Safety risk assessment method and platform based on deformation identification of dangerous slope
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CTI DISASTER PREVENTION TECH (SHENZHEN) CO LTD
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-09
Smart Images
Figure CN122176587A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image recognition technology, and in particular to a safety risk assessment method and platform based on dangerous slope deformation identification. Background Technology
[0002] Slope morphological instability is a common geological hazard in transportation and other fields. Currently, the mainstream risk assessment methods for slope morphological instability mainly rely on professional monitoring equipment and manual inspection. On the one hand, professional equipment monitoring, such as GNSS and InSAR, can provide accurate deformation data, but it generally has the disadvantages of complex deployment and maintenance and limited spatial coverage, making it difficult to apply on a large scale in remote areas or along existing infrastructure. On the other hand, regular manual inspection is constrained by low efficiency and often cannot achieve continuous real-time monitoring, and there are safety risks in steep and dangerous areas. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of the prior art. In view of the aforementioned problems, and in conjunction with the first aspect of this invention, a safety risk assessment method based on dangerous slope deformation identification is provided, the method comprising:
[0004] Acquire slope video stream data, and perform preprocessing operations such as correction mapping, color correction, and ROI determination on the slope video stream data in sequence to obtain standardized slope data;
[0005] Super-resolution reconstruction of standardized slope data was performed, and micro-motion video stream data was obtained through a difference amplification mechanism. The micro-motion video stream data was then analyzed to obtain micro-motion field sequence data.
[0006] Micro-motion component sequence data is generated from micro-motion field sequence data, and coherence map features, state features and dynamic characteristics are obtained based on the micro-motion component sequence data.
[0007] Risk quantification data is obtained based on coherence map features, state features, and dynamic characteristics, and through a slope assessment model.
[0008] Specifically, the preprocessing operations of performing correction mapping, normalization, and ROI determination on the slope video stream data in sequence include:
[0009] Set control points, obtain transformation matrices and distortion correction parameters based on the control points and slope video stream data respectively, and perform correction mapping on the slope video stream data through transformation matrices and distortion correction parameters;
[0010] Determine whether a reference template exists, and based on the determination result, choose to use white balance or adaptive histogram to perform color correction on the slope video stream data;
[0011] The slope video stream is initially divided based on semantic segmentation, and the ROI is determined through secondary segmentation.
[0012] Specifically, the process of performing super-resolution reconstruction on standardized slope data and obtaining micro-motion video stream data through a difference amplification mechanism, and analyzing the micro-motion video stream data to obtain micro-motion field sequence data, includes:
[0013] A super-resolution reconstruction model is selected, and the standardized slope data is reconstructed based on the super-resolution reconstruction model to obtain slope reconstruction data. Detail optimization is applied during the reconstruction process.
[0014] Based on slope reconstruction data, classified data groups were obtained, and time-frequency domain analysis was used to obtain differential motion data. The differential motion data was then amplified to obtain micro-motion video stream data.
[0015] Standardized slope data is applied to micro-motion video stream data to perform motion analysis on the micro-motion video stream data in order to obtain micro-motion field sequence data.
[0016] Specifically, the process of obtaining classified data groups based on slope reconstruction data, simultaneously using time-frequency domain analysis to obtain differential motion data, and amplifying the differential motion data to obtain micro-motion video stream data includes:
[0017] Semantic segmentation was performed on the slope reconstruction data to obtain categorized data groups including Class I, Class II, and Class III data.
[0018] Dominant frequency data and low frequency data are obtained based on Class I and Class II data respectively. Differential motion data are extracted from the dominant frequency data and low frequency data, while Class III data is ignored.
[0019] A first amplification factor and a second amplification factor are applied to the differential motion data respectively to obtain micro-motion video stream data.
[0020] Specifically, the step of generating micro-motion component sequence data from micro-motion field sequence data, and obtaining coherence map features, state features, and dynamic characteristic features based on the micro-motion component sequence data, includes:
[0021] Based on the micro-motion field sequence data, micro-motion component sequence data containing slope aspect component sequence data and intensity component data are obtained, and coherence analysis is performed to obtain coherence map characteristics;
[0022] Spectral analysis is performed based on micro-motion component sequence data, and an energy entropy evolution mechanism is introduced to obtain state characteristics.
[0023] Dynamic characteristics are obtained based on intensity component data and input excitation signal.
[0024] Specifically, the step of acquiring micro-motion component sequence data containing slope aspect component sequence data and intensity component data based on micro-motion field sequence data, and performing coherence analysis to obtain coherence map characteristics, includes:
[0025] The micro-dynamic field sequence data is extracted to obtain the aspect component sequence data and intensity component data respectively, and comparison data is defined at the same time;
[0026] The slope aspect coherence coefficient is obtained by comparing the data with the slope aspect component sequence data, and the intensity coherence coefficient is obtained by comparing the data with the intensity component data.
[0027] A coherence map is established based on the aspect coherence coefficient and the intensity coherence coefficient, and coherence map features are extracted from the coherence map.
[0028] Specifically, the spectral analysis based on micro-motion component sequence data, while introducing an energy entropy evolution mechanism to obtain state characteristics, includes:
[0029] Frequency domain transformation is performed on the aspect component sequence data and intensity component data to obtain the energy spectrum. The frequency band is then divided based on the energy spectrum to obtain the classification frequency band.
[0030] The full-band energy entropy is obtained based on the energy spectrum, and the energy ratio of the classification band is obtained by comparing the energy spectrum with the classification band. At the same time, directional analysis and mutation analysis are performed to obtain evolutionary characteristic data.
[0031] The full-band energy entropy, the energy ratio of the classified band, and the evolutionary characteristic data are integrated into state characteristics.
[0032] Specifically, the acquisition of dynamic characteristic features based on intensity component data and input excitation signal includes:
[0033] Define the input excitation signal, and simultaneously define the output response signal based on the intensity component data;
[0034] Estimating the frequency response function based on the input excitation signal and the output response signal;
[0035] The frequency response function is extracted to obtain the dominant resonant frequency, resonant peak gain, half-power bandwidth, and average gain, respectively.
[0036] The dominant resonant frequency, resonant peak gain, half-power bandwidth, and average gain are integrated into dynamic characteristic features.
[0037] Specifically, risk quantification data is obtained based on coherence map characteristics, state characteristics, and dynamic characteristics, and through a slope assessment model, including:
[0038] A pre-trained dataset is constructed, and a slope assessment model is obtained based on the pre-trained dataset and meta-learning. Coherence map features, state features, and dynamic characteristic features are input into the slope assessment model. Dimensionality reduction assessment is performed through the slope assessment model to generate risk quantification data. The slope assessment model includes a dimensionality reduction model and an assessment model.
[0039] In conjunction with the second aspect of the present invention, a safety risk assessment platform based on dangerous slope deformation identification is also provided, the platform comprising:
[0040] The preprocessing sub-platform is used to acquire slope video stream data and perform preprocessing operations to obtain standardized slope data.
[0041] The reconstruction sub-platform is used for super-resolution reconstruction of standardized slope data.
[0042] The amplification sub-platform is used to acquire micro-motion video stream data and simultaneously acquire micro-motion field sequence data through a differential amplification mechanism.
[0043] The extraction sub-platform is used to perform feature extraction to obtain coherence map features, state features, and dynamic characteristic features;
[0044] The quantitative sub-platform is used to obtain risk quantification data through slope assessment models.
[0045] In this invention, standardized slope data is first obtained through slope video stream data. Then, super-resolution reconstruction is performed on the slope data, and micro-motion video stream data is obtained through a difference amplification mechanism. Simultaneously, micro-motion field sequence data is obtained based on analysis. Combined with physically guided motion visual amplification, minute surface micro-motion responses of the slope are extracted. The deep stability of the slope can be inverted through the surface micro-motion responses. Then, feature extraction is performed to obtain coherence map features, state features, and dynamic characteristic features. Finally, risk quantification data is obtained through a meta-learning-driven slope assessment model to achieve early passive and large-scale risk assessment of slope morphological instability. Once risk information is available, more targeted and accurate monitoring can be carried out to improve safety. Attached Figure Description
[0046] Figure 1 This is a flowchart of the safety risk assessment method based on dangerous slope deformation identification according to the present invention.
[0047] Figure 2 This is a schematic diagram of the safety risk assessment system based on dangerous slope deformation identification of the present invention. Detailed Implementation
[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0049] Combined with appendix Figure 1 As shown, the first aspect of the present invention provides a safety risk assessment method based on dangerous slope deformation identification, the method comprising:
[0050] S1: Acquire slope video stream data, and perform preprocessing operations such as correction mapping, color correction, and ROI determination on the slope video stream data in sequence to obtain standardized slope data.
[0051] In the specific implementation of this invention, the preprocessing operations of sequentially performing correction mapping, color correction, and ROI determination on the slope video stream data include:
[0052] S11: Set control points, obtain transformation matrix and distortion correction parameters based on control points and slope video stream data respectively, and perform correction mapping on slope video stream data through transformation matrix and distortion correction parameters;
[0053] S12: Determine whether a reference template exists, and based on the determination result, select to use white balance or adaptive histogram to perform color correction on the slope video stream data;
[0054] S13: The slope video stream is initially divided based on semantic segmentation, and the ROI is determined through secondary segmentation.
[0055] In some possible embodiments, such as slope monitoring along a mountain highway, multiple cameras are already distributed along the highway, covering the slope area to be monitored. After acquiring real-time or historical video streams through the cameras, i.e., after acquiring the slope video stream data, it is synchronized along the timeline. Since the cameras are mostly installed on utility poles or gantry frames, their viewing angles are mostly oblique, which can cause perspective distortion in the slope images. Therefore, it is necessary to first correct and map the slope video stream. At least four easily identifiable fixed points with known geographical coordinates are selected within the camera's field of view as control points. These control points can be specific... The corner points of traffic signs or pre-placed simple markers can be accurately measured using a total station or differential GPS to obtain their precise three-dimensional coordinates. In a clear frame of the slope video stream, the pixel coordinates corresponding to the above control points are identified and marked. Using the correspondence between three-dimensional coordinates and two-dimensional pixel coordinates, the perspective transformation matrix and lens distortion correction coefficient for the camera are calculated through linear transformation. In subsequent real-time video processing, the above-solved transformation matrix and distortion correction coefficient are applied to each frame of input image to project the image that was originally distorted by perspective onto an imaging plane facing the slope.
[0056] After the correction mapping is completed, in order to eliminate the influence of changes in lighting and differences in camera parameters on the slope video stream, a neutral color region with a constant color over a long period of time is found in the camera's field of view, such as a light gray traffic facility base or a rock of a specific color, and used as a reference template. If such a reference template exists, the pixel color value of the region is extracted in each frame of the slope video stream, and its deviation from standard white (R=G=B) is calculated. Then, the gain coefficient for color correction of the entire image can be derived and applied to the entire frame image, so that while restoring the neutral color in the reference region, the hue is enhanced to maintain consistency. If there is no reference template, image-based statistical methods can be applied to process the slope video stream. The brightness channel histograms of multiple consecutive frames in the slope video stream are dynamically matched to adjust their mean and variance to a preset standard range, so as to reduce the overall brightness fluctuations caused by sunrise, sunset or cloudy / sunny weather. For the color channel, scale-invariant feature transformation can be used to maintain the relativity of color relationships.
[0057] Next, the corrected slope video stream data is processed using a semantic segmentation model. For example, a pre-trained DeepLabv3+ model can be used. The DeepLabv3+ model can classify pixels in the image into categories such as sky, road, slope vegetation, and slope rock and soil. This automatically extracts the pixel regions of slope vegetation and slope rock and soil, and merges them to form a binary slope main mask. Areas outside the slope main mask will be temporarily ignored in subsequent processing. Within the slope main mask, a regular analysis grid is automatically overlaid according to preset standards. For example, the slope area is divided into Nm × Nm grid cells. Each grid cell will serve as the basic spatial unit for subsequent feature extraction and other operations. Technicians can fine-tune the grid boundaries to ensure that they are aligned with the terrain features as much as possible. The above operations are used for ROI determination preprocessing.
[0058] The above preprocessing operation is performed on all slope video stream data acquired by cameras to obtain frames of all slope video stream data at the same time after preprocessing. These frames are then encapsulated into a data object, which should include each preprocessed image data, the corresponding slope main mask, grid information, camera number, and approximate mapping relationship after correction. All data objects are arranged in chronological order to obtain standardized slope data.
[0059] S2: Perform super-resolution reconstruction on the standardized slope data and obtain micro-motion video stream data through a difference amplification mechanism. Analyze the micro-motion video stream data to obtain micro-motion field sequence data.
[0060] In the specific implementation of this invention, the step of performing super-resolution reconstruction of standardized slope data and obtaining micro-motion video stream data through a difference amplification mechanism, and analyzing the micro-motion video stream data to obtain micro-motion field sequence data includes:
[0061] S21: Select a super-resolution reconstruction model, and reconstruct the slope standardized data based on the super-resolution reconstruction model to obtain slope reconstruction data, in which detail optimization is applied during the reconstruction process;
[0062] S22: Based on slope reconstruction data, classify data groups are obtained, and time-frequency domain analysis is used to obtain differential motion data. The differential motion data is then amplified to obtain micro-motion video stream data.
[0063] S23: Apply the standardized slope data to the micro-motion video stream data, perform motion analysis on the micro-motion video stream data, and obtain micro-motion field sequence data.
[0064] Furthermore, the process of obtaining classified data groups based on slope reconstruction data, simultaneously using time-frequency domain analysis to obtain differential motion data, and amplifying the differential motion data to obtain micro-motion video stream data includes:
[0065] S221: Perform semantic segmentation on the slope reconstruction data to obtain a categorized data set including Class I, Class II and Class III data;
[0066] S222: Obtain dominant frequency data and low frequency data based on Class I and Class II data respectively, and extract differential motion data through dominant frequency data and low frequency data, while ignoring Class III data;
[0067] S223: Apply the first amplification factor and the second amplification factor to the differential motion data respectively to obtain micro-motion video stream data.
[0068] In some possible embodiments, the video source used to analyze the slope is usually from an existing wide-angle camera deployed at a considerable distance. However, such long-distance, wide-angle shooting is prone to problems such as loss of detail and large motion quantization errors. If the low-resolution video is directly enlarged, it is equivalent to enlarging the blur, which often makes it impossible to extract real and useful minute motion signals. Therefore, super-resolution reconstruction is used here to reconstruct the standardized slope data.
[0069] To achieve high-quality super-resolution reconstruction, and considering the small displacements between adjacent frames in continuous video, which often contain complementary high-frequency details, a recursive or sliding window-based video super-resolution grid structure super-resolution reconstruction model can be selected. When processing the current frame in slope standardization data, this structure can utilize the information of the current frame and simultaneously combine it with cross-frame information from its adjacent previous and subsequent frames for alignment and fusion reconstruction. Compared with single-image reconstruction methods, it can more effectively reconstruct the details of the current frame.
[0070] When training the model, it is possible to use downsampled low-resolution video clips from slope inspection videos and close-up ground-shot videos, along with their corresponding high-resolution ground truth values, to train the model. This will enable it to better recover typical slope features such as vegetation texture and surface roughness of soil and rock. In addition to minimizing pixel value errors during training, perceptual loss, texture loss, and edge enhancement detail optimization should also be introduced. Perceptual loss can encourage the reconstructed image to be similar to the real high-resolution image in deep semantic features, improving the realism of the reconstructed vegetation and rocks. Texture loss can focus on maintaining the continuity of local textures to avoid generating overly smooth or false noise textures. After reconstruction, an adaptive sharpening filter can be selectively used for edge enhancement, which only moderately enhances the detected edge areas, making the contours of terrain lines and potential cracks clearer visually and in terms of gradient.
[0071] In one possible embodiment, a short time-series segment containing the current frame and N frames before and after it is extracted from the image sequence of the slope normalized data and centered on the current frame to be reconstructed. This segment is then input into the super-resolution reconstruction model, where N is typically 2 or 3. The super-resolution reconstruction model first performs optical flow estimation on each frame in the short time-series and aligns the features of the surrounding frames to the coordinate system of the central frame. Then, the aligned multi-frame features are fused with the features of the central frame in a depth feature channel to reconstruct lost details. For example, if there is a blurry green area in the low-resolution input, the model can reconstruct the outline and veins of a single leaf based on its micro-change pattern across frames. Finally, the fused features are passed through an upsampling channel to generate a high-resolution image of the current central frame, i.e., the slope reconstruction data. The reconstruction ratio can be determined according to the actual implementation requirements, such as reconstructing 1080p to 4K.
[0072] Based on the slope reconstruction data, the minute motion components related to slope stability are separated from the slope reconstruction data. Semantic segmentation is performed on the slope reconstruction data, and the model accuracy should be higher than that of the semantic segmentation model used in the preprocessing operation. In this way, the pixels in each frame of the slope reconstruction data are divided into physical regions, namely Class I, Class II, and Class III data. Class I data represents areas covered by flexible vegetation, such as shrubs and grasslands. These areas can be modeled as a large set of damped harmonic oscillators. The dominant motion is high-frequency, small-amplitude periodic oscillation driven by wind. The motion frequency is related to wind speed and the damping characteristics of the vegetation itself. Class II data represents rigid exposed areas, such as rock slopes and debris accumulation slopes. These areas can be modeled as a superposition of rigid bodies and local creep bodies. The dominant motion is low-frequency and quasi-static translation or slow deformation driven by gravity, thermal expansion and contraction, or internal stress. Class III data represents irrelevant background areas, such as the sky and road surface. These areas can be completely ignored.
[0073] Type labels are applied to each pixel in each frame of the slope reconstruction data, allowing for different time-domain and frequency-domain analysis methods for different data types. For one type of data, several feature points are selected, and their brightness or feature changes throughout the video are tracked to form a time series. Fourier transform is applied to this time series to identify the 1-3 main frequency data points with the most concentrated energy. These main frequency data points typically correspond to the fundamental frequency or low-order harmonics of wind-induced vibrations. Then, a narrow-bandpass filter is set for each type of data based on the estimated main frequency data. For example, if the main frequency component is 2.5 Hz, the filter can be set to 2.0 to 3.0 Hz. The purpose of the narrow-bandpass filter is to selectively extract the signals generated by the environment (e.g., wind-induced vibrations). The data is obtained by filtering out forced vibrations directly caused by wind, which characterize the collective properties of vegetation, and filtering out high-frequency noise and slow physiological growth changes at low frequencies. For the second type of data, the motion is usually extremely slow and at very low frequencies, typically below 0.5 Hz, possibly close to DC. Therefore, a high-pass filter is used to remove non-motion artifacts such as long-term drift caused by slow changes in sunlight, thereby obtaining low-frequency data. Then, a low-frequency filter with a passband of 0.01 to 0.05 Hz is set for the second type of data to filter the low-frequency data, thereby obtaining the second type of motion data to capture possible creep. The first and second types of motion data are then integrated into the second type of motion data.
[0074] The first differential motion data in the differential motion data is multiplied by a first amplification factor. This factor can be dynamically adjusted according to the wind speed. For example, a larger factor is used in light winds to enhance the first differential motion data, and a smaller factor is used in strong winds to avoid motion estimation saturation. The second differential motion data in the differential motion data is multiplied by a second amplification factor. Since the amplitude of the second differential motion data is smaller, the second amplification factor should be much larger than the first amplification factor. The amplified differential motion data is then combined with the unamplified three types of data. That is, for each pixel, its original intensity value is added to the corresponding amplified motion change. The above operation is performed on each frame in the slope reconstruction data to obtain amplified micro-motion video stream data. In the micro-motion video stream data, the first type of data should show clearly amplified forced vibration, while the second type of data should show slow overall offset or local distortion.
[0075] The grid from the slope standardization data is overlaid onto the micro-motion video stream data. For each grid cell, the block matching method is used to calculate the overall motion vector of the cell from the current frame to the next frame. Taking a certain grid cell as an example, feature points are selected within the cell grid of the current frame, or the grid is regarded as an image block. In the next frame, the position that best matches the feature point or image block is found within a predefined search range. The offset between the two positions in pixels can be converted into a two-dimensional plane vector by combining the mapping relationship in the slope standardization data. This vector contains magnitude and direction. For the first type of data, since its motion is high-frequency vibration, the focus can be on capturing its periodic instantaneous velocity direction. For the second type of data, the focus can be on capturing its slow cumulative displacement trend.
[0076] Simultaneously, a motion intensity is calculated for each grid cell. The motion intensity is the root mean square value or amplitude of the differential motion data of all pixels in the corresponding grid cell after being amplified by S223. It is used to reflect the activity level of small motions in the region. The above operation is repeated for each pair of consecutive frames (frame t and frame t+1) in the video sequence to generate a time-varying motion vector sequence and motion intensity sequence for each grid cell. Then, the obtained grid cell information can be encapsulated into micro-motion field sequence data.
[0077] S3: Generate micro-motion component sequence data from micro-motion field sequence data, and obtain coherence map features, state features and dynamic characteristic features based on the micro-motion component sequence data.
[0078] In a specific implementation of this invention, the step of generating micro-motion component sequence data from micro-motion field sequence data, and obtaining coherence map features, state features, and dynamic characteristic features based on the micro-motion component sequence data, includes:
[0079] S31: Based on the micro-motion field sequence data, obtain micro-motion component sequence data containing slope aspect component sequence data and intensity component data, and perform coherence analysis to obtain coherence map characteristics;
[0080] S32: Spectral analysis is performed based on micro-motion component sequence data, and an energy entropy evolution mechanism is introduced to obtain state characteristics;
[0081] S33: Obtain dynamic characteristic features based on intensity component data and input excitation signal.
[0082] In a specific implementation of this invention, the step of acquiring micro-motion component sequence data containing slope aspect component sequence data and intensity component data based on micro-motion field sequence data, and performing coherence analysis to obtain coherence spectrum features, includes:
[0083] S311: Extract the micro-dynamic field sequence data to obtain the aspect component sequence data and intensity component data respectively, and define comparison data;
[0084] S312: Obtain the slope aspect coherence coefficient by comparing the data with the slope aspect component sequence data, and obtain the intensity coherence coefficient by comparing the data with the intensity component data;
[0085] S313: Establish a coherence map based on the aspect coherence coefficient and intensity coherence coefficient, and extract coherence map features through the coherence map.
[0086] In one possible embodiment, the aspect component sequence data and intensity component data of each grid cell within a selected analysis period are extracted from the micro-motion field sequence data. For the aspect component sequence data, the motion vector sequence of each grid cell is projected onto the main slope direction, such as the downhill direction, according to the overall dip of the slope, thereby obtaining an aspect component sequence to represent the small motion of the grid cell along the most slippery direction. For the intensity component data, the motion intensity sequence in the micro-motion field sequence data is directly selected. Then, in order to analyze the spatial correlation, it is necessary to define the regions to be compared, i.e., the comparison data. Each grid cell is paired with the grid cells in the four directly adjacent directions (up, down, left, and right) to detect local small-scale inconsistencies. Then, according to the topography of the slope, two grid cells that are spatially separated but located on the same inferred path are selected along possible potential slip paths, such as along a suspected arc or plane, and paired. This can be used to detect overall inconsistencies over a larger area.
[0087] For each predefined pair of regions, the aspect coherence coefficient and intensity coherence coefficient are calculated separately. For the aspect coherence coefficient, the aspect component sequence data of the two regions are first transformed in the frequency domain to obtain the amplitude and phase information at each frequency component. Then, the normalized average value of the amplitude spectrum product of the two sequences at each frequency point within the selected frequency band is calculated. This is usually set to the dominant frequency band of the environmental excitation, such as the vibration frequency band determined for a type of data in S222. The aspect coherence coefficient can then be obtained. The closer the value is to 1, the more synchronous and waveform-similar the minute movements of the two regions are within the frequency band; the closer the value is to 0, the more unrelated or asynchronous the movements are. For the intensity coherence coefficient, the linear correlation coefficient can be directly calculated from the intensity component data of the two regions. This coefficient reflects whether the other region is synchronous when the minute movements of one region are intense. Its value is in the range of -1 to 1, with positive values indicating positive correlation and negative values indicating negative correlation. The larger the absolute value, the stronger the synchronicity of the changes. For stable slopes, adjacent regions should show a positive correlation under the same conditions.
[0088] For each grid cell, the aspect coherence coefficient and intensity coherence coefficient calculated with all adjacent regions are used as the attribute values of the grid cell. Then, two contour maps or heat maps, i.e., coherence maps, can be constructed based on the aspect coherence coefficient and intensity coherence coefficient. Then, a dynamic threshold is set on the map, for example, the coherence value is more than one standard deviation below the average value of all grid cells. This identifies low coherence grid cell clusters that meet the threshold condition. The low coherence grid cell clusters are connected and refined through morphological processing to obtain continuous low coherence boundaries. The low coherence boundaries may physically correspond to existing cracks, structural surfaces inside the rock and soil mass, or developing potential shear zones, because they block the transmission of vibration waves or coordinated motion. The coherence map and its derived low coherence boundaries are encapsulated as coherence map features.
[0089] In some possible embodiments, the coherence coefficients calculated locally and globally can also be compared. If the regions on both sides of a boundary have lost coordination at a small scale, and also show disharmony on the overall path across the boundary, then the credibility of the boundary as a potential slip surface indicator is increased.
[0090] Understandably, S31 can extract and quantify the spatial correlation characteristics that explain the structural integrity and connectivity of slopes, thereby identifying the spatial location of potential slip surfaces at an early stage without relying on absolute displacement thresholds or physical intrusion.
[0091] In the specific implementation of this invention, the step of performing spectral analysis based on micro-motion component sequence data, while simultaneously introducing an energy entropy evolution mechanism to obtain state characteristics, includes:
[0092] S321: Perform frequency domain transformation on the aspect component sequence data and intensity component data to obtain the energy spectrum, and divide the frequency bands through the energy spectrum to obtain the classification frequency bands;
[0093] S322: Obtain full-band energy entropy based on energy spectrum, obtain the energy ratio of classification bands through energy spectrum and classification bands, and simultaneously perform directional analysis and mutation analysis to obtain evolutionary characteristic data;
[0094] S323: Integrate full-band energy entropy, classified band energy ratio, and evolutionary characteristic data into state characteristics.
[0095] In one possible embodiment, a Fourier transform is performed on the aspect component sequence data and the intensity component data to convert the time-series signal into a frequency domain signal, thereby obtaining the energy spectrum. The energy spectrum represents the distribution of signal energy at different frequencies. Then, frequency bands can be divided based on the energy spectrum. Instead of uniform frequency band division, the frequency bands are divided into three characteristic frequency bands according to the physical source of the signal. The first is the dominant frequency band, which corresponds to the vibration frequency band determined for the first type of data in S222. The second is the low-frequency frequency band, which corresponds to the frequency band set for the second type of data in S222. The third is the background frequency band, which covers the remaining frequency range.
[0096] To quantify the concentration or dispersion of energy distribution, energy entropy is introduced. A higher entropy value indicates a more uniform but chaotic energy distribution across different frequencies; a lower entropy value indicates that energy is more concentrated and ordered at a few frequencies. The normalized probability of the energy spectrum across the entire frequency range is calculated by dividing the energy value of each characteristic frequency band by the total energy of all frequency points, thus obtaining the energy proportion of the characteristic frequency band. Using the definition formula of information entropy and based on the energy proportion of the characteristic frequency band, a full-band energy entropy is calculated. The full-band energy entropy is lower when in a stable state, and higher when internal damage causes resonance or disordered responses at multiple frequencies. After obtaining the full-band energy entropy, the ratio of the dominant frequency band energy to the total energy is calculated. A decrease in this ratio may indicate a reduction in the energy input efficiency of environmental excitation. Simultaneously, the ratio of the low-frequency band to the total energy is calculated. An increase in this ratio may indicate intensified static deformation activity. The ratios calculated by both methods together form the classification frequency band energy ratio.
[0097] Next, the directional distribution of the motion vectors of the grid cells throughout the entire time window is analyzed, and the average direction of their directional angles is calculated. The average direction is compared with the historical average direction to obtain the angular deviation. For bare areas, if the angular deviation continues to increase and points outward from the slope, it indicates potential slippage. For vegetated areas, abnormal changes in the angular deviation may reflect a weakening of the soil stabilization effect of the vegetation roots. Then, the time window is further divided into shorter sub-windows, and the energy ratio of the classification band and the energy entropy of the whole band are tracked within the sub-window. At the same time, the mutation point detection algorithm is applied to identify the moments when the statistical characteristics of the energy ratio of the classification band and the energy entropy of the whole band change significantly. For example, if the energy entropy of the whole band suddenly jumps and fails to recover, such mutations correspond to external disturbances or changes in the internal state, such as changes starting from a heavy rainfall. The mutation markers and the angular deviation together constitute the evolutionary feature data, and the energy entropy of the whole band, the energy ratio of the classification band, and the evolutionary feature data are integrated into state features.
[0098] Understandably, S32 can quantify the orderliness and disorder of the micro-motion signal of each grid cell and capture its subtle shifts in motion, thus providing characteristic indicators for judging whether the region deviates from a stable state.
[0099] In a specific implementation of this invention, the step of obtaining dynamic characteristic features based on intensity component data and input excitation signal includes:
[0100] S331: Define the input excitation signal and define the output response signal based on the intensity component data;
[0101] S332: Estimating the frequency response function based on the input excitation signal and the output response signal;
[0102] S333: Extract the frequency response function to obtain the dominant resonant frequency, resonant peak gain, half-power bandwidth, and average gain respectively;
[0103] S334: Integrates the dominant resonant frequency, resonant peak gain, half-power bandwidth, and average gain into a dynamic characteristic feature.
[0104] In one possible embodiment, the measured horizontal wind speed time series is used as the input excitation signal, and the wind speed vector is decomposed into components perpendicular to the slope direction according to the macroscopic orientation of the slope. The intensity component data is used as the corresponding output signal. Then, the autopower spectrum of the input signal, the autopower spectrum of the output signal, and the cross-power spectrum between the input and output are calculated respectively. At each frequency point, the value of the cross-power spectrum is divided by the value of the input autopower spectrum to obtain a complex estimate. The sequence of these complex estimates is the estimated frequency response function, which includes the gain and phase delay information of the excitation at that frequency. At the same time, the condensation function values of the input and output signals at each frequency point are calculated. This value is in the range of 0-1 and is used to measure what proportion of the output signal at a certain frequency point is caused by the corresponding input signal. A condensation function value close to 1 indicates that the frequency response function estimation at the frequency point is reliable, and a condensation function value close to 0 indicates that the response at the frequency point may mainly come from other unmeasured excitations or noise. The dominant resonant frequency, resonant peak gain, half-power bandwidth, and average gain are extracted from the frequency response function curves with condensation functions higher than the threshold.
[0105] The dominant resonant frequency is the frequency corresponding to the most significant peak on the frequency response function curve. For vegetated areas, this frequency reflects the inherent vibration frequency of the vegetation and soil. A decrease in this frequency may indicate a reduction in effective stiffness. The resonant peak gain is the gain corresponding to the dominant resonant frequency point. Changes in this value reflect changes in damping characteristics. The half-power bandwidth is the frequency bandwidth corresponding to a certain percentage decrease (e.g., a 3dB decrease) in the resonant peak gain. This bandwidth is inversely proportional to the damping ratio. A wider bandwidth usually indicates increased damping, and vice versa. The average gain is the average value of the gain within the dominant frequency band. The average gain reflects changes in sensitivity. Integrating the dominant resonant frequency, resonant peak gain, half-power bandwidth, and average gain into dynamic characteristic features, it can be understood that the dynamic transmission characteristics of each area of the slope can be non-invasively estimated using random excitation (horizontal wind speed time series) through S33.
[0106] S4: Based on coherence map features, state features and dynamic characteristics, risk quantification data is obtained through slope assessment models.
[0107] In the specific implementation of this invention, the step of obtaining risk quantification data based on coherence map features, state features, and dynamic characteristic features through a slope assessment model includes:
[0108] S41, construct a pre-trained dataset, obtain a slope assessment model based on the pre-trained dataset and using meta-learning, input coherence map features, state features and dynamic characteristic features into the slope assessment model, perform dimensionality reduction assessment through the slope assessment model to generate risk quantification data, wherein the slope assessment model includes a dimensionality reduction model and an assessment model.
[0109] In some possible embodiments, cooperation can first be established with organizations such as traffic management departments, nature reserves, or research institutions to obtain historical monitoring videos of slopes along highways and railways in different regions. These historical videos should have a long time span and include different seasons and weather conditions. Then, geotechnical engineering simulation software can be used to establish various typical slope geological models, simulating different environmental excitations such as wind loads and minor seismic vibrations. The vibration response at slope surface points can be calculated and rendered as a simulation. Parameters can be systematically changed, such as parameters simulating changes in stiffness and strength, to generate a stable slope model. Simulation data of the entire process from instability to stability is obtained. At the same time, visual technology is used to overlay different vegetation models, rock textures, and artificial structures on a 3D terrain model, and different lighting, rain and snow conditions are simulated to generate slope videos. Then, the micro-movements are simulated through algorithms. Each collected or generated video data segment is associated with tags and metadata. For real historical data, it is labeled as stable or unstable segments based on whether there are recorded landslide events. For simulation data, it is directly labeled according to the input parameters. The metadata includes information such as geographical location, dominant geological type, average slope, vegetation coverage, and main excitation type.
[0110] Next, the task structure and task generation strategy can be defined separately. The task structure includes a support set and a query set. The support set simulates the very limited amount of labeled data obtainable on the new slope. For example, it may only contain video clips of the slope in a stable state for a few hours, but unstable clips may also be included. The number of support sets should be limited, for example, to only 5-10 short video clips. The query set simulates other data on the same slope that the model needs to evaluate after rapid adaptation. It is used to evaluate the model's performance after adaptation during the meta-training phase. The task generation strategy includes three types of tasks. One type of task consists of all data from the same real slope, the same camera, and a continuous period of time, which is treated as a scene pool. From this pool, a scene is randomly selected. A small subset of segments serves as the support set, while another set of non-overlapping segments serves as the query set, thus forming a task to ensure the consistency of data distribution within the task. Different tasks also exhibit differences. The second type of task involves treating a group of simulation models with similar geological parameters but different specific values as a task cluster. Within the same task, the support set and query set come from simulations with similar geological parameter settings but different excitation conditions. Different tasks come from simulation models with significantly different parameter settings, thereby simulating adaptation across different geological conditions. The third type of task is a hybrid task; for example, the support set of a task might be a video of a real slope, while the query set might be another simulation data set with similar geological conditions, thus enhancing the model's generalization ability.
[0111] All collected real videos and simulation data are processed through S1 to S3 to generate coherence map features, state features, and dynamic characteristic features, thereby generating a pre-training dataset. At this point, adaptive training can be performed based on mature meta-learning techniques, such as the method of the invention patent "Continuous Learning Method and System Based on Generative Model and Meta-learning Optimization Method" with application number CN201910899856.X, to generate a slope assessment model.
[0112] The slope assessment model includes a dimensionality reduction model and an evaluation model. The dimensionality reduction model employs a stacked sparse autoencoder architecture, comprising an encoder and a decoder. The encoder consists of multiple fully connected layers, with non-linear activation functions used between layers to fuse and map high-dimensional global feature vectors to low-dimensional feature vectors. The decoder structure is symmetrical to the encoder, outputting a low-dimensional state vector. An attention network is also included between the encoder and decoder layers; its input is the current partial features or intermediate representations, and its output is a weight vector. This weight vector allows the dimensionality reduction model to focus more on currently relevant feature sources. The evaluation model employs a temporal convolutional network architecture, including a feature extraction layer and a dual-branch output layer. The feature extraction layer performs deep feature extraction on the input state vector of a fixed time length, capturing its contents. It includes short-term fluctuations, cyclical patterns, and long-term trends. For example, it identifies a coordinated pattern of pulses that first decrease and then increase before instability, usually appearing in several characteristic dimensions. Then, it calculates the dependency between states at different time points through causal convolution to understand whether the current state is part of a long-term stable trend or an abnormal turning point. The dual-branch output layer is the last layer of the evaluation model, which includes a probability branch and a directional branch. The probability branch outputs a risk probability value in the range of 0-1. This value represents the probability that the slope will become unstable within a preset time window based on the current and historical states. The directional branch outputs a risk label, which can be represented by 0, 1, and -1 to characterize different trends. For example, -1 can represent a decrease in risk, 0 can represent stability, and 1 can represent an increase in risk.
[0113] Specifically, the coherence map features, state features, and dynamic characteristic features acquired by S3 are input into the slope assessment model. The slope assessment model arranges all features according to the same grid cell order. For each grid cell, its corresponding features are concatenated into a cell feature vector in a predetermined order. Then, the cell feature vectors are further concatenated according to the spatial order of the grid cells to generate a global feature vector. Each feature dimension in the global feature vector is standardized, and then dimensionality reduction is performed using a dimensionality reduction model to output a low-dimensional and dense state vector. Finally, the state vectors of a fixed time period are... The vector input is fed into the evaluation model, which outputs a risk probability value and a risk label. The risk probability value is used as the base, and combined with the risk label, the final risk quantification data is output. For example, if the risk probability value is 0.4 and the risk label is -1, then 0.4 can be multiplied by a coefficient greater than 1, such as 1.5, to finally output a risk quantification data of 1.5 × 0.4 × 100 = 60%, which is within the probability range of 0-100. The risk quantification data can be used for early risk warning of slope instability. Once the risk information is available, more targeted and accurate monitoring can be carried out to improve safety.
[0114] Combined with appendix Figure 2 As shown, a second aspect of the present invention provides a safety risk assessment platform based on dangerous slope deformation identification, the platform comprising:
[0115] The preprocessing sub-platform 11 is used to acquire slope video stream data and perform preprocessing operations to obtain standardized slope data.
[0116] Reconstruction Sub-Platform 12 is used for super-resolution reconstruction of standardized slope data.
[0117] The amplification sub-platform 13 is used to acquire micro-motion video stream data and simultaneously acquire micro-motion field sequence data through a differential amplification mechanism.
[0118] Sub-platform 14 is used for feature extraction to obtain coherence map features, state features and dynamic characteristic features.
[0119] The quantitative sub-platform 15 is used to obtain risk quantification data through the slope assessment model.
[0120] In the specific implementation of this invention, the specific implementation methods of the platform item can be referred to the implementation methods of the above-mentioned method item, and will not be repeated here.
[0121] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.
[0122] Furthermore, the safety risk assessment method and platform based on dangerous slope deformation identification provided in the embodiments of the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A safety risk assessment method based on dangerous slope deformation identification, characterized in that, The method includes: S1: Acquire slope video stream data, and perform preprocessing operations such as correction mapping, color correction and ROI determination on the slope video stream data in sequence to obtain standardized slope data. S2: Perform super-resolution reconstruction on the standardized slope data and obtain micro-motion video stream data through the difference amplification mechanism; analyze the micro-motion video stream data to obtain micro-motion field sequence data. S3: Generate micro-motion component sequence data from micro-motion field sequence data, and obtain coherence map features, state features and dynamic characteristic features based on the micro-motion component sequence data; S4: Based on coherence map features, state features and dynamic characteristics, risk quantification data is obtained through slope assessment models.
2. The safety risk assessment method based on dangerous slope deformation identification according to claim 1, characterized in that, The preprocessing operations performed on the slope video stream data, including correction mapping, normalization, and ROI determination, include: Set control points, obtain transformation matrices and distortion correction parameters based on the control points and slope video stream data respectively, and perform correction mapping on the slope video stream data through transformation matrices and distortion correction parameters; Determine whether a reference template exists, and based on the determination result, choose to use white balance or adaptive histogram to perform color correction on the slope video stream data; The slope video stream is initially divided based on semantic segmentation, and the ROI is determined through secondary segmentation.
3. The safety risk assessment method based on dangerous slope deformation identification according to claim 1, characterized in that, The process involves super-resolution reconstruction of standardized slope data and acquisition of micro-motion video stream data via a difference amplification mechanism. Analysis of the micro-motion video stream data yields micro-motion field sequence data, including: A super-resolution reconstruction model is selected, and the standardized slope data is reconstructed based on the super-resolution reconstruction model to obtain slope reconstruction data. Detail optimization is applied during the reconstruction process. Based on slope reconstruction data, classified data groups were obtained, and time-frequency domain analysis was used to obtain differential motion data. The differential motion data was then amplified to obtain micro-motion video stream data. Standardized slope data is applied to micro-motion video stream data to perform motion analysis on the micro-motion video stream data in order to obtain micro-motion field sequence data.
4. The safety risk assessment method based on dangerous slope deformation identification according to claim 3, characterized in that, The process involves acquiring categorized data groups based on slope reconstruction data, simultaneously using time-frequency domain analysis to obtain differential motion data, and amplifying the differential motion data to obtain micro-motion video stream data, including: Semantic segmentation was performed on the slope reconstruction data to obtain categorized data groups including Class I, Class II, and Class III data. Dominant frequency data and low frequency data are obtained based on Class I and Class II data respectively. Differential motion data are extracted from the dominant frequency data and low frequency data, while Class III data is ignored. A first amplification factor and a second amplification factor are applied to the differential motion data respectively to obtain micro-motion video stream data.
5. The safety risk assessment method based on dangerous slope deformation identification according to claim 1, characterized in that, The process of generating micro-motion component sequence data from micro-motion field sequence data, and obtaining coherence map features, state features, and dynamic characteristic features based on the micro-motion component sequence data, includes: Based on the micro-motion field sequence data, micro-motion component sequence data containing slope aspect component sequence data and intensity component data are obtained, and coherence analysis is performed to obtain coherence map characteristics; Spectral analysis is performed based on micro-motion component sequence data, and an energy entropy evolution mechanism is introduced to obtain state characteristics. Dynamic characteristics are obtained based on intensity component data and input excitation signal.
6. The safety risk assessment method based on dangerous slope deformation identification according to claim 5, characterized in that, The process of acquiring micro-motion component sequence data containing slope aspect component sequence data and intensity component data based on micro-motion field sequence data, and performing coherence analysis to obtain coherence map characteristics, includes: The micro-dynamic field sequence data is extracted to obtain the aspect component sequence data and intensity component data respectively, and comparison data is defined at the same time; The slope aspect coherence coefficient is obtained by comparing the data with the slope aspect component sequence data, and the intensity coherence coefficient is obtained by comparing the data with the intensity component data. A coherence map is established based on the aspect coherence coefficient and the intensity coherence coefficient, and coherence map features are extracted from the coherence map.
7. The safety risk assessment method based on dangerous slope deformation identification according to claim 5, characterized in that, The method involves performing spectral analysis based on micro-motion component sequence data, while simultaneously introducing an energy entropy evolution mechanism to obtain state characteristics, including: Frequency domain transformation is performed on the aspect component sequence data and intensity component data to obtain the energy spectrum. The frequency band is then divided based on the energy spectrum to obtain the classification frequency band. The full-band energy entropy is obtained based on the energy spectrum, and the energy ratio of the classification band is obtained by comparing the energy spectrum with the classification band. At the same time, directional analysis and mutation analysis are performed to obtain evolutionary characteristic data. The full-band energy entropy, the energy ratio of the classified band, and the evolutionary characteristic data are integrated into state characteristics.
8. The safety risk assessment method based on dangerous slope deformation identification according to claim 5, characterized in that, The method for obtaining dynamic characteristic features based on intensity component data and input excitation signal includes: Define the input excitation signal, and simultaneously define the output response signal based on the intensity component data; Estimating the frequency response function based on the input excitation signal and the output response signal; The frequency response function is extracted to obtain the dominant resonant frequency, resonant peak gain, half-power bandwidth, and average gain, respectively. The dominant resonant frequency, resonant peak gain, half-power bandwidth, and average gain are integrated into dynamic characteristic features.
9. The safety risk assessment method based on dangerous slope deformation identification according to claim 1, characterized in that, Risk quantification data is obtained based on coherence map characteristics, state characteristics, and dynamic characteristics, and through a slope assessment model, including: A pre-trained dataset is constructed, and a slope assessment model is obtained based on the pre-trained dataset and meta-learning. Coherence map features, state features, and dynamic characteristic features are input into the slope assessment model. Dimensionality reduction assessment is performed through the slope assessment model to generate risk quantification data. The slope assessment model includes a dimensionality reduction model and an assessment model.
10. A safety risk assessment platform based on dangerous slope deformation identification, used to implement the safety risk assessment method based on dangerous slope deformation identification as described in any one of claims 1 to 9, characterized in that, The platform includes: Preprocessing sub-platform 11 is used to acquire slope video stream data and perform preprocessing operations to obtain standardized slope data; Reconstruction Sub-Platform 12 is used for super-resolution reconstruction of standardized slope data; Amplification sub-platform 13 is used to acquire micro-motion video stream data and simultaneously acquire micro-motion field sequence data through a differential amplification mechanism. Sub-platform 14 is used for feature extraction to obtain coherence map features, state features and dynamic characteristic features; The quantitative sub-platform 15 is used to obtain risk quantification data through the slope assessment model.