanism; The system employs a tiered early warning strategy, determining the early warning level based on a comprehensive risk index: The comprehensive risk score is calculated by weighting and summing four dimensions: geometric risk, development risk, environmental risk, and historical risk, with values ​​ranging from 0 to 1. The sum of the weight coefficients for each dimension is 1 to ensure the normalization of the score result. This comprehensive score can fully reflect the overall risk level of slope cracks.

[0086] Level 1 (Green Alert): Risk level: Low; Response measures: Routine monitoring; Reporting frequency: Weekly; Level 2 (Yellow Alert): Risk level: Medium; Response measures: Enhanced monitoring, on-site inspections; Reporting frequency: Daily; Level 3 (Orange Alert): Risk level: High; Response measures: Real-time monitoring, emergency response preparation; Reporting frequency: Every 4 hours; Level 4 (Red Alert): Risk level: Extremely high; Response measures: Immediate response, activate emergency plan; Reporting frequency: Real-time; Optionally, in some embodiments, a reinforcement learning-based monitoring strategy optimization method is employed. This method models the monitoring strategy selection as a Markov decision process and uses reinforcement learning to automatically optimize detection parameters. Reinforcement learning-based monitoring strategy optimization models the monitoring task as a Markov decision process. The state space includes the crack state and environmental conditions, while the action space includes strategy selections such as detection frequency and sensor configuration. The reward function comprehensively considers detection accuracy, efficiency, and cost. The Q-value is updated through temporal difference learning, enabling the agent to learn the optimal monitoring strategy.

[0087] Step 54: Intelligent Decision Support; The system provides corresponding decision-making suggestions for different levels of early warning: The decision tree-based suggestion generation system comprehensively considers the overall risk score, various risk components, current state parameters, and historical information, and automatically generates corresponding response suggestions through a trained decision tree model. This system can provide targeted response strategies based on different risk situations.

[0088] Decision tree construction: Information gain calculation:

[0089] in, Represents attributes For sample set Information gain Represents the sample set Information entropy Represents attributes The set of all possible values, Represents attributes A specific value, Represents attributes Values a subset of samples Representing a subset The number of samples, Represents the original sample set Total number of samples Representing a subset Information entropy.

[0090] Cost-benefit optimization decision-making seeks the course of action that minimizes costs while satisfying risk control constraints. This optimization model comprehensively considers labor costs, equipment costs, and time costs to ensure optimal resource allocation within an acceptable risk level.

[0091] Step 60: Based on the generated decision suggestions, the model parameters are continuously optimized using an online learning mechanism and an experience replay buffer. The detection threshold is dynamically adjusted through Bayesian optimization to improve the system's adaptability under different environmental conditions. Step 61: Online learning mechanism; The system employs an incremental learning method to continuously optimize model performance. Model parameter updates use a variant of stochastic gradient descent, the Adam optimizer:

[0092]

[0093]

[0094]

[0095]

[0096] in, Indicates time The gradient of the loss function, Indicates time The first moment estimate (exponential moving average of the gradient). Indicates time The second moment estimate (exponential moving average of the squared gradient). This represents the decay rate of the first moment. This represents the decay rate of the second moment. The learning rate represents the step size for updating parameters. This indicates a smoothing term used to prevent division by zero errors. This represents the first-moment estimate after bias correction. This represents the second-order moment estimate after bias correction. This represents the updated model parameters. This represents the current model parameters.

[0097] Step 62: Experience replay mechanism; An experience replay buffer is established to store historical experience, including a four-tuple of data: system state, detection strategy, reward signal, and next state. A time decay factor is used to calculate sample importance weights, giving higher learning priority to recent experiences. Batch data is randomly sampled from the buffer for model updates, and network parameters are optimized by minimizing the expected loss.

[0098] Step 63: Bayesian hyperparameter optimization; Automatic hyperparameter optimization using Gaussian processes: Gaussian process representation of the function:

[0099] in, Represented by hyperparameter vector The input is the objective function value (such as the model validation accuracy). Represents a Gaussian process. Indicates the Gaussian process at point The mean function at a given point is used to estimate the expected value of the objective function. This represents the covariance kernel function, used to measure the similarity between points with different hyperparameters.

[0100] Expected Improvement (EI) is used as the acquisition function:

[0101] in, This represents the expected improvement value, used to balance exploration and exploitation. This represents the mathematical expectation operation. Represented by hyperparameter vector The input is the objective function value (such as the model validation accuracy). This represents the currently observed optimal objective function value. This indicates the current optimal hyperparameter configuration. This indicates the operation of finding the maximum value.

[0102] Step 64: Model integration and robustness enhancement; Improving system robustness by employing ensemble learning methods: Boosting ensembles reduce model bias and improve prediction accuracy by combining the predictions of multiple weak learners in a weighted manner, with the weights dynamically adjusted based on the performance of each learner.

[0103] Uncertainty propagation quantifies prediction uncertainty by calculating the variance of predictions from each base learner, providing a credibility assessment for decision-making.

[0104] like Figure 2 As shown, in one embodiment of the present invention, a crack detection system for expansive soil slopes based on big data processing is provided, comprising: Data acquisition module: used to acquire high-resolution remote sensing images, UAV multispectral images, lidar point clouds, sensor network data, and meteorological data; Data preprocessing module: used for spatiotemporal registration, quality inspection, standardization, and noise filtering of multi-source data; Feature extraction module: Extracts multimodal features based on deep convolutional neural networks and performs adaptive fusion through an attention mechanism; Crack detection module: It uses a semantic segmentation network to identify crack regions and extracts geometric parameters through image processing algorithms; Predictive analysis module: It uses a time-series neural network to model the crack evolution process and combines chaos theory to predict trends; Early warning and decision-making module: Establish a multi-dimensional risk assessment system to generate tiered early warnings and decision-making recommendations; Self-optimization module: Continuously improves system performance through online learning and parameter optimization.

[0105] In one embodiment of the present invention, an application example of a method for detecting cracks in expansive soil slopes based on big data processing is provided; A 12-month monitoring and verification process was conducted at an Arctic research station to detect and validate cracks in expansive soil slopes. The monitoring area covered 2.5 square kilometers, where the expansive soil is subjected to both freeze-thaw cycles and expansion-contraction cycles, resulting in complex crack development. The monitoring objective was to promptly detect and predict crack development trends to ensure the safety of the research station's infrastructure.

[0106] Step 10 Implementation: Multi-source data acquisition and standardization processing; Data acquisition equipment configuration: 4 fixed LiDAR scanners with a scanning accuracy of ±2mm and a coverage range of 500m; 2 UAV platforms equipped with multispectral cameras with a resolution of 0.05m; 50 distributed sensor nodes, including sensors for temperature, humidity, strain, etc.; high-resolution satellite imagery with a resolution of 0.5m and an update cycle of 3 days; Table 1 shows the collected multi-source data; Table 1: Statistics of Multi-Source Data Collection

[0107] Coordinate system processing result: Using a seven-parameter coordinate transformation model, different data sources are unified into the WGS84 coordinate system, with a transformation accuracy of ±0.5m, which meets the requirements of multi-source data fusion.

[0108] Step 20: Intelligent feature extraction and multimodal fusion; Improved U-Net network training: The network was trained using an annotated crack dataset containing 5000 crack images under different lighting and weather conditions.

[0109] Table 2 shows the performance of the feature extraction network; Table 2: Performance of Feature Extraction Network

[0110] Multimodal fusion effect: Adaptive weight fusion strategy is adopted, with optical image weight of 0.4, radar data weight of 0.3, and multispectral data weight of 0.3. After fusion, the detection accuracy is improved to 94.2%.

[0111] Step 30: Crack semantic segmentation and geometric parameter extraction; The results of fracture segmentation and parameter extraction are shown in Table 3: Table 3: Statistics on the Extraction of Crack Geometric Parameters

[0112] Skeleton extraction accuracy verification: The Zhang-Suen refinement algorithm was used to extract the fracture skeleton. Compared with manual annotation, the skeleton extraction accuracy reached 96.8%, and the geometric parameter measurement error was controlled within ±3mm.

[0113] Step 40 Implementation: Time Series Modeling and Evolutionary Prediction; Bidirectional LSTM network configuration: The network consists of 2 layers of bidirectional LSTM, 256 hidden units, 8 heads of multi-head attention mechanism, and the training data is a 6-month continuous crack evolution sequence.

[0114] Table 4 shows the performance evaluation results of time series prediction. Table 4: Performance Evaluation of Time Series Prediction

[0115] Chaotic characteristics analysis results: The calculated maximum Lyapunov exponent of the system is 0.23, indicating that the system has weak chaotic characteristics, which verifies the necessity of using nonlinear modeling methods.

[0116] Step 50 Implementation: Intelligent Early Warning and Risk Assessment; Multidimensional risk assessment implementation: Establish a four-dimensional assessment system that includes geometric risk, development risk, environmental risk and historical risk, with weights of 0.3, 0.3, 0.2 and 0.2 for each dimension.

[0117] Table 5 shows the operational effectiveness of the early warning system; Table 5: Operational Effectiveness of the Early Warning System

[0118] Dynamic threshold adjustment effect: The system automatically adjusts the warning threshold according to environmental conditions. Compared with the fixed threshold method, the false alarm rate is reduced by 35% and the false alarm rate is reduced by 28%.

[0119] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for detecting cracks in expansive soil slopes based on big data processing, characterized in that, Includes the following steps: Multi-source data acquisition and standardization processing are carried out to obtain high-resolution remote sensing images, UAV multispectral images, lidar point clouds, sensor network data and meteorological and environmental data. Spatiotemporal registration and quality inspection of multi-source data are performed to obtain standardized multi-source data with a unified spatiotemporal coordinate system. Based on standardized multi-source data, a multi-scale convolutional neural network is used to extract image features, point cloud geometric features, and sensor temporal features. An attention mechanism is used to calculate adaptive fusion weights to perform weighted fusion of features from different modalities. The improved U-Net network is used to segment the crack region by weighted fusion features, the crack centerline is obtained by skeleton extraction algorithm, and the crack geometric parameters are calculated. Based on the calculated fracture parameters, a bidirectional LSTM network is constructed and combined with a multi-head attention mechanism. Chaos theory is introduced to model and predict the future evolution of fractures. Based on the prediction of future fracture evolution, a multi-dimensional risk assessment system is established, and the early warning level is determined according to the comprehensive risk index to generate decision-making recommendations. Based on the generated decision suggestions, an online learning mechanism and an experience replay buffer are used to continuously optimize the model parameters. The detection threshold is dynamically adjusted through Bayesian optimization to improve the system's adaptability under different environmental conditions.

2. The method for detecting cracks in expansive soil slopes based on big data processing according to claim 1, characterized in that, The spatiotemporal registration of the multi-source data adopts a seven-parameter coordinate transformation model, which uses three translation parameters, three rotation parameters and one scale factor to unify the coordinates of different data sources, and converts the three-dimensional coordinates of the source coordinate system into the corresponding coordinates in the WGS84 coordinate system.

3. The method for detecting cracks in expansive soil slopes based on big data processing according to claim 1, characterized in that, The adaptive fusion weights are calculated as follows: the context features of each data source are calculated, and the weight coefficients are obtained by normalization using the Softmax function. The final fusion features are the weighted sum of each modality feature and its corresponding weight, ensuring the dynamic adjustment of the contribution of each data source under different environmental conditions.

4. The method for detecting cracks in expansive soil slopes based on big data processing according to claim 1, characterized in that, The calculation of the fracture geometry parameters includes: length calculation, which involves path integration along the fracture skeleton line and summing the Euclidean distances between adjacent skeleton points; width calculation, which involves calculating the maximum vertical distance at each point on the skeleton line and taking the arithmetic mean of the widths of all points; orientation angle calculation, which involves calculating the principal orientation of the fracture skeleton points using principal component analysis; and density calculation, which involves calculating the total fracture length per unit area.

5. The method for detecting cracks in expansive soil slopes based on big data processing according to claim 1, characterized in that, The chaotic theory modeling is achieved by calculating the Lyapunov exponent, which is obtained by statistically calculating the number of sampling points, the sampling time interval, and the logarithm of the derivative of the dynamic system function. When the exponent is greater than zero, it indicates that the system has chaotic characteristics, and a nonlinear prediction model is adopted.

6. The method for detecting cracks in expansive soil slopes based on big data processing according to claim 1, characterized in that, The comprehensive risk index is calculated as follows: geometric risk factor, development trend risk factor, environmental risk factor and historical risk factor are calculated separately, and the comprehensive risk index is obtained by weighted summation. According to the magnitude of the risk index, it is divided into four warning levels: low risk, medium risk, relatively high risk and extremely high risk.

7. The method for detecting cracks in expansive soil slopes based on big data processing according to claim 1, characterized in that, The online learning mechanism uses the Adam optimizer to update model parameters with a learning rate of 0.

001. It also establishes an experience replay buffer to store historical detection experience and trains the model by randomly sampling batches of data to improve the generalization ability of the algorithm.

8. A crack detection system for expansive soil slopes based on big data processing, used to perform the steps in the crack detection method for expansive soil slopes based on big data processing as described in any one of claims 1-7, characterized in that, include: Data acquisition module: used to acquire high-resolution remote sensing images, UAV multispectral images, lidar point clouds, sensor network data, and meteorological data; Data preprocessing module: used for spatiotemporal registration, quality inspection, standardization, and noise filtering of multi-source data; Feature extraction module: Extracts multimodal features based on deep convolutional neural networks and performs adaptive fusion through an attention mechanism; Crack detection module: It uses a semantic segmentation network to identify crack regions and extracts geometric parameters through image processing algorithms; Predictive analysis module: It uses a time-series neural network to model the crack evolution process and combines chaos theory to predict trends; Early warning and decision-making module: Establish a multi-dimensional risk assessment system to generate tiered early warnings and decision-making recommendations; Self-optimization module: Continuously improves system performance through online learning and parameter optimization.

9. The expansive soil slope crack detection system based on big data processing according to claim 8, characterized in that, The feature extraction module adopts a multi-scale pyramid structure convolutional neural network, which includes an encoder and a decoder. The encoder uses residual connections to enhance feature learning ability, and the decoder restores spatial resolution through upsampling and skip connections. It also integrates channel attention mechanism and spatial attention mechanism.

10. The expansive soil slope crack detection system based on big data processing according to claim 8, characterized in that, The predictive analysis module includes a bidirectional LSTM network and a multi-head attention mechanism, which can capture the long-term dependencies and short-term fluctuation patterns of fracture evolution, support predictions at multiple time scales, including hourly, daily and weekly predictions, and provide uncertainty quantification of prediction results.

Images