Karst geological disaster multi-source fusion prediction method, device and equipment and storage medium

Abstract

The application discloses a karst geological disaster multi-source fusion prediction method, device and equipment and a storage medium. The method obtains the resistivity, wave velocity drop and line dissolution rate of historical karst geological samples, pre-processes the resistivity, wave velocity drop and line dissolution rate, and obtains training feature data. A target model is constructed and trained based on the training feature data, and the target model is obtained. Real-time geophysical prospecting data of a target area is pre-processed and input into the target model, and a karst development risk grade is output. The method can realize accurate prediction of karst risk, improve the accuracy of karst disaster detection, reduce the false negative rate of high-risk karst, significantly solve the high-risk false negative problem caused by the single data, the lack of spatial correlation and the unbalanced samples in the prior art, improve the prediction accuracy of karst disasters, avoid construction risks, improve the identification ability of hidden karst, and improve the speed and efficiency of karst geological disaster multi-source fusion prediction.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of geological exploration and artificial intelligence, and in particular to a method, device, equipment and storage medium for multi-source fusion prediction of karst geological disasters. Background Technology

[0002] Karst geological hazards (such as tunnel water inrush and ground subsidence) are major hidden dangers in engineering construction, and their detection accuracy directly affects construction safety and cost; existing technologies have the following shortcomings:

[0003] Limitations of data singularity: Currently, detection relies heavily on single physical field data, resulting in limited accuracy, prediction blind spots, and a lack of multi-physics coupling and borehole data verification.

[0004] Insufficient spatial correlation modeling: Karst development has a certain degree of spatial aggregation. Traditional LSTM and CNN models can only process sequence or local spatial features and cannot capture large-scale spatial topological relationships.

[0005] Imbalanced sample problem: Disaster samples (such as karst caves and collapse points) account for a small proportion in actual engineering. Traditional MSE loss function-dominated models tend to ignore minority class samples, resulting in a high rate of disaster underreporting.

[0006] Existing patent CN118962819A proposes a karst geological region prediction method based on the Inception-v4-Kansformer multi-module model, which mainly addresses the problem of "insufficient feature extraction accuracy of single ground-penetrating radar data". It improves the accuracy and efficiency of ground-penetrating radar data and image processing through multi-scale convolution of Inception-v4 and nonlinear fitting of KAN. However, it does not propose a systematic solution to problems such as data uniformity, insufficient modeling of spatial correlation, and imbalanced samples, and has defects in its use. Summary of the Invention

[0007] The main objective of this invention is to provide a multi-source fusion prediction method, device, equipment, and storage medium for karst geological hazards, aiming to solve the technical problems of high underreporting rate of high-risk karst and low accuracy of karst hazard detection caused by data uniformity, lack of spatial correlation, and sample imbalance in the existing technology.

[0008] In a first aspect, the present invention provides a multi-source fusion prediction method for karst geological hazards, the multi-source fusion prediction method for karst geological hazards comprising the following steps:

[0009] Resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples are obtained. The resistivity, wave velocity reduction, and linear dissolution rate are preprocessed to obtain training feature data.

[0010] Based on the training feature data, a Transformer model is constructed and trained to obtain the target model;

[0011] After preprocessing the real-time geophysical data of the target area, the data is input into the target model, and the karst development risk level is output.

[0012] Optionally, the acquisition of resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples, and the preprocessing of the resistivity, wave velocity reduction, and linear dissolution rate to obtain training feature data, include:

[0013] The resistivity of historical karst geological samples was obtained by an electrical resistivity analyzer, the wave velocity reduction of the historical karst geological samples was simultaneously collected by a micro-motion sensor, and the linear solubility of the historical karst geological samples was simultaneously collected by a borehole core scanner.

[0014] The resistivity, wave velocity reduction, and linear melting rate are standardized and subjected to wavelet denoising to obtain training feature data.

[0015] Optionally, the standardization and wavelet denoising of the resistivity, wave velocity reduction, and linear melting rate to obtain training feature data includes:

[0016] Z-Score normalization was performed on the resistivity, the wave velocity reduction, and the linear melting rate to obtain normalized data;

[0017] The standardized data is decomposed into three levels using the db4 wavelet basis. The detailed components containing karst features in the standardized data are then thresholded and enhanced. The enhanced components are then reconstructed to obtain training feature data.

[0018] Optionally, the step of constructing and training a Transformer model based on the training feature data to obtain the target model includes:

[0019] Based on the training feature data, a Transformer model is constructed that includes a multimodal embedding layer, address semantic location encoding, and a self-attention layer;

[0020] The proportion of disaster samples in the training feature data is dynamically adjusted according to a preset weighted loss function, and the Transformer model is trained according to the adjusted proportion of disaster samples to obtain the target model.

[0021] Optionally, the step of constructing a Transformer model based on the training feature data, comprising a multimodal embedding layer, address semantic location encoding, and a self-attention layer, includes:

[0022] The training feature data is mapped to a 256-dimensional space through a multimodal embedding layer, and the feature data of each detection point is combined into a three-dimensional vector, which is then embedded to form a 256-dimensional feature vector.

[0023] The planar coordinates of each detection point are converted into position vectors, and a 256-dimensional position encoding vector is calculated based on the position vectors using a preset geological semantic position encoding formula.

[0024] The 256-dimensional feature vector is added to the 256-dimensional position encoding vector and then input into the self-attention layer to construct the Transformer model.

[0025] Optionally, the step of converting the planar coordinates of each detection point into a position vector, and calculating a 256-dimensional position encoding vector based on the position vector using a preset geological semantic position encoding formula, includes:

[0026] The planar coordinates of each detection point are converted into position vectors. Then, a 256-dimensional position encoding vector is obtained by calculating the position vectors using a preset geological semantic position encoding formula:

[0027]

[0028]

[0029] in, The x-coordinate of the measuring point in the plane coordinate system. This represents the y-coordinate of the measuring point in the plane coordinate system. Dimension index for 256-dimensional features The frequency factor corresponding to the x-coordinate. The frequency factor corresponding to the y-coordinate. It is a 256-dimensional position encoding vector with an even number of dimensions. It is a 256-dimensional positional encoding vector with an odd number of dimensions.

[0030] Optionally, the step of preprocessing real-time geophysical data of the target area and inputting it into the target model to output the karst development risk level includes:

[0031] The real-time geophysical data of the target area is preprocessed to obtain structured input data, and the structured input data is then input into the target model.

[0032] Obtain the karst development risk probability output by the target model, match the karst development risk probability with a preset probability threshold range, and obtain the karst development risk level corresponding to the karst development risk probability.

[0033] Secondly, to achieve the above objectives, the present invention also proposes a multi-source fusion prediction device for karst geological hazards, the multi-source fusion prediction device for karst geological hazards comprising:

[0034] The data preprocessing module is used to obtain the resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples, and to preprocess the resistivity, wave velocity reduction, and linear dissolution rate to obtain training feature data.

[0035] The model training module is used to construct and train a Transformer model based on the training feature data to obtain the target model;

[0036] The risk analysis module is used to preprocess real-time geophysical data of the target area and input it into the target model, and output the risk level of karst development.

[0037] Thirdly, to achieve the above objectives, the present invention also proposes a karst geological hazard multi-source fusion prediction device, which includes: a memory, a processor, and a karst geological hazard multi-source fusion prediction program stored in the memory and executable on the processor. The karst geological hazard multi-source fusion prediction program is configured to implement the steps of the karst geological hazard multi-source fusion prediction method as described above.

[0038] Fourthly, to achieve the above objectives, the present invention also proposes a storage medium storing a multi-source fusion prediction program for karst geological hazards, wherein when the multi-source fusion prediction program for karst geological hazards is executed by a processor, the program implements the steps of the multi-source fusion prediction method for karst geological hazards as described above.

[0039] The multi-source fusion prediction method for karst geological hazards proposed in this invention obtains resistivity, wave velocity reduction, and linear dissolution rate from historical karst geological samples. The resistivity, wave velocity reduction, and linear dissolution rate are preprocessed to obtain training feature data. A Transformer model is constructed and trained based on the training feature data to obtain a target model. Real-time geophysical data of the target area is preprocessed and input into the target model to output the karst development risk level. This method enables accurate prediction of karst risk, improves the accuracy of karst hazard detection, reduces the underreporting rate of high-risk karst, and significantly solves the problems of data uniformity, lack of spatial correlation, and sample imbalance leading to high-risk underreporting in existing technologies. It improves the accuracy of karst hazard prediction, avoids construction risks, enhances the identification ability of concealed karst, and improves the speed and efficiency of multi-source fusion prediction of karst geological hazards. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention;

[0041] Figure 2 This is a flowchart illustrating the first embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention;

[0042] Figure 3 This is a flowchart illustrating the second embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention;

[0043] Figure 4 This is a flowchart illustrating the third embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention;

[0044] Figure 5 This is a schematic diagram of the multi-source data fusion logic in the multi-source fusion prediction method for karst geological hazards of the present invention;

[0045] Figure 6 This is a schematic diagram of the Transformer model architecture in the multi-source fusion prediction method for karst geological hazards of the present invention;

[0046] Figure 7 This is a flowchart illustrating the fourth embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention;

[0047] Figure 8 This is a comparison chart of the implementation effects of the present invention, BP neural network, and LSTM;

[0048] Figure 9 This is a functional block diagram of the first embodiment of the karst geological disaster multi-source fusion prediction device of the present invention.

[0049] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0050] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] The solution of this invention mainly involves: acquiring the resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples; preprocessing the resistivity, wave velocity reduction, and linear dissolution rate to obtain training feature data; constructing and training a Transformer model based on the training feature data to obtain a target model; inputting preprocessed real-time geophysical data of the target area into the target model to output the karst development risk level, thereby achieving accurate prediction of karst risk, improving the accuracy of karst disaster detection, reducing the underreporting rate of high-risk karst, significantly solving the problem of high-risk underreporting caused by data uniformity, lack of spatial correlation, and sample imbalance in existing technologies, improving the accuracy of karst disaster prediction, avoiding construction risks, enhancing the ability to identify concealed karst, and improving the speed and efficiency of multi-source fusion prediction of karst geological disasters. This solves the technical problems of high underreporting rate of high-risk karst and low accuracy of karst disaster detection caused by data uniformity, lack of spatial correlation, and sample imbalance in existing technologies.

[0052] Reference Figure 1 , Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention.

[0053] like Figure 1 As shown, the device may include: a processor 1001, such as a CPU; a communication bus 1002; a user interface 1003; a network interface 1004; and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0054] Those skilled in the art will understand that Figure 1 The device structure shown does not constitute a limitation on the device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0055] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating device, a network communication module, a user interface module, and a multi-source fusion prediction program for karst geological disasters.

[0056] The device of this invention calls the karst geological hazard multi-source fusion prediction program stored in the memory 1005 through the processor 1001, and performs the following operations:

[0057] Resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples are obtained. The resistivity, wave velocity reduction, and linear dissolution rate are preprocessed to obtain training feature data.

[0058] Based on the training feature data, a Transformer model is constructed and trained to obtain the target model;

[0059] After preprocessing the real-time geophysical data of the target area, the data is input into the target model, and the karst development risk level is output.

[0060] The device of this invention, through processor 1001 calling the multi-source fusion prediction program for karst geological hazards stored in memory 1005, also performs the following operations:

[0061] The resistivity of historical karst geological samples was obtained by an electrical resistivity analyzer, the wave velocity reduction of the historical karst geological samples was simultaneously collected by a micro-motion sensor, and the linear solubility of the historical karst geological samples was simultaneously collected by a borehole core scanner.

[0062] The resistivity, wave velocity reduction, and linear melting rate are standardized and subjected to wavelet denoising to obtain training feature data.

[0063] The device of this invention, through processor 1001 calling the multi-source fusion prediction program for karst geological hazards stored in memory 1005, also performs the following operations:

[0064] Z-Score normalization was performed on the resistivity, the wave velocity reduction, and the linear melting rate to obtain normalized data;

[0065] The standardized data is decomposed into three levels using the db4 wavelet basis. The detailed components containing karst features in the standardized data are then thresholded and enhanced. The enhanced components are then reconstructed to obtain training feature data.

[0066] The device of this invention, through processor 1001 calling the multi-source fusion prediction program for karst geological hazards stored in memory 1005, also performs the following operations:

[0067] Based on the training feature data, a Transformer model is constructed that includes a multimodal embedding layer, address semantic location encoding, and a self-attention layer;

[0068] The proportion of disaster samples in the training feature data is dynamically adjusted according to a preset weighted loss function, and the Transformer model is trained according to the adjusted proportion of disaster samples to obtain the target model.

[0069] The device of this invention, through processor 1001 calling the multi-source fusion prediction program for karst geological hazards stored in memory 1005, also performs the following operations:

[0070] The training feature data is mapped to a 256-dimensional space through a multimodal embedding layer, and the feature data of each detection point is combined into a three-dimensional vector, which is then embedded to form a 256-dimensional feature vector.

[0071] The planar coordinates of each detection point are converted into position vectors, and a 256-dimensional position encoding vector is calculated based on the position vectors using a preset geological semantic position encoding formula.

[0072] The 256-dimensional feature vector is added to the 256-dimensional position encoding vector and then input into the self-attention layer to construct the Transformer model.

[0073] The device of this invention, through processor 1001 calling the multi-source fusion prediction program for karst geological hazards stored in memory 1005, also performs the following operations:

[0074] The planar coordinates of each detection point are converted into position vectors. Then, a 256-dimensional position encoding vector is obtained by calculating the position vectors using a preset geological semantic position encoding formula:

[0075]

[0076]

[0077] in, The x-coordinate of the measuring point in the plane coordinate system. This represents the y-coordinate of the measuring point in the plane coordinate system. Dimension index for 256-dimensional features The frequency factor corresponding to the x-coordinate. The frequency factor corresponding to the y-coordinate. It is a 256-dimensional position encoding vector with an even number of dimensions. It is a 256-dimensional positional encoding vector with an odd number of dimensions.

[0078] The device of this invention, through processor 1001 calling the multi-source fusion prediction program for karst geological hazards stored in memory 1005, also performs the following operations:

[0079] The real-time geophysical data of the target area is preprocessed to obtain structured input data, and the structured input data is then input into the target model.

[0080] Obtain the karst development risk probability output by the target model, match the karst development risk probability with a preset probability threshold range, and obtain the karst development risk level corresponding to the karst development risk probability.

[0081] This embodiment, through the above-described scheme, obtains the resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples. The resistivity, wave velocity reduction, and linear dissolution rate are preprocessed to obtain training feature data. A Transformer model is constructed and trained based on this training feature data to obtain the target model. Real-time geophysical data of the target area is preprocessed and input into the target model to output the karst development risk level. This enables accurate prediction of karst risk, improves the accuracy of karst disaster detection, reduces the underreporting rate of high-risk karst, and significantly solves the problems of data uniformity, lack of spatial correlation, and sample imbalance leading to high-risk underreporting in existing technologies. It improves the accuracy of karst disaster prediction, avoids construction risks, enhances the identification ability of concealed karst, and improves the speed and efficiency of multi-source fusion prediction of karst geological disasters.

[0082] Based on the above hardware structure, an embodiment of the multi-source fusion prediction method for karst geological disasters of the present invention is proposed.

[0083] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention.

[0084] In the first embodiment, the multi-source fusion prediction method for karst geological hazards includes the following steps:

[0085] Step S10: Obtain the resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples; preprocess the resistivity, wave velocity reduction, and linear dissolution rate to obtain training feature data.

[0086] It should be noted that by collecting resistivity, wave velocity reduction, and linear dissolution rate data from historical karst geological samples, and then preprocessing the resistivity, wave velocity reduction, and linear dissolution rate, highly reliable and structured training feature data can be generated.

[0087] Step S20: Construct and train a Transformer model based on the training feature data to obtain the target model.

[0088] It should be understood that a Transformer model can be constructed based on the training feature data, and after training the Transformer model, a target model for high-precision karst risk prediction can be obtained.

[0089] Step S30: After preprocessing the real-time geophysical data of the target area, input it into the target model and output the karst development risk level.

[0090] Understandably, real-time geophysical data of the target area is preprocessed and then input into the target model to output the karst development risk level.

[0091] This embodiment, through the above-described scheme, obtains the resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples. The resistivity, wave velocity reduction, and linear dissolution rate are preprocessed to obtain training feature data. A Transformer model is constructed and trained based on this training feature data to obtain the target model. Real-time geophysical data of the target area is preprocessed and input into the target model to output the karst development risk level. This enables accurate prediction of karst risk, improves the accuracy of karst disaster detection, reduces the underreporting rate of high-risk karst, and significantly solves the problems of data uniformity, lack of spatial correlation, and sample imbalance leading to high-risk underreporting in existing technologies. It improves the accuracy of karst disaster prediction, avoids construction risks, enhances the identification ability of concealed karst, and improves the speed and efficiency of multi-source fusion prediction of karst geological disasters.

[0092] Furthermore, Figure 3 This is a flowchart illustrating the second embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention, as shown below. Figure 3 As shown, based on the first embodiment, a second embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention is proposed. In this embodiment, step S10 specifically includes the following steps:

[0093] Step S11: Obtain the resistivity of the historical karst geological sample using an electrical resistivity analyzer, simultaneously collect the wave velocity reduction of the historical karst geological sample using a micro-motion sensor, and simultaneously collect the linear solubility of the historical karst geological sample using a borehole core scanner.

[0094] It should be noted that the resistivity of historical karst geological samples is obtained using a high-density resistivity meter, the wave velocity reduction is collected simultaneously using a micro-motion detector, and the linear karst rate is collected simultaneously using a borehole core scanning system, thereby achieving accurate joint acquisition of data from three sources.

[0095] It is understandable that resistivity ρ, wave velocity reduction ΔVp, and linear dissolution rate λ are collected simultaneously in historical karst geological sample areas (such as karst development areas with marked risk levels) using a high-density electrical resistivity meter (electrode spacing 5m, grid survey line layout), a micro-motion detector (observation points are arranged in a triangular array with a side length of 20m), and a borehole core scanning system (calculating the proportion of dissolution length for every 5cm sample).

[0096] In practice, the multi-source data joint acquisition process is as follows:

[0097] Resistivity ρ was obtained using a high-density electrical resistivity meter with an electrode spacing of 5m. A grid survey line was laid out to obtain the resistivity ρ of the target area. The sampling frequency was 1kHz to suppress electromagnetic interference from the strata.

[0098] The wave velocity reduction ΔVp was obtained by a micrometer. The observation points were arranged in a triangular array with a spacing of 20m. The wave velocity reduction ΔVp of each point relative to the stable bedrock was calculated (formula: ΔVp=(Vp0-Vp1) / Vp0×100%, where Vp0 is the wave velocity of the intact bedrock and Vp1 is the measured wave velocity).

[0099] Core sampling was performed according to relevant requirements, with the drilling spacing and depth meeting the specifications. The linear dissolution rate λ was calculated for every 5 cm of the core sample, where λ = Σ(dissolution length) / sample length × 100%.

[0100] Step S12: Standardize and perform wavelet denoising on the resistivity, wave velocity reduction, and linear melting rate to obtain training feature data.

[0101] It should be understood that standardizing the resistivity, wave velocity reduction, and linear melting rate can eliminate dimensional differences. After wavelet denoising, high-reliability structured training feature data with improved signal-to-noise ratio can be generated.

[0102] Furthermore, step S12 specifically includes the following steps:

[0103] Z-Score normalization was performed on the resistivity, the wave velocity reduction, and the linear melting rate to obtain normalized data;

[0104] The standardized data is decomposed into three levels using the db4 wavelet basis. The detailed components containing karst features in the standardized data are then thresholded and enhanced. The enhanced components are then reconstructed to obtain training feature data.

[0105] It should be noted that the resistivity, wave velocity reduction, and linear dissolution rate were respectively subjected to Z-Score standardization (eliminating dimensional differences and normalizing to a mean of 0 and a standard deviation of 1). Then, a 3-level decomposition was performed using the db4 wavelet basis. After threshold enhancement processing was applied to the detail components containing karst features (such as dissolution fracture signals), the data were reconstructed. Finally, standardized training feature data with a signal-to-noise ratio increased by 2.3 times and spatial correlation enhanced was generated, providing structured input for the Transformer model.

[0106] In the specific implementation, the data preprocessing process is as follows:

[0107] Z-score normalization is applied to resistivity ρ, wave velocity reduction ΔVp, and linear solubility λ. The standard formula is:

[0108]

[0109] in, The Z-Score is the standardized data (mean 0, standard deviation 1). The original data values, The mean of the dataset. The standard deviation of the dataset;

[0110] A three-level decomposition was performed using the db4 wavelet basis to obtain three detail components and one approximate component. The feature information related to karst development in each component was analyzed. For the detail components containing obvious karst features, threshold enhancement was used to highlight the karst anomaly information. The processed components were then reconstructed to obtain data with enhanced karst features.

[0111] This embodiment uses the above-described scheme to obtain the resistivity of historical karst geological samples using an electrical resistivity analyzer, simultaneously collect the wave velocity reduction amplitude of the historical karst geological samples using a micro-motion sensor, and simultaneously collect the linear dissolution rate of the historical karst geological samples using a borehole core scanner. The resistivity, wave velocity reduction amplitude, and linear dissolution rate are standardized and subjected to wavelet denoising to obtain training feature data. This generates highly reliable and structured training feature data, laying a data foundation for the subsequent accurate modeling of karst spatial topology by the Transformer model, and effectively solving the prediction blind spot problem caused by traditional single data sources.

[0112] Furthermore, Figure 4 This is a flowchart illustrating the third embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention. Figure 4 As shown, based on the first embodiment, a third embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention is proposed. In this embodiment, step S20 specifically includes the following steps:

[0113] Step S21: Construct a Transformer model containing a multimodal embedding layer, address semantic location encoding, and self-attention layer based on the training feature data.

[0114] It should be noted that a Transformer model containing a multimodal embedding layer, a geological semantic location encoding layer, and a self-attention layer can be constructed based on the training feature data.

[0115] Furthermore, step S21 specifically includes the following steps:

[0116] The training feature data is mapped to a 256-dimensional space through a multimodal embedding layer, and the feature data of each detection point is combined into a three-dimensional vector, which is then embedded to form a 256-dimensional feature vector.

[0117] The planar coordinates of each detection point are converted into position vectors, and a 256-dimensional position encoding vector is calculated based on the position vectors using a preset geological semantic position encoding formula.

[0118] The 256-dimensional feature vector is added to the 256-dimensional position encoding vector and then input into the self-attention layer to construct the Transformer model.

[0119] It is understandable that by using a multimodal embedding layer, resistivity, wave velocity reduction, and linear karst rate data are mapped to a unified 256-dimensional feature space; geological semantic location encoding is introduced to process the coordinates of the measuring points, preserving the spatial topological relationship of the karst structure; an 8-head self-attention layer can be designed to split the input features into 8 sub-vectors and calculate the attention weights in parallel, effectively capturing long-range spatial dependencies, and finally forming a Transformer model architecture that can accurately fuse multi-source data.

[0120] In its specific implementation, the Transformer model construction process is as follows:

[0121] For data of three different modes—resistivity ρ, wave velocity reduction ΔVp, and linear solubility λ—an embedding layer is designed to convert the data of each mode into an embedding vector of the same dimension. This is achieved through linear transformation and activation function, so that the data of different modes are in the same feature space.

[0122] By combining the coordinate information of data collection points and the geological structural features of the region, a location encoding function is designed to enable the model to perceive the spatial location semantics of the data. The location encoding vector is added to the embedding vector and used as the input sequence for the model.

[0123] An 8-head self-attention mechanism is constructed, which divides the feature vector of the input sequence into 8 different subspaces for parallel attention computation.

[0124] The attention weight of each head can be calculated using the Query, Key, and Value matrices. Finally, the outputs of the eight heads are concatenated and a linear transformation is performed to obtain the output of the self-attention layer.

[0125] In the specific implementation, see Figure 5 , Figure 5 This is a schematic diagram of the multi-source data fusion logic in the multi-source fusion prediction method for karst geological hazards of the present invention, as shown below. Figure 5 As shown, after the data is input, data of the same scale can be preprocessed at the data layer, and features of the same scale can be embedded into the same space through multimodal embedding at the feature layer. At the model layer, cross-source association can be achieved through self-attention mechanism, thereby outputting the multi-source data fusion result.

[0126] It should be noted that, see Figure 6 , Figure 6This is a schematic diagram of the Transformer model architecture in the multi-source fusion prediction method for karst geological hazards of the present invention, as shown below. Figure 6 As shown, the input layer embeds location labels through multi-source data sequences, encodes them through an encoding component composed of encoders, and decodes them through a decoding component composed of decoders, thereby outputting predicted values ​​at the output layer.

[0127] Furthermore, the step converts the planar coordinates of each detection point into a position vector, and calculates a 256-dimensional position encoding vector based on the position vector using a preset geological semantic position encoding formula, including:

[0128] The planar coordinates of each detection point are converted into position vectors. Then, a 256-dimensional position encoding vector is obtained by calculating the position vectors using a preset geological semantic position encoding formula:

[0129]

[0130]

[0131] in, The x-coordinate of the measuring point in the plane coordinate system. This represents the y-coordinate of the measuring point in the plane coordinate system. Dimension index for 256-dimensional features The frequency factor corresponding to the x-coordinate. The frequency factor corresponding to the y-coordinate. It is a 256-dimensional position encoding vector with an even number of dimensions. It is a 256-dimensional positional encoding vector with an odd number of dimensions.

[0132] In the specific implementation, the three types of data are mapped to a 256-dimensional feature space, with weight matrices Wρ∈R256×1, WΔVp∈R256×1, and Wλ∈R256×1. Dimensionality is unified through linear transformation, as shown in the formula:

[0133]

[0134] in, This is a 256-dimensional feature vector of resistivity after processing by a multimodal embedding layer. This is a 256×1 weight matrix corresponding to resistivity. The input is the standardized resistivity value. This is a 256-dimensional eigenvector representing the wave velocity reduction amplitude. This is the 256×1 weight matrix corresponding to the wave velocity reduction. The input is the standardized wave velocity reduction value. This is a 256-dimensional eigenvector representing the linear melting rate. This is the 256×1 weight matrix corresponding to the linear dissolution rate. The input is the standardized linear solubility value;

[0135] Geological semantic location encoding converts spatial coordinates (x, y) into location vectors PE; where (x, y) are the latitude and longitude or projected coordinates of the measurement point, and i is the dimension index of the 256-dimensional feature. This encoding, by coupling geographic coordinates with hyperbolic sine waves, preserves the relative distances and topological relationships between measurement points in high-dimensional space, solving the problem of karst spatial correlation modeling.

[0136] The multi-head self-attention layer uses 8 heads to calculate attention weights in parallel, as shown in the formula:

[0137]

[0138] Where Q=K=V=E+PE, dk=32, multi-scale feature association is captured in parallel by 8 heads, Q is the query vector, K is the key vector, V is the value vector, softmax is the normalization function, dK is the dimension of the key vector, and T represents the matrix transpose.

[0139] Step S22: Dynamically adjust the proportion of disaster samples in the training feature data according to the preset weighted loss function, and train the Transformer model according to the adjusted proportion of disaster samples to obtain the target model.

[0140] It is understandable that the proportion of disaster samples in the training feature data is dynamically adjusted according to a pre-set weighted loss function, and the Transformer model is trained based on the adjusted proportion of disaster samples to obtain the target model.

[0141] It should be understood that a dynamic weighted loss function is used during the training process, which is automatically adjusted according to the proportion of disaster samples. This effectively solves the problem of sample imbalance, significantly reduces the underreporting rate of high-risk karst, and finally outputs an optimized target model for accurate prediction of karst risk.

[0142] In practice, the model training process is as follows:

[0143] The dataset is partitioned by dividing the preprocessed multi-source data into training, validation, and test sets according to a certain ratio. The training set is used to learn the model parameters, the validation set is used to monitor overfitting during model training and adjust hyperparameters, and the test set is used to evaluate the final performance of the model.

[0144] Training parameters are set using the Adam optimizer, with an initial learning rate of 0.001, dynamically adjusted based on the loss changes on the validation set (e.g., using a learning rate decay strategy); batch size is set to 32-128 (adjusted according to hardware performance); the number of training epochs is determined based on model convergence, generally no less than 100 epochs.

[0145] Weighted loss function application: The training process uses a weighted loss function.

[0146] To address the imbalanced sample problem, the weighted loss function is L = 0.5MSE + 0.5FL, where MSE is the mean squared error and FL is the focal loss, which measures the overall error between the model's predicted values ​​and the true values.

[0147] Focus loss:

[0148]

[0149] in, The predicted probability of disaster samples output by the model is obtained by... This approach increases the weight of low-probability disaster samples, balances the bias of MSE towards the majority class (non-disaster samples), and reduces the false negative rate.

[0150] This embodiment, through the above-described scheme, constructs a Transformer model based on the training feature data, which includes a multimodal embedding layer, address semantic location encoding, and a self-attention layer; dynamically adjusts the proportion of disaster samples in the training feature data according to a preset weighted loss function; trains the Transformer model based on the adjusted proportion of disaster samples to obtain the target model. This effectively solves the problem of missed reports caused by the scarcity of disaster samples in karst data, reduces the missed report rate of high-risk karst, and significantly improves the accuracy and reliability of karst risk prediction.

[0151] Furthermore, Figure 7 This is a flowchart illustrating the fourth embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention. Figure 7 As shown, based on the first embodiment, a fourth embodiment of the multi-source fusion prediction method for karst geological hazards of the present invention is proposed. In this embodiment, step S30 specifically includes the following steps:

[0152] Step S31: Preprocess the real-time geophysical data of the target area to obtain structured input data, and input the structured input data into the target model.

[0153] It should be noted that the real-time geophysical data of the target area (including resistivity ρ obtained by high-density resistivity meter, wave velocity reduction ΔVp obtained by micro-motion detector, and linear karst rate λ calculated by borehole core scanning system) undergoes the same preprocessing as in the training phase: first, Z-Score standardization is performed (eliminating dimensional differences and normalizing to mean 0 and standard deviation 1), then 3-level decomposition is performed using db4 wavelet basis, and the detailed components containing karst features are reconstructed after threshold enhancement, which can finally generate structured input data with high signal-to-noise ratio (increased to 2.3 times).

[0154] Step S32: Obtain the karst development risk probability output by the target model, match the karst development risk probability with a preset probability threshold range, and obtain the karst development risk level corresponding to the karst development risk probability.

[0155] It can be understood that this data is input into the trained Transformer target model, and the model outputs the karst development risk probability p based on the geological semantic position encoding and the dynamic weighted loss function. Matching the karst development risk probability with the preset probability threshold range means automatically dividing the risk levels according to different preset thresholds: low risk (p ≤ 0.3), medium risk (0.3 < p ≤ 0.7), and high risk (p > 0.7), so as to achieve accurate real-time early warning of karst disasters.

[0156] It should be understood that the real-time geophysical exploration data of the target area is input, and the karst development risk level is output through the trained model: the last layer of the model outputs the risk probability p between 0 and 1 through the Sigmoid function; level division: set the thresholds p1 = 0.3 and p2 = 0.7, and divide them into: low risk (p ≤ 0.3), medium risk (0.3 < p ≤ 0.7), high risk (p > 0.7). Of course, other values can also be set, and this embodiment does not limit this.

[0157] In specific implementation, taking a highway bridge project in a mountainous area in the southwest as an example, the limestone formation is developed in this area, with a total length of 3.2 km. The pile foundation is adopted, and the pile length is 40 - 66 m. It is necessary to identify the karst development situation in the bridge site area to provide a basis for determining the pile length. The precipitation in the project area is large in the rainy season, the corrosion effect is active, and historical surveys show that there are karst hidden dangers.

[0158] 1. Data collection implementation

[0159] (1) High-density electrical method detection

[0160] Equipment: WDJD-4 type electrical method instrument, configured with 64 channels of electrodes, using the Wenner device;

[0161] Survey line layout: Arrange 2 survey lines along the tunnel axis direction, and arrange 1 transverse survey line every 500 m perpendicular to the axis direction, forming a 5 m × 50 m grid;

[0162] Collection parameters: Supply voltage 120 V, sampling interval 0.5 s, suppress the electromagnetic interference of the formation, and a total of 1800 groups of resistivity data are obtained, with a range of 50 - 6000 Ω·m.

[0163] (2) Microtremor detection

[0164] Equipment: SE2404 type microtremor instrument, 48 channels, sampling rate 100 Hz;

[0165] Observation point layout: 60 points are arranged in a triangular array with a side length of 20m, covering the tunnel axis and a range of 100m on both sides;

[0166] Data processing: Surface wave velocity was calculated using the spectrum analysis method and compared with the wave velocity of intact bedrock (Vp0=3800m / s) to obtain the wave velocity reduction ΔVp, which ranged from 5-45%, with 520 sets of valid data.

[0167] (3) Core sampling verification

[0168] Equipment: XY-1 drilling rig, diamond drill bit, core diameter 91mm;

[0169] Drilling layout: Drill holes are laid out along the bridge foundation location, with a depth of 50-75m;

[0170] Linear dissolution rate calculation: For each 1m unit of the core sample, the length of the dissolution fracture was measured and λ was calculated, ranging from 0.5% to 15%, with 230 sets of valid data.

[0171] 2. Data Preprocessing Implementation

[0172] (1) Z-Score standardization

[0173] Resistivity ρ: μρ = 1900 Ω·m, σρ = 1300 Ω·m, normalized range -1.46 to 2.38;

[0174] Wave velocity reduction ΔVp: μΔVp=17%, σΔVp=11%, normalized range -1.09 to 2.55; Linear dissolution rate λ: μλ=4.5%, σλ=3.2%, normalized range -1.25 to 3.28.

[0175] (2) db4 wavelet 3-level decomposition

[0176] Using the db4 wavelet from the PyWavelets library, the standardized data was decomposed into three levels, preserving high-frequency components (detail coefficients d1-d3). The signal-to-noise ratio of karst features in the reconstructed data increased from 1.3 to 2.3, and the signal intensity of dissolution fissures was enhanced by 1.9 times.

[0177] 3. Model Building and Training Implementation

[0178] (1) Multimodal embedding layer

[0179] Wρ, WΔVp, and Wλ are initialized using Xavier, mapping the three types of data to a 256-dimensional space; the (ρ, ΔVp, λ) of each probe point are combined into a 3-dimensional vector, which is then embedded to form a 256-dimensional feature vector.

[0180] (2) Geological semantic location coding

[0181] Convert the plane coordinates (x, y) of the detection points into position vectors, where x ∈ [0, 3200] and y ∈ [-100, 100]. Calculate the 256-dimensional position encoding through a formula to make the cosine similarity of the position vectors of points within 30m > 0.8.

[0182] (3) Multi-Head Self-Attention Layer

[0183] The network structure contains 6 layers of Transformer encoders, with 8 heads of self-attention in each layer and a feed-forward neural network dimension of 1024; the Adam optimizer is used, with a learning rate of 0.001, a batch size of 32, and 500 iterations, and the training duration is about 4.5 hours.

[0184] (4) Loss Function Training

[0185] The weighting coefficient α = 0.5 and the focal loss γ = 2; for the training set division, 80% of the data is used for training and 20% of the data is used for validation; stop training when the validation set MSE < 0.115 and the F1 score > 0.90.

[0186] 4. Risk Prediction and Level Classification Implementation

[0187] Input the preprocessed 256-dimensional feature vector. After 6 layers of Transformer encoding, output the risk probability p through a fully connected layer and the Sigmoid function; set p ≤ 0.3 (low risk), 0.3 < p ≤ 0.7 (medium risk), p > 0.7 (high risk). A total of 15 high-risk areas are identified, accounting for 21%.

[0188] 5. Effect Verification

[0189] See Figure 8 , Figure 8 which is the comparison chart of the implementation effects of the present invention, the BP neural network, and the LSTM. As Figure 8 shown, it can be seen that the present invention is superior to the Backpropagation Neural Network (BP) and the Long Short-Term Memory (LSTM) in terms of prediction accuracy and high-risk identification rate.

[0190] (1) Comparison with Traditional Methods

[0191]

[0192] (2) Engineering Verification Results

[0193] Advanced drilling was conducted on 6 out of 15 high-risk areas to verify the presence of karst caves (0.8-4.1m high) in 5 areas and a dense zone of dissolution fissures in 1 area, with a verification accuracy of 83.3%. In contrast, the traditional BP neural network missed 2 karst caves (1.5-2.8m high). This invention improves the accuracy of karst disaster prediction and avoids construction risks.

[0194] This embodiment, through the above-described scheme, obtains structured input data by preprocessing real-time geophysical data of the target area, and inputs the structured input data into the target model; obtains the karst development risk probability output by the target model, matches the karst development risk probability with a preset probability threshold range, and obtains the karst development risk level corresponding to the karst development risk probability; it can achieve accurate prediction of karst risk, improve the accuracy of karst disaster detection, reduce the false negative rate of high-risk karst, significantly solve the problems of high-risk false negatives caused by data uniformity, lack of spatial correlation, and sample imbalance in the prior art, improve the accuracy of karst disaster prediction, avoid construction risks, enhance the ability to identify hidden karst, and improve the speed and efficiency of multi-source fusion prediction of karst geological disasters.

[0195] Accordingly, the present invention further provides a multi-source fusion prediction device for karst geological disasters.

[0196] Reference Figure 9 , Figure 9 This is a functional block diagram of the first embodiment of the karst geological disaster multi-source fusion prediction device of the present invention.

[0197] In the first embodiment of the karst geological hazard multi-source fusion prediction device of the present invention, the karst geological hazard multi-source fusion prediction device includes:

[0198] The data preprocessing module 10 is used to acquire the resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples, and to preprocess the resistivity, wave velocity reduction, and linear dissolution rate to obtain training feature data.

[0199] The model training module 20 is used to construct and train a Transformer model based on the training feature data to obtain the target model.

[0200] The risk analysis module 30 is used to preprocess real-time geophysical data of the target area and input it into the target model, and output the karst development risk level.

[0201] The data preprocessing module 10 is also used to obtain the resistivity of historical karst geological samples through an electrical resistivity meter, synchronously collect the wave velocity reduction amplitude of the historical karst geological samples through a micro-motion sensor, and synchronously collect the linear dissolution rate of the historical karst geological samples through a borehole core scanner; and to standardize and perform wavelet noise reduction on the resistivity, the wave velocity reduction amplitude and the linear dissolution rate to obtain training feature data.

[0202] The data preprocessing module 10 is further used to perform Z-Score standardization on the resistivity, the wave velocity reduction, and the linear karst rate to obtain standardized data; to perform three-level decomposition on the standardized data using the db4 wavelet basis; to perform threshold enhancement on the detail components containing karst features in the standardized data; and to reconstruct the enhanced components to obtain training feature data.

[0203] The model training module 20 is also used to construct a Transformer model containing a multimodal embedding layer, address semantic location encoding and self-attention layer based on the training feature data; dynamically adjust the proportion of disaster samples in the training feature data according to a preset weighted loss function; train the Transformer model according to the adjusted proportion of disaster samples to obtain the target model.

[0204] The model training module 20 is further configured to map the training feature data to a 256-dimensional space through a multimodal embedding layer, combine the feature data of each detection point into a three-dimensional vector, and form a 256-dimensional feature vector after embedding; convert the planar coordinates of each detection point into a position vector, calculate the 256-dimensional position encoding vector according to the position vector using a preset geological semantic position encoding formula; and add the 256-dimensional feature vector and the 256-dimensional position encoding vector and input them into the self-attention layer to construct a Transformer model.

[0205] The model training module 20 is also used to convert the planar coordinates of each detection point into a position vector, and to calculate a 256-dimensional position encoding vector based on the position vector using a preset geological semantic position encoding formula:

[0206]

[0207]

[0208] in, The x-coordinate of the measuring point in the plane coordinate system. This represents the y-coordinate of the measuring point in the plane coordinate system. Dimension index for 256-dimensional features The frequency factor corresponding to the x-coordinate. The frequency factor corresponding to the y-coordinate. It is a 256-dimensional position encoding vector with an even number of dimensions. It is a 256-dimensional positional encoding vector with an odd number of dimensions.

[0209] The risk analysis module 30 is further configured to preprocess real-time geophysical data of the target area to obtain structured input data, input the structured input data into the target model, obtain the karst development risk probability output by the target model, match the karst development risk probability with a preset probability threshold range, and obtain the karst development risk level corresponding to the karst development risk probability.

[0210] The steps for implementing each functional module of the karst geological hazard multi-source fusion prediction device can be referred to in the various embodiments of the karst geological hazard multi-source fusion prediction method of the present invention, and will not be repeated here.

[0211] Furthermore, this embodiment of the invention also proposes a storage medium storing a multi-source fusion prediction program for karst geological hazards. When the multi-source fusion prediction program for karst geological hazards is executed by a processor, it performs the following operations:

[0212] Resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples are obtained. The resistivity, wave velocity reduction, and linear dissolution rate are preprocessed to obtain training feature data.

[0213] Based on the training feature data, a Transformer model is constructed and trained to obtain the target model;

[0214] After preprocessing the real-time geophysical data of the target area, the data is input into the target model, and the karst development risk level is output.

[0215] Furthermore, when the karst geological hazard multi-source fusion prediction program is executed by the processor, it also performs the following operations:

[0216] The resistivity of historical karst geological samples was obtained by an electrical resistivity analyzer, the wave velocity reduction of the historical karst geological samples was simultaneously collected by a micro-motion sensor, and the linear solubility of the historical karst geological samples was simultaneously collected by a borehole core scanner.

[0217] The resistivity, wave velocity reduction, and linear melting rate are standardized and subjected to wavelet denoising to obtain training feature data.

[0218] Furthermore, when the karst geological hazard multi-source fusion prediction program is executed by the processor, it also performs the following operations:

[0219] Z-Score normalization was performed on the resistivity, the wave velocity reduction, and the linear melting rate to obtain normalized data;

[0220] The standardized data is decomposed into three levels using the db4 wavelet basis. The detailed components containing karst features in the standardized data are then thresholded and enhanced. The enhanced components are then reconstructed to obtain training feature data.

[0221] Furthermore, when the karst geological hazard multi-source fusion prediction program is executed by the processor, it also performs the following operations:

[0222] Based on the training feature data, a Transformer model is constructed that includes a multimodal embedding layer, address semantic location encoding, and a self-attention layer;

[0223] The proportion of disaster samples in the training feature data is dynamically adjusted according to a preset weighted loss function, and the Transformer model is trained according to the adjusted proportion of disaster samples to obtain the target model.

[0224] Furthermore, when the karst geological hazard multi-source fusion prediction program is executed by the processor, it also performs the following operations:

[0225] The training feature data is mapped to a 256-dimensional space through a multimodal embedding layer, and the feature data of each detection point is combined into a three-dimensional vector, which is then embedded to form a 256-dimensional feature vector.

[0226] The planar coordinates of each detection point are converted into position vectors, and a 256-dimensional position encoding vector is calculated based on the position vectors using a preset geological semantic position encoding formula.

[0227] The 256-dimensional feature vector is added to the 256-dimensional position encoding vector and then input into the self-attention layer to construct the Transformer model.

[0228] Furthermore, when the karst geological hazard multi-source fusion prediction program is executed by the processor, it also performs the following operations:

[0229] The planar coordinates of each detection point are converted into position vectors. Then, a 256-dimensional position encoding vector is obtained by calculating the position vectors using a preset geological semantic position encoding formula:

[0230]

[0231]

[0232] in, The x-coordinate of the measuring point in the plane coordinate system. This represents the y-coordinate of the measuring point in the plane coordinate system. Dimension index for 256-dimensional features The frequency factor corresponding to the x-coordinate. The frequency factor corresponding to the y-coordinate. It is a 256-dimensional position encoding vector with an even number of dimensions. It is a 256-dimensional positional encoding vector with an odd number of dimensions.

[0233] Furthermore, when the karst geological hazard multi-source fusion prediction program is executed by the processor, it also performs the following operations:

[0234] The real-time geophysical data of the target area is preprocessed to obtain structured input data, and the structured input data is then input into the target model.

[0235] Obtain the karst development risk probability output by the target model, match the karst development risk probability with a preset probability threshold range, and obtain the karst development risk level corresponding to the karst development risk probability.

[0236] Those skilled in the art will understand that all or part of the steps in the methods described above can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium is a computer-readable storage medium, including: USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and other media that can store program code.

[0237] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0238] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0239] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A multi-source fusion prediction method for karst geological hazards, characterized in that, The multi-source fusion prediction method for karst geological hazards includes: Resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples are obtained. The resistivity, wave velocity reduction, and linear dissolution rate are preprocessed to obtain training feature data. Based on the training feature data, a Transformer model is constructed and trained to obtain the target model; After preprocessing, the real-time geophysical data of the target area is input into the target model, and the karst development risk level is output. The step of constructing and training a Transformer model based on the training feature data to obtain the target model includes: Based on the training feature data, a Transformer model is constructed that includes a multimodal embedding layer, address semantic location encoding, and a self-attention layer; The proportion of disaster samples in the training feature data is dynamically adjusted according to a preset weighted loss function, and the Transformer model is trained according to the adjusted proportion of disaster samples to obtain the target model. The construction of the Transformer model based on the training feature data, comprising a multimodal embedding layer, address semantic location encoding, and a self-attention layer, includes: The training feature data is mapped to a 256-dimensional space through a multimodal embedding layer, and the feature data of each detection point is combined into a three-dimensional vector, which is then embedded to form a 256-dimensional feature vector. The planar coordinates of each detection point are converted into position vectors, and a 256-dimensional position encoding vector is calculated based on the position vectors using a preset geological semantic position encoding formula. The 256-dimensional feature vector is added to the 256-dimensional position encoding vector and then input into the self-attention layer to construct the Transformer model; The step of converting the planar coordinates of each detection point into a position vector, and calculating a 256-dimensional position encoding vector based on the position vector using a preset geological semantic position encoding formula, includes: The planar coordinates of each detection point are converted into position vectors. Then, a 256-dimensional position encoding vector is obtained by calculating the position vectors using a preset geological semantic position encoding formula: in, The x-coordinate of the measuring point in the plane coordinate system. This represents the y-coordinate of the measuring point in the plane coordinate system. Dimension index for 256-dimensional features The frequency factor corresponding to the x-coordinate. The frequency factor corresponding to the y-coordinate. It is a 256-dimensional position encoding vector with an even number of dimensions. It is a 256-dimensional positional encoding vector with an odd number of dimensions.

2. The multi-source fusion prediction method for karst geological hazards as described in claim 1, characterized in that, The process involves acquiring resistivity, wave velocity reduction, and linear dissolution rate from historical karst geological samples, and preprocessing these resistivity, wave velocity reduction, and linear dissolution rate to obtain training feature data, including: The resistivity of historical karst geological samples was obtained by an electrical resistivity analyzer, the wave velocity reduction of the historical karst geological samples was simultaneously collected by a micro-motion sensor, and the linear solubility of the historical karst geological samples was simultaneously collected by a borehole core scanner. The resistivity, wave velocity reduction, and linear melting rate are standardized and subjected to wavelet denoising to obtain training feature data.

3. The multi-source fusion prediction method for karst geological hazards as described in claim 2, characterized in that, The process of standardizing and performing wavelet denoising on the resistivity, wave velocity reduction, and linear melting rate to obtain training feature data includes: Z-Score normalization was performed on the resistivity, the wave velocity reduction, and the linear melting rate to obtain normalized data; The standardized data is decomposed into three levels using the db4 wavelet basis. The detailed components containing karst features in the standardized data are then thresholded and enhanced. The enhanced components are then reconstructed to obtain training feature data.

4. The multi-source fusion prediction method for karst geological hazards as described in claim 1, characterized in that, The process of preprocessing real-time geophysical data of the target area and inputting it into the target model to output the karst development risk level includes: The real-time geophysical data of the target area is preprocessed to obtain structured input data, and the structured input data is then input into the target model. Obtain the karst development risk probability output by the target model, match the karst development risk probability with a preset probability threshold range, and obtain the karst development risk level corresponding to the karst development risk probability.

5. A multi-source fusion prediction device for karst geological hazards, characterized in that, The multi-source fusion prediction device for karst geological hazards includes: The data preprocessing module is used to obtain the resistivity, wave velocity reduction, and linear dissolution rate of historical karst geological samples, and to preprocess the resistivity, wave velocity reduction, and linear dissolution rate to obtain training feature data. The model training module is used to construct and train a Transformer model based on the training feature data to obtain the target model; The risk analysis module is used to preprocess real-time geophysical data of the target area and input it into the target model, and output the karst development risk level. The model training module is also used to construct a Transformer model containing a multimodal embedding layer, address semantic location encoding, and self-attention layer based on the training feature data; dynamically adjust the proportion of disaster samples in the training feature data according to a preset weighted loss function; train the Transformer model according to the adjusted proportion of disaster samples to obtain the target model; The model training module is further configured to map the training feature data to a 256-dimensional space through a multimodal embedding layer, combine the feature data of each detection point into a three-dimensional vector, and form a 256-dimensional feature vector after embedding; convert the planar coordinates of each detection point into a position vector, calculate the 256-dimensional position encoding vector according to the position vector using a preset geological semantic position encoding formula; and add the 256-dimensional feature vector and the 256-dimensional position encoding vector and input them into the self-attention layer to construct the Transformer model. The model training module is also used to convert the planar coordinates of each detection point into a position vector, and to calculate a 256-dimensional position encoding vector based on the position vector using a preset geological semantic position encoding formula: in, The x-coordinate of the measuring point in the plane coordinate system. This represents the y-coordinate of the measuring point in the plane coordinate system. Dimension index for 256-dimensional features The frequency factor corresponding to the x-coordinate. The frequency factor corresponding to the y-coordinate. It is a 256-dimensional position encoding vector with an even number of dimensions. It is a 256-dimensional positional encoding vector with an odd number of dimensions.

6. A multi-source fusion prediction device for karst geological hazards, characterized in that, The karst geological hazard multi-source fusion prediction device includes: a memory, a processor, and a karst geological hazard multi-source fusion prediction program stored in the memory and executable on the processor. The karst geological hazard multi-source fusion prediction program is configured to implement the steps of the karst geological hazard multi-source fusion prediction method as described in any one of claims 1 to 4.

7. A storage medium, characterized in that, The storage medium stores a multi-source fusion prediction program for karst geological hazards, which, when executed by a processor, implements the steps of the multi-source fusion prediction method for karst geological hazards as described in any one of claims 1 to 4.

Images