Multi-parameter coupling surface collapse risk identification and early warning method based on time sequence convolutional network

Abstract

The invention discloses a multi-parameter coupling surface collapse risk identification and early warning method based on a time sequence convolutional network, and the method comprises the steps: arranging an FBG sensor array and a parallel electrical method system, and collecting the one-dimensional strain time sequence data and the two-dimensional resistivity space distribution data of a pipeline and a surrounding soil body; constructing a double-branch deep learning network, and performing modal specific feature extraction on the two types of data by using a one-dimensional time sequence convolutional network and a two-dimensional convolutional network respectively; introducing a cross attention feature deep fusion mechanism to extract feature information sensitive to surface collapse in the two types of data, and establishing a reliable mapping relationship between the monitoring data and the collapse risk; a risk level is output and an early warning signal is generated through a risk identification and early warning framework based on an attention mechanism, so that accurate identification and early warning of surface collapse are realized. According to the method, the accuracy of surface collapse risk identification and the early warning timeliness are remarkably improved, and the defects of high false alarm rate, early warning delay and the like in a traditional method are effectively overcome.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, and in particular to a method for identifying and warning of ground subsidence risk based on a multi-parameter coupled temporal convolutional network. Background Technology

[0002] Currently, my country's urban development has shifted from large-scale incremental construction to a stage of improving the quality of existing infrastructure with safety as the core. As a complex mega-system, the safe operation of urban underground space has become a crucial aspect of urban governance. Against this backdrop, ground subsidence, a typical geological hazard characterized by its sudden onset and high destructiveness, presents significant challenges to accurate identification and early warning.

[0003] In existing technologies, fiber optic sensing and parallel electrical resistivity methods are two effective detection methods: the former can sensitively capture stress redistribution and structural deformation within the soil by monitoring micro-strain, while the latter can directly reflect soil erosion and cavity development by detecting resistivity changes. However, traditional analysis methods process these two aspects independently, making it difficult to distinguish between local disturbances and precursors of overall instability in strain data, leading to false alarms; and making it difficult to distinguish between geological heterogeneity and actual hazards in resistivity anomalies, leading to missed diagnoses. This data barrier results in a severe lag in the system's response to comprehensive precursor characteristics, and a severely insufficient early warning window.

[0004] The development of artificial intelligence technology, especially the deep learning paradigm, has provided a new path to solve the above problems. By constructing a deep learning architecture capable of simultaneously processing time-series strain data and spatial resistivity images, not only can modality-specific features of the two types of data be extracted separately, but also mutual verification of mechanical and electrical signals can be achieved through an advanced feature interaction mechanism. This dual-drive analysis method can accurately capture weak but significantly enhanced pre-collapse features in a single data stream under multi-parameter coupling, thereby driving a fundamental shift in ground subsidence early warning from post-event statistics to pre-event prediction. Summary of the Invention

[0005] To address the problems of high false alarm rate and delayed early warning in traditional methods, this invention proposes a multi-parameter coupled ground subsidence risk identification and early warning method based on temporal convolutional networks. This method integrates fiber optic grating sensing and parallel electrical resistivity data, and uses a deep learning model for multi-parameter coupled analysis, effectively improving the accuracy of ground subsidence risk identification and the timeliness of early warning.

[0006] This invention adopts the following technical solution: a method for identifying and warning of multi-parameter coupled ground subsidence risk based on temporal convolutional networks, comprising the following steps:

[0007] S10. One-dimensional strain time-series data of the pipeline and surrounding soil are collected by fiber optic grating (FBG) sensor arrays deployed in the monitoring area, and two-dimensional resistivity spatial distribution data of the soil are collected by parallel electrical resistivity test (ERT) system. FBG data and parallel electrical resistivity test data are synchronously collected through a unified timing device.

[0008] S20. The original FBG data is denoised using wavelet transform, and the original parallel electrical method is inverted and imaged using smooth constrained least squares. The preprocessed multi-source data is unified to the same timestamp sequence. The input sample dataset is constructed using the sliding window method and divided into training set, validation set and test set according to the time series.

[0009] S30. Construct a dual-branch deep learning network, use a one-dimensional temporal convolutional network (1D-TCN) to extract one-dimensional temporal features of FBG data and capture long-term dependencies in FBG data; use a two-dimensional convolutional neural network (2D-CNN) to extract two-dimensional spatial features of parallel electrical resistivity data and capture spatial anomalous patterns in resistivity images.

[0010] S40. Flatten the extracted one-dimensional temporal features and two-dimensional spatial features and use them as input. Calculate the correlation weight between the two types of features dynamically through a cross-attention mechanism to generate a unified spatiotemporal fusion feature.

[0011] S50. Input the spatiotemporal fusion features into the classifier and output a continuous probability value representing the collapse risk. During the training phase, optimize the classifier parameters by minimizing the binary cross-entropy loss function.

[0012] S60. Based on the risk probability value at the current moment and the linear regression slope of the most recent time windows, the trend of change is calculated. Combined with the preset probability threshold and rate of change threshold, different levels of risk warning signals are triggered to achieve dynamic graded warning from safe to high risk.

[0013] Preferably, in step S10, the FBG sensor array is arranged at intervals along the pipeline and connected in series by optical fibers to form a one-dimensional distributed strain monitoring network to collect FBG strain signals.

[0014] The parallel electrical resistivity system deploys an electrode array on the ground directly above the pipeline, using Wenner devices arranged in a grid layout to collect apparent resistivity data. Synchronous data acquisition is achieved through a multi-channel resistivity meter.

[0015] The unified time synchronization device uses a GPS synchronous clock module to provide a unified time reference for the FBG sensor array and the parallel electrical simulation system.

[0016] Preferably, the preprocessing and standardization of FBG data and parallel electrical resistivity data in step S20 specifically includes the following sub-steps:

[0017] S21. Signal denoising: Wavelet transform is used to denoise the original FBG strain signal. The wavelet coefficient thresholding function is as follows:

[0018] ;

[0019] In the formula, The wavelet coefficients and λ are threshold values, adaptively determined using the Stein unbiased risk estimation criterion. It is a symbolic function.

[0020] S22. Time-series sample construction: The sliding window method is used to construct the preprocessed continuous one-dimensional strain sequence. Reconstructed into model input samples, Where L is the window length and N is the number of sensor channels.

[0021] S23. Inversion imaging: The smoothed constrained least squares method is used to invert the original apparent resistivity data. The objective function is:

[0022] ;

[0023] In the formula, For the observed data vector, For the orthogonal operator, This is the model parameter vector (true resistivity). For the data weight matrix, The smooth constraint matrix is, This is the regularization parameter.

[0024] S24: Image normalization, for the inverted resistivity image Perform pixel-level Z-score normalization:

[0025] ;

[0026] In the formula, and These are the pixels in the training set. The mean and standard deviation of the resistivity values.

[0027] S25: Perform time alignment and spatial correlation on fiber optic monitoring data and parallel electrical resistivity data.

[0028] Preferably, time alignment is based on the time points of parallel electrical resistivity data acquisition. Based on the baseline, extract the time window. The FBG data within the window is used to calculate the statistical characteristics (mean, variance, maximum value) of the FBG data within that window, forming a feature vector that is time-aligned with the electrical resistivity data. ;

[0029] Preferably, spatial association is achieved through a coordinate mapping function. The three-dimensional spatial coordinates of each FBG sensor are projected onto the pixel coordinates of the two-dimensional resistivity cross-sectional image to establish the correspondence between the spatial position of the sensor and the image region.

[0030] Preferably, in step S30, the network used to extract the one-dimensional temporal features of the FBG data is a one-dimensional temporal convolutional network, which is composed of multiple stacked temporal convolutional blocks. Each temporal convolutional block includes a one-dimensional convolutional layer, a batch normalization layer, and... Activation function.

[0031] The one-dimensional convolutional layer uses dilated convolution to capture long-term dependencies, and its output features are represented as follows:

[0032] ;

[0033] In the formula, Indicates the first The output features of the layer The input features (for the first layer, the input is FBG data) are the input features. , For time steps, (for feature dimensions) and These are the convolution weights and biases, respectively. This represents a one-dimensional convolution operation, and BN represents batch normalization. The activation function is used; the dilation factor of the dilated convolution increases exponentially with the network depth, i.e., the first... The expansion factor of the layer is .

[0034] Preferably, in step S30, the network used to extract the two-dimensional spatial features of the parallel electrical resistivity data is a two-dimensional convolutional neural network, consisting of multiple alternating convolutional layers and pooling layers, with each convolutional layer followed by a batch normalization layer and... Live function.

[0035] The output feature map of the two-dimensional convolutional neural network is represented as follows:

[0036] ;

[0037] In the formula, Indicates the first The output feature map of the layer, The input features (for the first layer, the input is parallel electrical resistivity data) are the input features. , and For the height and width of the space, (Number of channels) and These are the convolution weights and biases, respectively. This represents a two-dimensional convolution operation.

[0038] Preferably, in step S40, the correlation weights between the one-dimensional temporal features and the two-dimensional spatial features are calculated using a cross-attention mechanism to generate unified spatiotemporal fusion features, specifically including the following sub-steps:

[0039] S41. Feature preprocessing: extracting two-dimensional spatial feature maps... The sequence is flattened and converted into a spatial feature sequence, where... and The height and width of the feature map respectively. The number of feature channels, The total number of spatial locations;

[0040] S42. For one-dimensional temporal and spatial feature sequences respectively. Perform linear transformations to generate the query matrix, key matrix, and value matrix:

[0041] ;

[0042] ;

[0043] ;

[0044] In the formula, For learnable weight matrix, and These are the dimensions of the key and the value, respectively. Representing one-dimensional time series features Feature dimensions;

[0045] S43. Based on the query matrix and key matrix, calculate the cross-attention weight matrix A between one-dimensional temporal features and two-dimensional spatial features:

[0046] ;

[0047] In the formula, The function is computed along each row of the matrix to ensure that the sum of the attention weights for all spatial locations at each time step is 1; the attention weight matrix, in which each element... Each element This represents the correlation weight between the i-th time step and the j-th spatial location.

[0048] S44. Using the attention weight matrix The value matrix V is weighted and summed to generate a unified spatiotemporal fusion feature F, which in turn generates a corresponding context vector for each time step.

[0049] ;

[0050] In the formula, For unified spatiotemporal fusion characteristics.

[0051] Preferably, the classifier in step S50 is a fully connected neural network layer followed by... The activation function is calculated as follows:

[0052] S51. The spatiotemporal fusion feature F is passed through a global average pooling layer to aggregate the entire temporal context information, resulting in a global feature vector:

[0053] ;

[0054] In the formula, Let F represent the eigenvector of F at time step t.

[0055] S52, global feature vector Input is fed into the fully connected layer, and through The function is activated, and the output is a continuous probability value representing the risk of collapse. :

[0056] ;

[0057] In the formula, and The weights and bias parameters of the fully connected layer are... for function, The closer the value is to 1, the higher the risk of collapse.

[0058] S53. During the training phase, minimize the probability of predicted risk. With true binary labels The classifier's parameters are optimized using a binary cross-entropy loss function (0 for safe, 1 for collapsed). Defined as:

[0059] ;

[0060] in, The number of samples in a training batch. and The first The true label and predicted probability of each sample.

[0061] Preferably, in step S60, based on the risk probability value at the current moment... and its rate of change Joint decision-making, specifically including:

[0062] S61, based on the probability at the current moment Probability compared to the previous moment The difference is used to approximate the instantaneous rate of change of risk probability. :

[0063] ;

[0064] S62, Set probability threshold (in and rate of change threshold ;

[0065] S63. Trigger different levels of risk warning signals according to the following logical rules:

[0066] like If this occurs, a red alert (the highest level) will be immediately triggered, indicating a high risk.

[0067] like and If this occurs, an orange alert (medium level) will be triggered, indicating that the risk is increasing significantly;

[0068] like and If this occurs, a yellow alert (attention level) will be triggered, indicating that there is a risk but the trend is stable;

[0069] like If the indicator is blue, the system status is safe (no warning).

[0070] S64. The calculation of the risk probability change rate Rt is further optimized to be based on the slope of a linear regression over the most recent k time windows, in order to smooth short-term fluctuations and capture stable trend directions:

[0071] ;

[0072] In the formula, This is the average time index within the time window. The slope is the average of the risk probability values ​​within the corresponding window. The magnitude of the sign indicates the rate of change, and the sign indicates the direction of change.

[0073] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0074] 1. The method of this invention breaks through the limitations of traditional single-parameter monitoring. It couples one-dimensional strain time-series data reflecting mechanical behavior with two-dimensional resistivity spatial data reflecting the electrical properties of soil and rock media. This enables cross-verification and complementary information, and more comprehensively captures multiple precursor features such as soil loosening, water content changes, and void development before collapse, fundamentally improving the dimension and reliability of risk perception.

[0075] 2. The method of the present invention adopts a dual-branch architecture consisting of temporal convolutional networks and convolutional neural networks, which performs modality-specific deep feature extraction for the characteristics of temporal and spatial data respectively, significantly improving the ability and efficiency of capturing key risk features.

[0076] 3. This invention introduces a cross-attention mechanism, which differs from simple feature splicing or addition. It dynamically and adaptively allows for deep fusion of features from two modes, enabling intelligent discovery of the intrinsic coupling relationship between strain and resistivity change, thereby greatly reducing the false alarm rate.

[0077] 4. The method of this invention can accurately identify early and weak nonlinear anomaly patterns, enabling early warning of ground subsidence and significantly improving the timeliness of the warning. At the same time, the decision-making mechanism based on multi-parameter cross-validation ensures the high reliability of the warning signal. It can achieve full automation from data collection, processing, analysis to warning release, reducing manual intervention and improving the efficiency and stability of the system's engineering applications. Attached Figure Description

[0078] Figure 1 This is a flowchart of the multi-parameter coupled ground subsidence risk identification and early warning method of the present invention;

[0079] Figure 2 This is a flowchart of the dual-branch network of the present invention;

[0080] Figure 3 This is a diagram of the D-TCN network structure of the present invention;

[0081] Figure 4 This is a diagram of the D-CNN network structure of the present invention. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0083] In one embodiment of the present invention, the method of the present invention is used for monitoring and early warning of ground subsidence risk in urban underground pipeline areas. The monitoring area is an underground water supply pipeline section in the city center, approximately 500 meters long and 5 meters deep. The soil in this area is soft clay, which is prone to ground subsidence due to pipeline leakage or soil erosion.

[0084] In this embodiment, the process of the multi-parameter coupled ground subsidence risk identification and early warning method based on temporal convolutional networks is as follows: Figure 1 As shown, the steps are as follows:

[0085] Step 1: Sensor Deployment

[0086] (1) Deploy the FBG sensor array

[0087] An FBG sensor is deployed every 10 meters along the pipeline, with a total of 50 sensor nodes forming a one-dimensional distributed strain monitoring network. The sensor sampling frequency is set to 10Hz, which means that strain data is collected once every 0.1 seconds. The sensors are connected in series via optical fibers to collect FBG strain signals, and the data is transmitted to the central processing unit in real time via a demodulator.

[0088] (2) Deploy the parallel electrical tracer system (ERT)

[0089] An electrode array was installed on the ground directly above the pipeline, using Wenner devices arranged with a 2-meter spacing between electrodes to form a 20-row × 10-column grid layout, totaling 200 electrodes. The sampling frequency was set to 1 Hz, meaning that apparent resistivity data was collected once per second, with synchronous data acquisition achieved through a multi-channel resistivity meter.

[0090] (3) Deploy a unified time synchronization device

[0091] A GPS-synchronized clock module is used to provide a unified time reference for the FBG demodulator and electrical simulation system, with a time synchronization accuracy better than 1 millisecond, ensuring the time consistency of multi-source data.

[0092] Step 2: Data Preprocessing and Sample Construction

[0093] (1) FBG signal noise reduction processing

[0094] The original FBG strain signal contains high-frequency noise (such as traffic vibration interference). Wavelet transform is used for noise reduction, with the wavelet basis function being "db4". The threshold λ is adaptively calculated using the Stein unbiased risk estimation criterion.

[0095] (2) ERT data inversion imaging

[0096] The inversion is performed using the smooth constrained least squares method, with the objective function being:

[0097] ;

[0098] In the formula, For the observed data vector, For the orthogonal operator, This is the model parameter vector (true resistivity). For the data weight matrix, The smooth constraint matrix is, This is the regularization parameter.

[0099] Set regularization parameters If the inversion converges, the true resistivity two-dimensional cross-sectional image will be output. .

[0100] (3) Spatiotemporal alignment of multi-source data

[0101] Time alignment and spatial correlation are performed on fiber optic monitoring data and parallel electrical resistivity data.

[0102] Time alignment: based on the time points of parallel electrical resistivity data acquisition. Based on the baseline, extract the time window. The FBG data within, including Seconds. Calculate the statistical characteristics (mean, variance, maximum value) of each FBG sensor within this window, forming a feature vector. .

[0103] Spatial association: through coordinate mapping functions The three-dimensional spatial coordinates of each FBG sensor are projected onto the pixel coordinates of the ERT resistivity image.

[0104] For example, sensor 1 located at (0,0,5) is mapped to the pixel region of the 5th row and 25th column of the resistivity image.

[0105] (4) Sample construction and dataset partitioning

[0106] The input samples are constructed using the sliding window method, and the length of the FBG data window is determined. (Corresponding to 10 seconds of data), number of channels Input Sample Each ERT data point corresponds to a resistivity image. The result was normalized using Z-Score.

[0107] The collected data from 30 consecutive days was further divided into training, validation, and test sets in a ratio of 7:1.5:1.5.

[0108] Step 3: Dual-branch deep learning network

[0109] First, construct a two-branch deep learning network, such as... Figure 2 As shown, a one-dimensional temporal convolutional network and a two-dimensional convolutional network are used to extract modality-specific features from the two types of data, respectively. The collected data is preprocessed in step S20 and then input into the dual-branch network.

[0110] Then, through, as Figure 3 The time-series branch 1D-TCN shown extracts one-dimensional time-series features from FBG data. The input dimensions are: L=100 time steps and N=20 sensor channels.

[0111] One-dimensional temporal convolutional networks (1D-TCNs) are composed of multiple stacked temporal convolutional blocks. Each temporal convolutional block includes a one-dimensional convolutional layer, a batch normalization layer, and... Activation function.

[0112] At the same time, through such Figure 4 The spatial branch of 2D-CNN is shown to extract two-dimensional spatial features from parallel electrical resistivity data, with the input being a normalized resistivity image.

[0113] A two-dimensional convolutional neural network consists of multiple alternating convolutional and pooling layers, with each convolutional layer followed by a batch normalization layer and... Activation function.

[0114] The pooling layer uses max pooling, and the pooling kernel size is [size missing]. The step size is 2.

[0115] Step 4: Through the feature fusion module, the 256-dimensional one-dimensional temporal features output by 1D-TCN and the flattened 1024-dimensional two-dimensional spatial features output by 2D-CNN are subjected to cross-attention dynamic calculation to generate a unified 256-dimensional spatiotemporal fusion feature.

[0116] Step 5: Input the spatiotemporal fusion features into a classifier, which is a fully connected neural network layer followed by... The activation function is used to classify the risk probability P that the classifier outputs.

[0117] The warning threshold is set as follows: , , The time window k=5.

[0118] Step 6: Risk Classification and Early Warning

[0119] Based on the probability at the current moment Probability compared to the previous moment The difference is used to approximate the instantaneous rate of change of risk probability. Based on the risk probability value at the current moment and its rate of change Joint decision-making risk classification is as follows:

[0120] like If this occurs, a red alert (the highest level) will be immediately triggered, indicating a high risk.

[0121] like and If this occurs, an orange alert (medium level) will be triggered, indicating that the risk is increasing significantly;

[0122] like and If this occurs, a yellow alert (attention level) will be triggered, indicating that there is a risk but the trend is stable;

[0123] like If the color is displayed in blue, the system status is safe (no warning).

[0124] In summary, the present invention provides a method for identifying and warning of ground subsidence risk based on multi-parameter coupled temporal convolutional networks. This method significantly improves the accuracy of ground subsidence risk identification and the timeliness of warning, effectively overcoming the shortcomings of traditional methods such as high false alarm rate and warning delay. It can realize a fundamental shift in ground subsidence warning from post-event statistics to pre-event prediction.

[0125] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying and warning of ground subsidence risk based on multi-parameter coupled convolutional networks, characterized in that, Includes the following steps: S10. One-dimensional strain time-series data of the pipeline and surrounding soil are collected by FBG sensor arrays deployed in the monitoring area, and two-dimensional resistivity spatial distribution data of the soil are collected by parallel electrical resistivity system. FBG data and parallel electrical resistivity data are synchronously collected by a unified timing device. S20. The original FBG data is denoised using wavelet transform, and the original parallel electrical method is inverted and imaged using smooth constrained least squares. The preprocessed multi-source data is unified to the same timestamp sequence. The input sample dataset is constructed using the sliding window method and divided into training set, validation set and test set according to the time series. S30. Construct a dual-branch deep learning network, use a one-dimensional temporal convolutional network to extract one-dimensional temporal features of FBG data and capture long-term dependencies in FBG data; use a two-dimensional convolutional neural network to extract two-dimensional spatial features of parallel electrical resistivity data and capture spatial anomaly patterns in resistivity images. S40. Flatten the extracted one-dimensional temporal features and two-dimensional spatial features and use them as input. Calculate the correlation weight between the two types of features dynamically through a cross-attention mechanism to generate a unified spatiotemporal fusion feature. S50. Input the spatiotemporal fusion features into the classifier and output a continuous probability value representing the collapse risk. During the training phase, optimize the classifier parameters by minimizing the binary cross-entropy loss function. S60. Based on the risk probability value at the current moment and the linear regression slope of the most recent time windows, the trend of change is calculated. Combined with the preset probability threshold and rate of change threshold, different levels of risk warning signals are triggered to achieve dynamic graded warning from safe to high risk.

2. The method for identifying and warning of ground subsidence risk based on multi-parameter coupled convolutional networks according to claim 1, characterized in that, In step S10, the FBG sensor array is arranged at intervals along the pipeline and connected in series by optical fibers to form a one-dimensional distributed strain monitoring network to collect FBG strain signals. The parallel electrical resistivity system deploys an electrode array on the ground directly above the pipeline, using Wenner devices arranged in a grid layout to collect apparent resistivity data. Synchronous data acquisition is achieved through a multi-channel resistivity meter. The unified time synchronization device uses a GPS synchronous clock module to provide a unified time reference for the FBG sensor array and the parallel electrical simulation system.

3. The method for identifying and warning of ground subsidence risk based on multi-parameter coupled convolutional networks according to claim 2, characterized in that, Step S20 involves preprocessing and standardizing the raw FBG data and parallel electrical resistivity data, including the following sub-steps: S21. Signal Denoising: The original FBG strain signal is denoised using wavelet transform. The wavelet coefficient thresholding function is as follows: ； In the formula, λ is the wavelet coefficient, and λ is the threshold, which is adaptively determined by the Stein unbiased risk estimation criterion. It is a symbolic function; S22. Time-series sample construction: The sliding window method is used to construct the preprocessed continuous one-dimensional strain sequence. Reconstructed into model input samples Where L is the window length and N is the number of sensor channels; S23. Inversion Imaging: The original apparent resistivity data is inverted using the smooth constrained least squares method. The objective function is: ； In the formula, For the observed data vector, For the orthogonal operator, For model parameter vectors, For the data weight matrix, The smooth constraint matrix is For regularization parameters; S24. Image normalization: Normalize the resistivity image obtained from the inversion. Perform pixel-level Z-score normalization: ； In the formula, and Pixels on the training set The mean and standard deviation of the resistivity values; S25. Perform time alignment and spatial correlation on FBG data and parallel electrical resistivity data.

4. The method for identifying and warning of ground subsidence risk based on multi-parameter coupled convolutional networks according to claim 3, characterized in that, In step S25, the time alignment is based on the parallel electrical resistivity data acquisition time point. Based on the baseline, extract the time window. The FBG data within the window is used to calculate the statistical characteristics of the data within that window, including the mean, variance, and maximum value, forming a feature vector that is time-aligned with the electrical resistivity data. ; The spatial association is achieved through a coordinate mapping function. The three-dimensional spatial coordinates of each FBG sensor Pixel coordinates projected onto a two-dimensional resistivity cross-sectional image The correspondence between the spatial location of the FBG sensor and the image region is established.

5. The method for identifying and warning of ground subsidence risk based on multi-parameter coupled ground subsidence risk according to claim 3, characterized in that, In step S30, the one-dimensional temporal convolutional network is formed by stacking multiple temporal convolutional blocks, each temporal convolutional block including: a one-dimensional convolutional layer, a batch normalization layer, and... Activation function; The one-dimensional convolutional layer uses dilated convolution to capture long-term dependencies, and the output feature is represented as follows: ； In the formula, Indicates the first The output features of the layer These are input features. and These are the convolution weights and biases, respectively. This represents a one-dimensional convolution operation, and BN represents batch normalization. For activation functions; The dilation factor of the dilated convolution increases exponentially with the network depth, the th... The expansion factor of the layer is .

6. The method for identifying and warning of ground subsidence risk based on multi-parameter coupled convolutional networks according to claim 3, characterized in that, In step S30, the two-dimensional convolutional neural network consists of multiple alternating convolutional layers and pooling layers, with each convolutional layer followed by a batch normalization layer and... Activation function; The output feature map of the two-dimensional convolutional neural network is represented as follows: ； In the formula, Indicates the first The output feature map of the layer, These are input features. and These are the convolution weights and biases, respectively. Represents a two-dimensional convolution operation; The pooling layer uses max pooling, and the pooling kernel size is [size missing]. The step size is 2.

7. The method for identifying and warning of ground subsidence risk based on multi-parameter coupled ground subsidence risk according to claim 3, characterized in that, In step S40, generating unified spatiotemporal fusion features includes the following sub-steps: S41. Feature preprocessing: extracting two-dimensional spatial feature maps... Flatten the image and convert it into a spatial feature sequence. ,in, and The height and width of the feature map respectively. The number of feature channels, The total number of spatial locations; S42. For one-dimensional time series features respectively and spatial feature sequences Perform a linear transformation to generate the query matrix. Key matrix Sum matrix : ； ； ； In the formula, For learnable weight matrix, and These are the dimensions of the key and the value, respectively. Representing one-dimensional time series features Feature dimensions; S43. Based on the query matrix and key matrix, calculate the cross-attention weight matrix A between one-dimensional temporal features and two-dimensional spatial features: ； In the formula, The function is computed along each row of the matrix to ensure that the sum of the attention weights for all spatial locations at each time step is 1; S44. Utilizing the cross-attention weight matrix The value matrix V is weighted and summed to generate a unified spatiotemporal fusion feature F, which in turn generates a corresponding context vector for each time step. ； In the formula, For unified spatiotemporal fusion characteristics.

8. The method for identifying and warning of ground subsidence risk based on multi-parameter coupled convolutional networks according to claim 7, characterized in that, In step S50, the classifier is a fully connected neural network layer followed by... The activation function is calculated as follows: S51. The spatiotemporal fusion feature F is passed through a global average pooling layer to aggregate the entire temporal context information, resulting in a global feature vector. : ； In the formula, This represents the eigenvector of F at time step t; S52, global feature vector Input is fed into the fully connected layer, and through The function is activated, and outputs a continuous probability value P representing the risk of collapse: ； In the formula, and The weights and bias parameters of the fully connected layer are... for function, The closer the value is to 1, the higher the risk of collapse. S53. During the training phase, the parameters of the classifier are optimized by minimizing the binary cross-entropy loss between the predicted risk probability P and the true binary label Y. The loss function is... Defined as: ； Where M is the number of samples in a training batch. and Let be the true label and the predicted probability of the i-th sample, respectively.

9. The method for identifying and warning of ground subsidence risk based on multi-parameter coupled convolutional networks according to claim 8, characterized in that, In step S60, based on the risk probability value at the current moment... and its rate of change The joint decision-making process for determining the level of risk warning signals includes the following sub-steps: S61, based on the probability at the current moment Probability compared to the previous moment The difference is used to approximate the instantaneous rate of change of risk probability. : ； S62, Set probability threshold and rate of change threshold ,in, , ; S63, Based on the probability at the current time The relationship between probability thresholds and rate of change thresholds triggers different levels of risk warning signals; S64. Calculate the rate of change of risk probability Rt, based on the linear regression slope of the most recent k time windows, to smooth short-term fluctuations and capture stable trend directions: ； In the formula, This is the average time index within the time window. The slope is the average of the risk probability values ​​within the corresponding window. The magnitude of the sign indicates the rate of change, and the sign indicates the direction of change.

10. The method for identifying and warning of ground subsidence risk based on multi-parameter coupled ground subsidence risk according to claim 9, characterized in that, In step S63, risk warning signals of different levels are triggered, and the logical rules are as follows: like This immediately triggers the highest level of warning, indicating a high risk. like and This triggers a medium-level warning, indicating that the risk is increasing significantly. like and This triggers a level-of-concern alert, indicating that there is risk but the trend is stable. like No warning was issued; the system status is safe.

Images