Timing network ground subsidence detection method and device based on distributed optical fiber phase signal

Abstract

The present invention provides a time-series network ground subsidence detection method and device based on distributed optical fiber phase signals. The detection method comprises: obtaining axial and radial strain signals of a distributed optical fiber in a pre-detected ground subsidence zone; performing noise reduction processing on the axial strain signal to obtain a first strain signal; performing enhancement processing on the radial strain signal to obtain a second strain signal; performing multi-dimensional feature extraction on the first and second strain signals to obtain first and second feature data; performing multi-scale time-series feature extraction on the first feature data to obtain time-series feature data; performing spatial feature mining on the second feature data to obtain feature mining data; fusing the time-series feature data and the feature mining data to obtain a fusion result; and performing ground subsidence detection based on the fusion result to obtain a ground subsidence detection result. This embodiment of the present invention can more accurately capture precursor information of ground subsidence.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of urban geological disaster detection, and in particular to a method and device for detecting ground subsidence using a time-series network based on distributed optical fiber phase signals. Background Art

[0002] Ground collapse detection is the process of monitoring, identifying, and assessing potential or existing ground collapses. Traditional distributed fiber optic sensing systems (DAS) are insensitive to radial strain, making it difficult to capture the multidimensional deformation characteristics of ground collapses. Furthermore, microvibrations that signal ground collapses are easily masked by noise such as traffic loads. Furthermore, existing methods lack the ability to analyze dynamic events coupled with multiple physical fields. Summary of the Invention

[0003] The technical problem to be solved by the embodiments of the present invention is to provide a time-series network ground subsidence detection method and device based on distributed optical fiber phase signals, which can more accurately capture the precursor information of ground collapse and improve the accuracy of detection.

[0004] To solve the above technical problems, the technical solutions of the embodiments of the present invention are as follows:

[0005] A time-series network ground sink detection method based on distributed optical fiber phase signals includes:

[0006] Acquiring an axial strain signal and a radial strain signal of a distributed optical fiber for pre-detecting a ground subsidence zone;

[0007] performing noise reduction processing on the axial strain signal to obtain a first strain signal;

[0008] performing enhancement processing on the radial strain signal to obtain a second strain signal;

[0009] performing multi-dimensional feature extraction on the first strain signal to obtain first feature data;

[0010] performing multi-dimensional feature extraction on the second strain signal to obtain second feature data;

[0011] Performing multi-scale time series feature extraction on the first feature data to obtain time series feature data;

[0012] Performing feature mining of a spatial dimension on the second feature data to obtain feature mining data;

[0013] Performing fusion processing on the time series feature data and the feature mining data to obtain a fusion processing result;

[0014] According to the fusion processing result, ground collapse detection is performed to obtain a ground collapse detection result.

[0015] Optionally, performing noise reduction processing on the axial strain signal to obtain a first strain signal includes:

[0016] performing preliminary screening and preprocessing on the axial strain signal to obtain a preprocessed axial strain signal;

[0017] Axial channel noise reduction processing is performed on the preprocessed axial strain signal to obtain a first strain signal.

[0018] Optionally, performing enhancement processing on the radial strain signal to obtain a second strain signal includes:

[0019] performing preliminary screening and preprocessing on the radial strain signal to obtain a preprocessed radial strain signal;

[0020] The pre-processed radial strain signal is subjected to radial channel enhancement processing to obtain a second strain signal.

[0021] Optionally, performing multi-dimensional feature extraction on the first strain signal to obtain first feature data includes:

[0022] performing time dimension feature extraction on the first strain signal to obtain first dimension feature data;

[0023] performing spatial dimension feature extraction on the first strain signal to obtain second dimension feature data;

[0024] fusing the first strain signal with a physical characteristic vector of the soil detected on site to obtain third-dimensional characteristic data;

[0025] First feature data is obtained according to the first dimensional feature data, the second dimensional feature data, and the third dimensional feature data.

[0026] Optionally, performing multi-dimensional feature extraction on the second strain signal to obtain second feature data includes:

[0027] Performing time dimension feature extraction on the first strain signal to obtain fourth dimension feature data;

[0028] performing spatial dimension feature extraction on the first strain signal to obtain fifth dimension feature data;

[0029] fusing the first strain signal with a physical characteristic vector of the soil detected on site to obtain sixth-dimensional characteristic data;

[0030] Second feature data is obtained according to the fourth dimensional feature data, the fifth dimensional feature data, and the sixth dimensional feature data.

[0031] Optionally, performing multi-scale time series feature extraction on the first feature data to obtain time series feature data includes:

[0032] Multi-scale time series feature extraction is performed on the first feature data to obtain time series feature data containing the complete time series change law.

[0033] Optionally, performing feature mining of a spatial dimension on the second feature data to obtain feature mining data includes:

[0034] Perform feature mining of the spatial dimension on the second feature data to obtain feature mining data that can reflect the spatial distribution characteristics.

[0035] Optionally, fusing the time series feature data and the feature mining data to obtain a fusion processing result includes:

[0036] Dynamically assigning weights to the time series feature data and the feature mining data;

[0037] Taking the soil mechanics constitutive equation as the physical constraint condition, the weighted time series feature data and feature mining data are fused to obtain the fusion processing results that include the time series change law and spatial propagation characteristics.

[0038] Optionally, performing ground collapse detection based on the fusion processing result to obtain a ground collapse detection result includes:

[0039] Extract the time series variation patterns and spatial propagation characteristics from the fusion processing results;

[0040] Comparing the spatial propagation characteristics with the set characteristics to determine the specific type of the ground subsidence;

[0041] Determine the occurrence area of ​​the landslide according to the time series variation law and spatial propagation characteristics;

[0042] An estimated energy value of the ground sink is determined according to the ground sink occurrence area.

[0043] An embodiment of the present invention also provides a time-series network ground sink detection device based on distributed optical fiber phase signals, comprising:

[0044] An acquisition module, used for acquiring an axial strain signal and a radial strain signal of a distributed optical fiber for pre-detecting a ground subsidence zone;

[0045] A processing module is configured to perform noise reduction processing on the axial strain signal to obtain a first strain signal; perform enhancement processing on the radial strain signal to obtain a second strain signal; perform multi-dimensional feature extraction on the first strain signal to obtain first feature data; perform multi-dimensional feature extraction on the second strain signal to obtain second feature data; perform multi-scale time series feature extraction on the first feature data to obtain time series feature data; perform spatial dimension feature mining on the second feature data to obtain feature mining data; perform fusion processing on the time series feature data and the feature mining data to obtain a fusion processing result; and perform ground collapse detection based on the fusion processing result to obtain a ground collapse detection result.

[0046] The above solution of the embodiment of the present invention has at least the following beneficial effects:

[0047] The above-mentioned scheme of the embodiment of the present invention enhances the detection capability of radial strain by obliquely cross-laying auxiliary sensing optical fibers. In order to solve the problem that the precursor micro-vibration of ground subsidence is easily masked by noise such as traffic load, the axial strain signal is subjected to noise reduction processing and the radial strain signal is subjected to enhancement processing, which effectively distinguishes the precursor micro-vibration of ground subsidence from noise, thereby improving the accuracy of detection. By performing multi-dimensional feature extraction on the first strain signal and the second strain signal, the multi-dimensional deformation characteristics of the ground subsidence can be fully captured, thereby achieving more accurate detection of ground collapse. By performing fusion processing on the time series feature data and the feature mining data, the complementarity between features is enhanced. The fusion processing result contains multi-dimensional feature information. Ground collapse detection based on this feature information can more accurately capture the precursor information of ground collapse, thereby improving the accuracy of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 The present invention provides a flow chart of a method for detecting ground sinkholes in a time-series network based on distributed optical fiber phase signals.

[0049] Figure 2 Schematic diagram of the positional relationship between the main sensing optical fiber and the auxiliary sensing optical fiber in the time-series network ground subsidence detection method based on distributed optical fiber phase signals provided by an embodiment of the present invention;

[0050] Figure 3 It is a schematic diagram of the positional relationship between the main sensing optical fiber and the temperature compensation node of the timing network ground subsidence detection method based on distributed optical fiber phase signals provided by an embodiment of the present invention.

[0051] Figure 4 It is a module schematic diagram of a timing network ground subsidence detection device based on distributed optical fiber phase signals provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0053] like Figure 1 As shown, an embodiment of the present invention provides a method for detecting ground sinks in a timing network based on a distributed optical fiber phase signal, comprising:

[0054] Step 11, obtaining an axial strain signal and a radial strain signal of a distributed optical fiber in a pre-detection ground subsidence zone;

[0055] Step 12: performing noise reduction processing on the axial strain signal to obtain a first strain signal;

[0056] Step 13: performing enhancement processing on the radial strain signal to obtain a second strain signal;

[0057] Step 14: performing multi-dimensional feature extraction on the first strain signal to obtain first feature data;

[0058] Step 15: performing multi-dimensional feature extraction on the second strain signal to obtain second feature data;

[0059] Step 16: performing multi-scale time series feature extraction on the first feature data to obtain time series feature data;

[0060] Step 17: performing feature mining of the spatial dimension on the second feature data to obtain feature mining data;

[0061] Step 18: fusing the time series feature data and the feature mining data to obtain a fusion processing result;

[0062] Step 19: Perform ground subsidence detection based on the fusion processing result to obtain a ground subsidence detection result.

[0063] like Figure 2 、 3 As shown, the distributed optical fiber may include: a plurality of main sensing optical fibers 1 arranged axially along the ground subsidence zone, for detecting the axial strain signal of the ground subsidence zone; a plurality of auxiliary sensing optical fibers 2 arranged axially along the ground subsidence zone, the plurality of auxiliary sensing optical fibers 2 being arranged obliquely across the plurality of main sensing optical fibers 1, and the auxiliary sensing optical fibers 2 being used to detect the radial strain signal of the ground subsidence zone; and temperature compensation nodes 3 arranged axially along the ground subsidence zone at set intervals, for monitoring ambient temperature changes and eliminating signal interference of thermal noise on the axial strain signal and the radial strain signal.

[0064] Specifically, the auxiliary sensing optical fiber 2 is arranged at a set angle of 45 degrees, the set spacing between multiple temperature compensation nodes 3 is 50 meters, and the main sensing optical fiber 1 can adopt a high-sensitivity single-mode optical fiber with a sampling rate of 1 kHz.

[0065] In this embodiment, the ability to detect radial strain is enhanced by cross-laying the auxiliary sensing optical fibers 2 in an oblique manner. In order to solve the problem that the microvibration that signals the beginning of a ground subsidence is easily masked by noise such as traffic load, the axial strain signal is subjected to noise reduction processing and the radial strain signal is subjected to enhancement processing, so that the microvibration that signals the beginning of a ground subsidence is effectively distinguished from the noise, thereby improving the accuracy of detection. By performing multi-dimensional feature extraction on the first strain signal and the second strain signal, the multi-dimensional deformation characteristics of the ground subsidence can be fully captured, thereby achieving more accurate detection of ground subsidence. By fusing the time series feature data and the feature mining data, the complementarity between the features is enhanced. The fusion processing result contains multi-dimensional feature information. Based on this feature information, ground subsidence detection can more accurately capture the precursor information of ground subsidence, thereby improving the accuracy of detection.

[0066] In an optional embodiment of the present invention, the step 11 may include:

[0067] Step 111, obtaining the axial strain signal of the pre-detection ground subsidence zone detected by the main sensing optical fiber 1;

[0068] Step 112 : obtaining a radial strain signal of the pre-detection ground subsidence zone detected by the auxiliary sensing optical fiber 2 .

[0069] In this embodiment, the axial strain signal is detected by the main sensing optical fiber 1, and the auxiliary sensing optical fibers 2 are arranged at a 45-degree oblique cross to enhance radial strain sensitivity, thereby capturing the multi-dimensional deformation characteristics of the ground subsidence and more comprehensively monitoring the deformation information of the ground collapse.

[0070] In an optional embodiment of the present invention, in step 12, performing noise reduction processing on the axial strain signal to obtain a first strain signal includes:

[0071] Step 121: Preliminary screening and preprocessing are performed on the axial strain signal to obtain a preprocessed axial strain signal. Specifically, the preprocessing may include: eliminating abnormal data caused by equipment failure or transient interference based on the amplitude and frequency characteristics of the signal;

[0072] Step 122: Perform axial channel noise reduction on the preprocessed axial strain signal to obtain a first strain signal. Specifically, the axial strain signal can be subjected to initial Kalman filtering to remove high-frequency noise (>500 Hz). The Rauch-Tung-Striebel (RTS) smoothing algorithm is then used to optimize the time series signal, retaining the low-frequency effective components (0.1-100 Hz). The preprocessed axial strain signal undergoes initial Kalman filtering and RTS smoothing to obtain the first strain signal.

[0073] In this embodiment, abnormal data generated by equipment failure or transient interference can introduce significant noise into the strain signal. By eliminating this abnormal data, the impact of this interference can be effectively reduced, making the strain signal clearer, facilitating subsequent accurate signal analysis, and improving the ability to detect weak ground subsidence signals.

[0074] Initial Kalman filtering effectively removes high-frequency noise (>500Hz) from the axial strain signal, resulting in a purer signal and reducing noise interference with subsequent analysis. The RTS smoothing algorithm optimizes the time series signal, preserving the low-frequency components (0.1-100Hz). These components are crucial for monitoring conditions such as slow subsidence and help more accurately capture the signal characteristics associated with ground subsidence. This processing yields a higher-quality first strain signal that better reflects the actual ground subsidence, providing a more reliable data foundation for subsequent analysis and early warning. Effectively removing noise and preserving key signal components enhances the overall ground subsidence detection system's ability to detect subtle changes, improving detection accuracy and reliability.

[0075] In an optional embodiment of the present invention, in step 13, the radial strain signal is enhanced to obtain a second strain signal, including:

[0076] Step 131: Preliminary screening and preprocessing are performed on the radial strain signal to obtain a preprocessed radial strain signal. Specifically, the preprocessing may include: eliminating abnormal data caused by equipment failure or transient interference based on the amplitude and frequency characteristics of the signal;

[0077] Step 132: Perform radial channel enhancement processing on the pre-processed radial strain signal to obtain a second strain signal. Specifically, the second strain signal can be obtained by E n =r n 2 +r （n-1） r （n+1） Determine the instantaneous energy of the radial strain signal at each sampling point,

[0078] Among them, E nis the instantaneous energy of the radial strain signal at each sampling point, n=1,2,3...N, N is the number of signal sampling points, r n is the discrete sequence of the nth sampling points of the radial strain signal, r （n-1） is the discrete sequence of the n-1th sampling points of the radial strain signal, r （n+1） is the discrete sequence of the n+1th signal sampling points of the radial strain signal;

[0079] You can use f n =f min + (f max -f min ) Dynamically adjust the filter cutoff frequency of each sampling point of the radial strain signal, where f n is the filter cutoff frequency of each sampling point of the radial strain signal, f min 、f max is the minimum and maximum value of the cutoff frequency. 、 is the minimum and maximum value of instantaneous energy, E n is the instantaneous energy of the radial strain signal at each sampling point;

[0080] The radial strain signal is filtered using a bandpass filter, and the filter cutoff frequency f of each sampling point of the radial strain signal is set to n Dynamically adjust the passband range to achieve dynamic filtering enhancement of the signal.

[0081] In this embodiment, the elimination of abnormal data based on amplitude and frequency characteristics in step 131 can effectively remove erroneous signals caused by equipment failure or transient interference (such as lightning, mechanical vibration, etc.), preventing these abnormal data from interfering with subsequent analysis. This allows the preprocessed radial strain signal to more realistically reflect geological changes, laying a reliable data foundation for the entire ground subsidence detection process.

[0082] Step 132, by calculating the instantaneous energy, effectively captures transient vibration characteristics in the radial strain signal, such as vibrations caused by the formation of soil microcracks. These transient vibrations are often important precursors to land subsidence. By enhancing the ability to extract such signals, early signs of land subsidence can be detected more promptly and accurately, buying valuable time for subsequent warnings.

[0083] Dynamically adjusting the filter cutoff frequency based on instantaneous energy allows the bandpass filter to flexibly adapt to signal characteristics under different operating conditions. When the signal's instantaneous energy is high, indicating the presence of strong transient vibrations, increasing the cutoff frequency can better preserve high-frequency transient information. When the instantaneous energy is low, lowering the cutoff frequency can suppress low-frequency noise. This dynamic filtering method enhances the processing capabilities of complex signals and improves the targetedness and effectiveness of signal processing.

[0084] After preliminary screening, preprocessing, and dynamic filtering enhancement, the impact of noise on the radial strain signal is effectively reduced, improving the signal-to-noise ratio. This enables the system to accurately identify valid signals related to ground subsidence from numerous interfering signals even in complex environments, such as busy urban roads, enhancing the system's anti-interference capabilities and environmental adaptability.

[0085] In an optional embodiment of the present invention, step 13 may further include:

[0086] Step 133: Calculate the coherence of the first strain signal and the second strain signal. Specifically, the coherence can be calculated by R m = , determine the cross-correlation function of the first strain signal and the second strain signal, where R m is the cross-correlation function of the first strain signal and the second strain signal, m is the time delay, m=0, ±1, ±2, ..., is the discrete sequence of the nth signal sampling point of the first strain signal, is the second strain signal nth A discrete sequence of signal sampling points, n=1,2,3...N, where N is the number of signal sampling points;

[0087] By ρ= , determining a coherence coefficient between the first strain signal and the second strain signal, wherein ρ is the coherence coefficient between the first strain signal and the second strain signal, is the cross-correlation function of the first strain signal and the second strain signal when the time delay is 0;

[0088] Step 134: Eliminate interference between the first strain signal and the second strain signal based on the coherence; specifically, the interference can be eliminated by λ a = ,λ b = , determine the weight of the first strain signal and the weight of the second strain signal, where λ a is the weight of the first strain signal, λ b is the weight of the second strain signal, and ρ is the coherence coefficient between the first strain signal and the second strain signal;

[0089] By w n =λ a +λ b , and obtain the total strain signal of the first strain signal and the second strain signal, where w n is the total strain signal, λ a is the weight of the first strain signal, λ b is the weight of the second strain signal, is the discrete sequence of the nth signal sampling point of the first strain signal, is a discrete sequence of the nth signal sampling points of the second strain signal;

[0090] By x n = ,y n = , obtain the first strain signal and the second strain signal after eliminating interference, where x n is the first strain signal after eliminating interference, y n The second strain signal after eliminating interference is w n is the total strain signal, λ a is the weight of the first strain signal, λ b is the weight of the second strain signal, is the discrete sequence of the nth signal sampling point of the first strain signal, is a discrete sequence of the nth signal sampling points of the second strain signal.

[0091] In this embodiment, step 133 quantifies the relationship between the first and second strain signals. The cross-correlation function reflects the degree of similarity between the two signals at different time delays, while the coherence coefficient provides a normalized representation of the degree of linear correlation between them. This quantitative analysis enables subsequent processing based on accurate inter-signal correlation characteristics, avoiding information loss or erroneous enhancement caused by blind processing. This provides a reliable basis for interference elimination and signal fusion, and facilitates more accurate understanding and utilization of the ground subsidence-related information contained in the two signals.

[0092] In step 134, weights are determined based on the coherence coefficient, achieving weighted fusion of the two signals. When the coherence coefficient indicates strong correlation or interference between the signals, weight adjustment can reduce duplicate information and mutual interference, preserving valid information. When the signal correlation is weak, weight allocation can fully utilize the different valid information components of the two signals. This adaptive weighting strategy can specifically eliminate inter-channel interference, making the fused total strain signal purer, reducing the impact of noise and interference on the signal, and significantly improving signal quality.

[0093] By separating the interference-free primary and secondary strain signals from the total strain signal, the accuracy and reliability of each signal are further ensured. This separation process, a rational inverse operation based on weighted fusion, preserves the individual characteristics of the signals while eliminating any interference that may remain during the fusion process. This allows for a more realistic reflection of the axial and radial strain conditions during a landslide, providing higher-quality data for subsequent landslide detection, feature extraction, and early warning decision-making, thereby improving the accuracy and effectiveness of the entire landslide detection system.

[0094] In an optional embodiment of the present invention, in step 14, performing multi-dimensional feature extraction on the first strain signal to obtain first feature data includes:

[0095] Step 141: extract the time dimension feature of the first strain signal to obtain the first dimension feature data; specifically, extract the time dimension feature of the first strain signal Perform empirical mode decomposition (EMD) to obtain j intrinsic mode function (IMF) components, and calculate the energy proportion p of each IMF component i :

[0096] According to H= Calculate the intrinsic mode function energy entropy,

[0097] Where H is the energy entropy of the intrinsic mode function, p i is the ratio of the energy of each intrinsic mode function component to the total energy, i=1,2,3...j, j is the total number of all intrinsic mode function components;

[0098] The instantaneous frequency of each IMF is obtained through Hilbert transform, and the statistical characteristics of the instantaneous frequency (such as mean and variance) are further extracted. The first dimension feature data includes the energy entropy of the intrinsic mode function and the instantaneous frequency of each IMF.

[0099] Step 142 extracts spatial features from the first strain signal to obtain second-dimensional feature data. Specifically, the spatial coherence between sensing fibers at different locations is calculated and a coherence threshold is set to determine the coherence radius. This radius can identify abnormal areas of strain wave propagation (e.g., high coherence attenuation occurs around voids). A spatial adjacency matrix is ​​constructed based on the fiber network topology. The matrix elements comprehensively consider sensor spacing and coherence to quantify the attenuation characteristics of strain propagation and obtain the attenuation coefficient. The second-dimensional feature data includes the coherence radius and the attenuation coefficient.

[0100] Step 143: The first strain signal is fused with the physical feature vectors of the soil detected on-site to obtain third-dimensional feature data. Specifically, the soil parameters (water content, pore pressure, shear modulus, and density) monitored on-site are nonlinearly mapped using a multi-layer perceptron (MLP) to convert them into more expressive physical feature vectors. To enhance the correlation between the physical and spatiotemporal features, weights are calculated based on the variance of the spatiotemporal features, and the physical feature vectors output by the MLP are weighted and adjusted to form the final physical features. The third-dimensional feature data includes the final physical features.

[0101] Step 144: First feature data is obtained based on the first, second, and third dimensional feature data. The first feature data includes the intrinsic mode function energy entropy, the instantaneous frequency of each IMF, the coherence radius, the attenuation coefficient, and the final physical characteristics. Specifically, the feature vectors of the three dimensions of time, space, and physical are concatenated, and the feature expression capability is enhanced through nonlinear transformation. The final output first feature data includes: time features such as IMF energy entropy and instantaneous frequency, which reflect the complex characteristics of the signal in the time domain; spatial features such as the coherence radius and attenuation coefficient, which reveal the propagation anomalies of strain in space; and physical features after MLP mapping and weighting of soil parameters, which reflect the influence of geophysical properties on the signal.

[0102] In this embodiment, the signal is adaptively separated into IMF components of varying scales. IMF energy entropy quantifies energy distribution uniformity, while instantaneous frequency statistics capture time-varying characteristics, making it more sensitive to subtle anomalies. Anomalous regions are located using coherence radius (e.g., coherence decay around cavities), and the adjacency matrix quantifies strain propagation attenuation, enabling regionalized monitoring capabilities. MLP nonlinear mapping converts soil parameters into high-dimensional feature vectors, and a spatiotemporal weighting strategy dynamically correlates physical characteristics with monitoring signals, enhancing targeted monitoring.

[0103] Temporal features capture temporal changes, spatial features locate abnormal areas, and physical features provide geological support. The combination of these three reduces single-dimensional misjudgments. IMF energy entropy, coherence radius, and MLP features synergistically improve the early detection rate of ground subsidence, especially for small-scale cavities and weak vibrations.

[0104] In an optional embodiment of the present invention, in step 15, performing multi-dimensional feature extraction on the second strain signal to obtain second feature data includes:

[0105] Step 151: extracting time dimension features from the first strain signal to obtain fourth dimension feature data;

[0106] Step 152: extracting spatial dimension features from the first strain signal to obtain fifth dimensional feature data;

[0107] Step 153: fusing the physical characteristic vector of the soil detected on site with the first strain signal to obtain sixth-dimensional characteristic data;

[0108] Step 154 ​​: Obtain second feature data based on the fourth dimensional feature data, the fifth dimensional feature data, and the sixth dimensional feature data.

[0109] The specific implementation process of step 15 can refer to step 14.

[0110] In this embodiment, multi-dimensional feature extraction of the second strain signal deeply explores signal information from the temporal, spatial, and physical dimensions. This information complements the characteristic data of the first strain signal to construct a more complete map of ground subsidence characteristics. The temporal dimension captures the time-varying patterns of the signal, the spatial dimension locates abnormal areas, and the physical dimension, combined with soil parameters, verifies ground subsidence signs from multiple angles, reducing misjudgment of single signals and improving data reliability.

[0111] In an optional embodiment of the present invention, in step 16, performing multi-scale time series feature extraction on the first feature data to obtain time series feature data includes:

[0112] In step 161, multi-scale time series feature extraction is performed on the first feature data to obtain time series feature data that contains the complete time series variation pattern. Specifically, a dilated causal convolutional network is used, with dilation factors set to 2, 4, and 8, respectively. As the convolution proceeds, different dilation factors cause the receptive field of the convolution kernel to continuously change. Smaller dilation factors focus on the short-term fluctuation details of the signal, capturing information such as sudden high-frequency vibrations in the axial strain signal; larger dilation factors focus on long-term trends, such as long-term soil settlement or slow stress changes. In this way, the axial strain signal is fully covered, from high-frequency vibrations within a few seconds to long-term trend changes of several days or even months, and multi-scale time series features are extracted.

[0113] In this embodiment, the small expansion factor (2) captures high-frequency vibrations (seconds), and the large expansion factor (8) tracks long-term trends (day / month), completely preserving the full-cycle signal characteristics from microcracks in the early stage of the sinkhole to gradual settlement, avoiding the loss of single-scale information. The short-term fluctuation details (such as sudden strain peaks) and the long-term stress change trends (such as continuous settlement slope) are extracted in a coordinated manner, which not only sensitively captures precursor anomalies but also accurately depicts the evolution trajectory, enhancing the model's ability to analyze complex sinkhole processes. Multi-scale parallel processing enables the model to adapt to different signal modes and improve its noise resistance. When a single scale is interfered with, the other scale features can still provide reliable support, reduce the false alarm rate, and ensure the detection stability in complex geological environments. The rich multi-scale features directly improve the accuracy of sinkhole classification and positioning, especially the sensitivity to early weak signals.

[0114] In an optional embodiment of the present invention, in step 17, performing feature mining of a spatial dimension on the second feature data to obtain feature mining data includes:

[0115] Step 171 performs spatial feature mining on the second feature data to obtain feature-mined data that reflects spatial distribution characteristics. Specifically, based on the Graph Attention Network (GAT), each fiber optic sensing unit is considered a node, and edge weights are determined based on spatial coherence. Spatial coherence reflects the degree of correlation between strain signals between sensing units at different locations. Nodes with high coherence have larger edge weights, indicating smoother signal propagation between these nodes and greater mutual influence. Nodes with low coherence have smaller edge weights. Through GAT, the model can adaptively focus on key sensing unit nodes in space and their interrelationships, thereby accurately modeling the propagation patterns of radial strain signals in space.

[0116] In this implementation, edge weights are defined based on spatial coherence, adaptively focusing on highly correlated nodes (such as those surrounding voids) to accurately identify anomalous regions of strain propagation (such as coherence attenuation zones). This approach transcends the traditional method's reliance on fixed grid structures and enhances its ability to characterize complex geological structures. GAT dynamically adjusts inter-node weights through an attention mechanism, tracking the spatial diffusion path of strain waves in real time. The spatial heterogeneity of edge weight distribution enables precise localization of anomaly sources (such as void centers). Combined with the coherence radius metric, this approach achieves millimeter-level localization accuracy. The attention mechanism automatically suppresses noise in low-coherence regions, enhancing sensitivity to weak anomalous signals.

[0117] In an optional embodiment of the present invention, in step 18, the time series feature data and the feature mining data are fused to obtain a fusion result, including:

[0118] Step 181: Dynamically assign weights to the temporal feature data and the feature mining data; specifically, through a cross-modal attention mechanism, the weight q is flexibly adjusted according to the importance of the temporal feature data and the feature mining data to the current sinkhole detection task. a and q b ,

[0119] Among them, q a is the weight of the time series feature data, q b Weights for feature mining data;

[0120] Step 182: Using the soil mechanics constitutive equation as a physical constraint, the weighted time series feature data and the feature mining data are fused to obtain a fusion result that includes the time series variation law and spatial propagation characteristics.

[0121] Specifically, according to F f =q aF a +q b F b , determine the preliminary fusion results, where F f is the preliminary fusion result, q a is the weight of the time series feature data, q b is the weight of feature mining data, F a is the time series feature data, F b mining data for features;

[0122] Determine the physical constraint loss according to the physical constraint condition, determine the correction term Z of the fusion process according to the physical constraint loss, and determine the correction term Z of the fusion process according to F final =F f +Z, determine the final fusion processing result, where F final is the final fusion processing result, F f is the preliminary fusion result, and Z is the correction term of the fusion process.

[0123] In this embodiment, the cross-modal attention mechanism dynamically adjusts weights based on task requirements, enabling the model to flexibly focus on key information. For example, under different geological conditions, if the initial signs of a landslide are high-frequency vibration anomalies in time series, the model automatically increases the weight of the time series features. If the anomaly is concentrated in a localized area, the weight of the feature mining data is increased, enhancing the model's adaptability to complex scenarios.

[0124] The weighted fusion strategy leverages the complementary nature of time series feature data and feature mining data. Time series features record the temporal evolution of strain signals, while feature mining data characterizes their spatial distribution. The fusion of these two fully preserves the temporal and spatial variations in landslide events, avoiding misjudgments caused by missing features in a single dimension.

[0125] The soil mechanics constitutive equation is introduced as a constraint to ensure that the fusion results conform to actual physical laws. The correction term is calculated through the physical constraint loss to adjust the initial fusion results that do not conform to physical rules, reducing the deviation between the model output and the actual physical phenomena, making the prediction results more scientific and credible.

[0126] Dynamic weights and physical constraints work together to effectively suppress the effects of noise and interference. Even if some data fluctuates abnormally, the model can output reasonable fusion results through weight adjustments and physical constraint corrections, reducing the interference of external factors on detection accuracy and improving the model's stability and robustness.

[0127] In an optional embodiment of the present invention, in step 19, performing ground collapse detection based on the fusion processing result to obtain a ground collapse detection result includes:

[0128] Step 191: extract the time series variation pattern and spatial propagation characteristics from the fusion processing result; specifically, extract the fusion processing result F final The time series variation law AX and spatial propagation characteristics BY,

[0129] Among them, the time series change law AX can include the trend slope k, volatility σ and mutation point arg, and the spatial propagation characteristics BY can include the coherence radius coh, strain gradient tr and abnormal aggregation C.

[0130] Step 192: Compare the time series variation pattern and spatial propagation characteristics with the set characteristics to determine the specific type of the sinkhole; specifically, construct a feature vector V = {k, σ, arg, coh, tr, C},

[0131] Among them, V is the characteristic vector, k is the trend slope, σ is the volatility, arg is the mutation point, r coh is the coherence radius, s tr is the strain gradient, C l is abnormal aggregation;

[0132] The characteristic vector V is combined with the set ground subsidence type vector set U t ={k t , σ t , arg t , coh t , tr t , C t}, perform similarity analysis and obtain similarity coefficient. Specifically, according to DL t =

[0133]

[0134] Determine the similarity coefficient DL between the feature vector V and each vector in the set sinkhole type vector set t ,

[0135] Among them, the sinkhole type vector set U t is a collection of various sinkhole type vectors;

[0136] Compare all similarity coefficients DL t , get the land subsidence type corresponding to the minimum value, and determine the specific type of land subsidence.

[0137] Step 193: Determine the area where the ground subsidence occurs based on the spatial propagation characteristics. Specifically, if the strain gradient tr>trs, C>Cs in a certain area, then mark the area as an abnormal point.

[0138] Where tr is the strain gradient of a certain area, C is the abnormal aggregation degree of a certain area, trs is the set strain gradient threshold, and Cs is the set abnormal aggregation degree threshold;

[0139] Cluster the outliers whose distances between them are less than the set distance to form outlier clusters;

[0140] Compare the area and number of points of each abnormal point cluster to obtain the abnormal point cluster with the largest area or the most points, and determine the area where the ground subsidence occurs;

[0141] Step 194: Determine the estimated energy value of the sinkhole according to the sinkhole occurrence area; specifically, determine the time correction coefficient α according to the trend slope k and the duration, and determine the mutation point correction coefficient arg according to the mutation point intensity. ;

[0142] According to E=α , determine the energy estimate of the sinkhole,

[0143] Where E is the estimated energy of the sinkhole, α is the time correction coefficient, is the mutation point correction coefficient, is the total number of points in the sinkhole occurrence area, Points within the sinkhole area The strain gradient, Points within the sinkhole area The abnormal aggregation of is the unit volume.

[0144] In this embodiment, in the feature extraction step (step 191), the time series variation pattern and spatial propagation characteristics are extracted from the fusion results. The former includes trend slope, volatility, and mutation point, while the latter includes coherence radius, strain gradient, and abnormal aggregation, etc., which comprehensively capture the spatiotemporal information of the collapse process. In the type determination step (192), a vector containing the above features is constructed and similarity analysis is performed with the set ground subsidence type vector set. By comparing the similarity coefficients, the collapse type is accurately determined, and different types such as karst collapse and goaf collapse are effectively distinguished.

[0145] Regional positioning (step 193) sets thresholds based on strain gradient and anomaly clustering, filters and clusters outliers, and identifies the collapse region using the cluster with the largest area or number of points. This achieves sub-meter positioning accuracy and reduces false alarm rates. Energy estimation (step 194) determines a correction factor based on the trend slope, duration, and intensity of the time series. Energy is calculated by comprehensively considering the strain gradient, anomaly clustering, total number of points, and unit volume within the collapse region, in accordance with the laws of soil mechanics.

[0146] In an optional embodiment of the present invention, the detection method further comprises:

[0147] Step 20: Make a multi-level early warning decision based on the ground collapse detection result.

[0148] Specifically, in step 20, the multi-level warning decision-making is performed based on the ground collapse detection result, including:

[0149] Step 201, determining the strain rate, coherence radius and energy entropy mutation value of the ground subsidence area included in the ground subsidence detection result;

[0150] Step 202: Compare the strain rate with a first set trigger threshold, and determine whether the strain rate exceeds the first set trigger threshold for more than a first set time.

[0151] Step 203: If the conditions are not met, continue to monitor the strain rate of the subsidence area in real time;

[0152] Step 204: if the conditions are met, determine the coherence radius of the subsidence area based on the fused features;

[0153] Step 205, comparing the coherence radius with a second set verification threshold;

[0154] Step 206: If the conditions are not met, continue to monitor the strain rate of the subsidence area in real time;

[0155] Step 207: If the conditions are met, determine the energy entropy mutation value of the subsidence area based on the fused features;

[0156] Step 208, comparing the energy entropy mutation value with a third set determination threshold;

[0157] Step 209: If the conditions are not met, continue to monitor the strain rate of the subsidence area in real time;

[0158] Step 2010: If the conditions are met, a red alert for landslide is issued;

[0159] Step 2011: Dynamically adjust the first set trigger threshold during heavy rain or traffic rush hour.

[0160] In this embodiment, the multi-level early warning decision-making mechanism of step 20 significantly improves the reliability and timeliness of ground collapse warning through multi-indicator verification and dynamic threshold adjustment. First, the strain rate, coherence radius, and energy entropy mutation value are used to construct a progressive judgment logic to avoid misjudgment of a single indicator. The red warning is triggered only when multiple conditions are met at the same time to ensure the accuracy of the warning. Secondly, the trigger threshold is dynamically adjusted for special scenarios such as heavy rain and traffic rush, so that the early warning system is more in line with the actual risks. For example, the threshold is lowered during heavy rain to capture risk signals in advance and reserve sufficient emergency time. Furthermore, when the warning conditions are not met, continuous real-time monitoring is performed to form a closed-loop feedback mechanism to capture the dynamic changes of ground subsidence in a timely manner. This mechanism uses the complementary multi-dimensional indicators to comprehensively assess risks from the perspectives of development speed, impact range, energy anomalies, etc., providing a scientific basis for emergency decision-making.

[0161] In terms of data acquisition and preprocessing, the present invention uses obliquely cross-laid optical fibers to enhance radial strain detection, and uses Kalman filtering, RTS smoothing algorithm, and dynamic filtering enhancement processing to effectively reduce noise while retaining key signals, remove abnormal interference such as equipment failure, improve the signal-to-noise ratio, and enhance the ability to capture weak ground subsidence signals.

[0162] In the feature extraction and fusion stage, multi-dimensional and multi-scale feature extraction comprehensively captures the spatiotemporal and physical characteristics of ground subsidence. The dilated causal convolutional network and graph attention network extract temporal and spatial features respectively. The cross-modal attention mechanism is dynamically weighted and fused, combined with the correction of the soil mechanics constitutive equation, to enhance feature complementarity and physical rationality, and improve the model stability and adaptability to complex scenarios.

[0163] During ground collapse detection, the system extracts and compares features to determine the type of subsidence, clusters the location based on thresholds, and estimates energy using multiple factors. This allows for precise detection, achieving sub-meter accuracy, and ensuring energy estimation complies with physical laws. Multi-level early warning decision-making utilizes a multi-metric, progressive judgment process combined with dynamic threshold adjustment to avoid misjudgments, improve the accuracy and timeliness of early warnings, and provide proactive responses to specific scenarios, providing a reliable basis for emergency decision-making.

[0164] like Figure 4 As shown, an embodiment of the present invention further provides a timing network ground sink detection device 40 based on distributed optical fiber phase signals, comprising:

[0165] An acquisition module 41 is configured to acquire an axial strain signal and a radial strain signal of a distributed optical fiber in a pre-detection ground subsidence zone;

[0166] The processing module 42 is used to perform noise reduction processing on the axial strain signal to obtain a first strain signal; perform enhancement processing on the radial strain signal to obtain a second strain signal; perform multi-dimensional feature extraction on the first strain signal to obtain first feature data; perform multi-dimensional feature extraction on the second strain signal to obtain second feature data; perform multi-scale time series feature extraction on the first feature data to obtain time series feature data; perform spatial dimension feature mining on the second feature data to obtain feature mining data; perform fusion processing on the time series feature data and the feature mining data to obtain a fusion processing result; and perform ground collapse detection based on the fusion processing result to obtain a ground collapse detection result.

[0167] Optionally, performing noise reduction processing on the axial strain signal to obtain a first strain signal includes:

[0168] performing preliminary screening and preprocessing on the axial strain signal to obtain a preprocessed axial strain signal;

[0169] Axial channel noise reduction processing is performed on the preprocessed axial strain signal to obtain a first strain signal.

[0170] Optionally, performing enhancement processing on the radial strain signal to obtain a second strain signal includes:

[0171] performing preliminary screening and preprocessing on the radial strain signal to obtain a preprocessed radial strain signal;

[0172] The pre-processed radial strain signal is subjected to radial channel enhancement processing to obtain a second strain signal.

[0173] Optionally, performing multi-dimensional feature extraction on the first strain signal to obtain first feature data includes:

[0174] performing time dimension feature extraction on the first strain signal to obtain first dimension feature data;

[0175] performing spatial dimension feature extraction on the first strain signal to obtain second dimension feature data;

[0176] fusing the first strain signal with a physical characteristic vector of the soil detected on site to obtain third-dimensional characteristic data;

[0177] First feature data is obtained according to the first dimensional feature data, the second dimensional feature data, and the third dimensional feature data.

[0178] Optionally, performing multi-dimensional feature extraction on the second strain signal to obtain second feature data includes:

[0179] Performing time dimension feature extraction on the first strain signal to obtain fourth dimension feature data;

[0180] performing spatial dimension feature extraction on the first strain signal to obtain fifth dimension feature data;

[0181] fusing the first strain signal with a physical characteristic vector of the soil detected on site to obtain sixth-dimensional characteristic data;

[0182] Second feature data is obtained according to the fourth dimensional feature data, the fifth dimensional feature data, and the sixth dimensional feature data.

[0183] Optionally, performing multi-scale time series feature extraction on the first feature data to obtain time series feature data includes:

[0184] Multi-scale time series feature extraction is performed on the first feature data to obtain time series feature data containing the complete time series change law.

[0185] Optionally, performing feature mining of a spatial dimension on the second feature data to obtain feature mining data includes:

[0186] Perform feature mining of the spatial dimension on the second feature data to obtain feature mining data that can reflect the spatial distribution characteristics.

[0187] Optionally, fusing the time series feature data and the feature mining data to obtain a fusion processing result includes:

[0188] Dynamically assigning weights to the time series feature data and the feature mining data;

[0189] Taking the soil mechanics constitutive equation as the physical constraint condition, the weighted time series feature data and feature mining data are fused to obtain the fusion processing results that include the time series change law and spatial propagation characteristics.

[0190] Optionally, performing ground collapse detection based on the fusion processing result to obtain a ground collapse detection result includes:

[0191] Extract the time series variation patterns and spatial propagation characteristics from the fusion processing results;

[0192] Comparing the spatial propagation characteristics with the set characteristics to determine the specific type of the ground subsidence;

[0193] Determine the occurrence area of ​​the landslide according to the time series variation law and spatial propagation characteristics;

[0194] An estimated energy value of the ground sink is determined according to the ground sink occurrence area.

[0195] It should be noted that this device is a device corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0196] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A time-series network ground sink detection method based on distributed optical fiber phase signals, characterized in that: include: Acquiring an axial strain signal and a radial strain signal of a distributed optical fiber for pre-detecting a ground subsidence zone; performing noise reduction processing on the axial strain signal to obtain a first strain signal; performing enhancement processing on the radial strain signal to obtain a second strain signal; performing multi-dimensional feature extraction on the first strain signal to obtain first feature data; performing multi-dimensional feature extraction on the second strain signal to obtain second feature data; Performing multi-scale time series feature extraction on the first feature data to obtain time series feature data; Performing feature mining of a spatial dimension on the second feature data to obtain feature mining data; Performing fusion processing on the time series feature data and the feature mining data to obtain a fusion processing result; Performing ground collapse detection based on the fusion processing result to obtain a ground collapse detection result; The step of performing multi-dimensional feature extraction on the first strain signal to obtain first feature data includes: performing time dimension feature extraction on the first strain signal to obtain first dimension feature data; performing spatial dimension feature extraction on the first strain signal to obtain second dimension feature data; fusing the first strain signal with a physical characteristic vector of the soil detected on site to obtain third-dimensional characteristic data; Obtaining first feature data according to the first dimensional feature data, the second dimensional feature data, and the third dimensional feature data; The multi-dimensional feature extraction is performed on the second strain signal to obtain second feature data, including: performing time dimension feature extraction on the second strain signal to obtain fourth dimension feature data; performing spatial dimension feature extraction on the second strain signal to obtain fifth dimension feature data; fusing the second strain signal with a physical characteristic vector of the soil detected on site to obtain sixth-dimensional characteristic data; Obtaining second feature data according to the fourth dimension feature data, the fifth dimension feature data, and the sixth dimension feature data; The fusion processing of the time series feature data and the feature mining data to obtain the fusion processing result includes: Dynamically assigning weights to the time series feature data and the feature mining data; Taking the soil mechanics constitutive equation as the physical constraint, the weighted time series feature data and feature mining data are fused to obtain the fusion results that include the time series variation law and spatial propagation characteristics. The method of performing ground collapse detection based on the fusion processing result to obtain the ground collapse detection result includes: Extract the time series variation patterns and spatial propagation characteristics from the fusion processing results; Comparing the time series variation pattern with the spatial propagation characteristics and the set characteristics to determine the specific type of the ground subsidence; determining an occurrence area of ​​the ground subsidence according to the spatial propagation characteristics; An estimated energy value of the ground sink is determined according to the ground sink occurrence area.

2. The method for detecting ground sinkholes in a time-series network based on distributed optical fiber phase signals according to claim 1, characterized in that: Performing noise reduction processing on the axial strain signal to obtain a first strain signal includes: performing preliminary screening and preprocessing on the axial strain signal to obtain a preprocessed axial strain signal; Axial channel noise reduction processing is performed on the preprocessed axial strain signal to obtain a first strain signal.

3. The method for detecting ground sinkholes in a time-series network based on distributed optical fiber phase signals according to claim 1, characterized in that: Performing enhancement processing on the radial strain signal to obtain a second strain signal includes: performing preliminary screening and preprocessing on the radial strain signal to obtain a preprocessed radial strain signal; The pre-processed radial strain signal is subjected to radial channel enhancement processing to obtain a second strain signal.

4. The method for detecting ground sinkholes in a time-series network based on distributed optical fiber phase signals according to claim 1, wherein: Performing multi-scale time series feature extraction on the first feature data to obtain time series feature data includes: Multi-scale time series feature extraction is performed on the first feature data to obtain time series feature data containing the complete time series change law.

5. The method for detecting ground sinkholes in a time-series network based on distributed optical fiber phase signals according to claim 1, characterized in that: Performing feature mining of a spatial dimension on the second feature data to obtain feature mining data includes: Perform feature mining of the spatial dimension on the second feature data to obtain feature mining data that can reflect the spatial distribution characteristics.

6. A time-series network ground subsidence detection device based on distributed optical fiber phase signals, characterized in that: include: An acquisition module, used for acquiring an axial strain signal and a radial strain signal of a distributed optical fiber for pre-detecting a ground subsidence zone; a processing module configured to perform noise reduction processing on the axial strain signal to obtain a first strain signal; perform enhancement processing on the radial strain signal to obtain a second strain signal; and perform multi-dimensional feature extraction on the first strain signal to obtain first feature data; Perform multi-dimensional feature extraction on the second strain signal to obtain second feature data; perform multi-scale time series feature extraction on the first feature data to obtain time series feature data; perform spatial dimension feature mining on the second feature data to obtain feature mining data; perform fusion processing on the time series feature data and the feature mining data to obtain a fusion processing result; perform ground collapse detection based on the fusion processing result to obtain a ground collapse detection result; wherein, performing multi-dimensional feature extraction on the first strain signal to obtain first feature data includes: performing time dimension feature extraction on the first strain signal to obtain first dimension feature data; performing spatial dimension feature extraction on the first strain signal to obtain second dimension feature data; performing fusion of physical feature vectors of soil detected on site on the first strain signal to obtain third dimension feature data; obtaining first feature data based on the first dimension feature data, the second dimension feature data and the third dimension feature data; wherein, performing multi-dimensional feature extraction on the second strain signal to obtain second feature data includes: performing time dimension feature extraction on the second strain signal to obtain fourth dimension feature data According to the present invention, the present invention relates to a method for detecting a ground collapse according to the present invention; extracting spatial dimensional features of the second strain signal to obtain fifth dimensional feature data; fusing the physical feature vectors of the soil detected on site to obtain sixth dimensional feature data; and obtaining second feature data based on the fourth dimensional feature data, the fifth dimensional feature data, and the sixth dimensional feature data. The method further comprises fusing the time series feature data and the feature mining data to obtain a fusion processing result, including: dynamically assigning weights to the time series feature data and the feature mining data; fusing the weighted time series feature data and the feature mining data using the soil mechanics constitutive equation as a physical constraint condition to obtain a fusion processing result including time series variation patterns and spatial propagation characteristics; and performing ground collapse detection based on the fusion processing result to obtain a ground collapse detection result, including: extracting time series variation patterns and spatial propagation characteristics from the fusion processing result; comparing the time series variation patterns and spatial propagation characteristics with set characteristics to determine the specific type of ground collapse; determining the occurrence area of ​​the ground collapse based on the spatial propagation characteristics; and determining the energy estimation value of the ground collapse based on the occurrence area of ​​the ground collapse.

Images