An Artificial Intelligence-Based Diagnostic Method and System for Dam Hazards Based on Three-Dimensional High-Density Electrical Resonance Method

Abstract

This invention discloses an artificial intelligence-based method and system for diagnosing potential hazards in dams based on three-dimensional high-density electrical resistivity tomography (EDT), belonging to the field of water conservancy engineering safety monitoring technology. The method includes: deploying a three-dimensional electrode array and environmental sensors along the dam, and simultaneously collecting resistivity, polarizability, and natural potential data through intelligent switching modes; processing the collected data and fusing it with environmental parameters to generate a multi-parameter data cube of the dam; inputting this data cube into a U-shaped network model trained with a physical constraint loss function to obtain a high-resolution three-dimensional resistivity distribution model; automatically extracting features of potential hazards using a three-dimensional convolutional neural network, combining environmental parameters weighted by an attention mechanism, and achieving intelligent identification and risk classification of potential hazard types through an integrated classifier, and generating a diagnostic report. This invention achieves full automation from data acquisition to intelligent diagnosis, significantly improving the accuracy and efficiency of hazard detection, and providing reliable technical support for the safe operation of dams.

Description

Technical Field

[0001] This invention relates to the field of water conservancy project safety monitoring technology, specifically to an artificial intelligence-based method and system for diagnosing potential hazards in dams based on three-dimensional high-density electrical resistivity tomography. Background Technology

[0002] As a crucial component of flood control systems, the structural safety of dams directly impacts the safety of people's lives and property. However, dams are chronically affected by seepage erosion, loose soil, animal burrows, and other hidden defects. Traditional manual inspections and drilling methods suffer from low efficiency, limited coverage, and difficulty in detecting deep-seated problems. In recent years, high-density electrical resistivity tomography (EDT) has been applied to dam inspection due to its non-destructive testing advantages, but existing technologies still have the following limitations: (1) Insufficient data acquisition dimensions: Traditional two-dimensional electrical resistivity tomography can only provide cross-sectional information, which is difficult to characterize the spatial distribution of hidden dangers. Conventional three-dimensional electrical resistivity tomography is limited by the electrode layout, which makes it difficult to balance detection depth and resolution. In addition, the electrode durability is poor in humid environments. (2) Low efficiency and accuracy of inversion algorithm: Traditional inversion methods based on least squares, finite element and other methods rely on initial model assumptions, have long calculation time, and are not good at distinguishing local anomalies and are easily affected by noise. (3) Lack of intelligent diagnostic capabilities: Existing systems rely heavily on manual experience to interpret resistivity images, lack fusion analysis of multi-source electrical parameters, and have insufficient automation and dynamism in risk assessment, making it difficult to provide real-time decision support for engineering disposal.

[0003] Furthermore, while artificial intelligence technology has made progress in fields such as medical imaging and geological exploration, it still faces challenges in the diagnosis of hidden dangers in dams: first, the complex nonlinear relationship between electrical resistivity data and geological parameters is difficult to characterize using shallow models; second, the generalization ability of models is insufficient under small sample conditions; and third, the lack of physical constraints leads to distortion of inversion results. Therefore, there is an urgent need for a system that integrates high-precision three-dimensional detection, intelligent inversion, and dynamic diagnosis to overcome the technical bottleneck in identifying hidden dam defects. Summary of the Invention

[0004] To address the problems of low detection efficiency and insufficient accuracy in existing two-dimensional high-density electrical resistivity tomography (EDT) methods, this invention proposes an artificial intelligence-based method and system for diagnosing potential hazards in dams based on three-dimensional high-density EDT.

[0005] The first aspect is an artificial intelligence-based method for diagnosing potential hazards in dams based on three-dimensional high-density electrical resistivity tomography, which includes the following steps: Step S1: Deploy a three-dimensional multi-electrode array along the dam body axis and key areas, and simultaneously deploy environmental sensors. By intelligently switching the electrode arrangement mode, synchronously collect resistivity, polarizability and natural potential data to obtain raw electrical resistivity data. Step S2: Perform terrain correction, noise filtering and bad pixel removal on the original electrical resistivity data to obtain the processed original electrical resistivity data. Then, use a spatiotemporal interpolation algorithm to fuse the processed original electrical resistivity data and environmental parameters collected by environmental sensors into a unified three-dimensional grid coordinate system to generate a multi-parameter data cube of the dam. Step S3: Construct a U-shaped network model by inputting the multi-parameter data cube of the dam into the U-shaped network model and outputting a high-resolution three-dimensional resistivity distribution model; wherein, the U-shaped network model is obtained by training using a total loss function that includes physical constraint loss; Step S4: Based on the high-resolution three-dimensional resistivity distribution model, a three-dimensional convolutional neural network is used to automatically extract the spatial morphological features, connectivity features, and electrical anomaly features of the hazard body; the spatial morphological features, connectivity features, and electrical anomaly features are fused with environmental parameters weighted by an attention mechanism, and an ensemble classifier is used to achieve intelligent identification and risk rating of the hazard body type, and finally generate a hazard diagnosis report.

[0006] Furthermore, in step S1, a three-dimensional multi-electrode array is deployed along the dam body axis and key areas, while environmental sensors are also deployed. By intelligently switching the electrode arrangement mode, resistivity, polarizability, and natural potential data are collected synchronously to obtain raw electrical resistivity data. The specific steps are as follows: Step S11: Install stainless steel silver-plated electrodes in a grid pattern along the axis of the dam body, with a longitudinal spacing of 5-20 meters and a transverse spacing of 2-5 meters, to form a three-dimensional observation network; Step S12: Densify the deployment of stainless steel silver-plated electrodes in key areas and dynamically adjust the spacing of the stainless steel silver-plated electrodes according to the detection depth requirements. The selection of the spacing of the stainless steel silver-plated electrodes meets the following empirical criteria: the effective detection depth of the stainless steel silver-plated electrodes is 1 / 6 to 1 / 8 of the spacing of the stainless steel silver-plated electrodes, and the spacing of the stainless steel silver-plated electrodes is not greater than the minimum design target size to be detected. Step S13: After the three-dimensional multi-electrode array is deployed, environmental sensors are simultaneously deployed at key nodes and key areas of the three-dimensional multi-electrode array. The environmental sensors include temperature sensors, humidity sensors and pore water pressure sensors. Step S14: Simultaneously collect resistivity, polarizability, and natural potential data, which is achieved through intelligent switching of the arrangement mode; wherein, intelligent switching includes automatically selecting and switching between the Wenner device arrangement mode and the dipole device arrangement mode; wherein, the natural potential data is directly measured.

[0007] Further, in step S2, the original electrical resistivity data undergoes terrain correction, noise filtering, and bad pixel removal to obtain processed original electrical resistivity data. Then, a spatiotemporal interpolation algorithm is used to fuse the processed original electrical resistivity data and environmental parameters collected by environmental sensors into a unified three-dimensional grid coordinate system, generating a multi-parameter data cube for the dam. The specific steps are as follows: Step S21: Perform terrain correction on the resistivity to obtain the corrected resistivity, and use the Gauss-Seidel iterative method to eliminate the influence of terrain undulations. Step S22: The corrected resistivity is filtered. The noise filtering process adopts the wavelet threshold denoising method. The wavelet coefficients are obtained by decomposing the wavelet basis function into 5 levels and then processing the wavelet coefficients using the hard threshold function. Step S23: Based on the Laida criterion, perform defect removal on the filtered resistivity, removing outliers that exceed ±3 times the standard deviation range, to obtain the processed resistivity; Step S24: Perform the same filtering and bad pixel removal processes as in steps S22 and S23 on the polarizability and natural potential data to obtain the processed polarizability and natural potential data. Step S25: Establish a three-dimensional grid coordinate system with a resolution of 0.5m × 0.5m × 0.2m. Use the Kriging spatiotemporal interpolation algorithm to interpolate the processed resistivity, processed polarizability, processed natural potential data, and temperature, humidity, and pore water pressure data collected by environmental sensors onto the grid of the unified three-dimensional grid coordinate system to generate a multi-parameter data cube of the dam.

[0008] Further, in step S3, a U-shaped network model is constructed. The multi-parameter data cube of the dam is input into the U-shaped network model, and a high-resolution three-dimensional resistivity distribution model is output. The U-shaped network model is trained using a total loss function that includes physical constraint losses. The specific steps are as follows: Step S31: The U-shaped network model consists of an encoder and a decoder; the encoder uses 4 three-dimensional convolutional layers, i.e., the kernel size is 3×3×3 and the stride is 2 for downsampling; the decoder uses 4 transposed convolutional layers for upsampling. Step S32: The total loss function of the U-shaped network model is composed of a weighted sum of the data fitting loss and the physical constraint loss; the data fitting loss is calculated using the mean square error to determine the difference between the predicted and measured values; the physical constraint loss is calculated based on Maxwell's equations; a regularization parameter is used to balance the weights of the data fitting loss and the physical constraint loss, and the value of the regularization parameter ranges from 0.1 to 0.5. Physical constraint loss The formula is as follows: ; Where M is the total number of grid points, and ∇ is the gradient operator. Here is the conductivity value at the i-th grid point, in Siemens per meter. The potential value at the i-th grid point, in volts; Step S33: The constructed U-shaped network model is trained using the Adam optimizer with an initial learning rate of 0.001 and an exponential decay strategy. The training batch size is set to 16, and the number of training rounds is 200. After training is completed, the multi-parameter data cube of the dam is input into the trained U-shaped network model, and a high-resolution three-dimensional resistivity distribution model is output. Based on the high-resolution three-dimensional resistivity distribution model, three-dimensional visualization processing is performed to generate a three-dimensional image of the resistivity distribution inside the dam.

[0009] Furthermore, in step S4, based on a high-resolution three-dimensional resistivity distribution model, a three-dimensional convolutional neural network is used to automatically extract the spatial morphological features, connectivity features, and electrical anomaly features of the hazard. These features are then fused with environmental parameters weighted through an attention mechanism, and an ensemble classifier is used to achieve intelligent identification and risk rating of the hazard type, ultimately generating a hazard diagnosis report. The specific steps are as follows: Step S41: Based on the high-resolution three-dimensional resistivity distribution model, a three-dimensional convolutional neural network is used to automatically extract the hidden danger features. The three-dimensional convolutional neural network contains 4 convolutional layers with a kernel size of 3×3×3 and a stride of 1. The pooling layer uses max pooling with a pooling window of 2×2×2. A three-dimensional convolutional neural network extracts spatial morphological features, connectivity features, and electrical anomaly features, respectively. Spatial morphological features include the volume, surface area, and flattening of the hazard body. Connectivity features include Euler number and pore-throat ratio. Electrical anomaly features include resistivity anomaly degree and resistivity contrast. Step S42: Environmental parameters are weighted and fused using an attention mechanism. These parameters include temperature, humidity, and pore water pressure. The formula is as follows: ; in, For the first The attention weights for each environmental parameter, where exp is an exponential function. Weight vector The transpose of , where tanh is the hyperbolic tangent activation function. This is the weight matrix. For the first Feature vectors of environmental parameters For bias vectors, The number of environmental parameters; Step S43: The spatial morphological features, connectivity features, and electrical anomaly features are fused with environmental parameters weighted by the attention mechanism, and the ensemble classifier is used to output the hazard type and confidence probability. Step S44: Based on the confidence probability, the hidden danger risk is divided into three risk levels: low risk, medium risk, and high risk; at the same time, a pre-set hidden danger-disposal rule knowledge base is established, with hidden danger type and risk level as joint input keys. Step S45: Based on the combination of the hazard type and risk level, match and obtain corresponding targeted handling suggestions from the hazard-handling rule knowledge base; automatically generate a hazard diagnosis report, which includes the hazard type, hazard spatial location, risk level, and handling suggestions; wherein, the spatial location of the hazard in the hazard diagnosis report is determined based on the spatial coordinate system of the high-resolution three-dimensional resistivity distribution model.

[0010] Furthermore, the ensemble classifier integrates the outputs of the extreme gradient boosting classifier and the graph neural network classifier through weighted voting.

[0011] Secondly, an AI-based dam hazard diagnosis system based on three-dimensional high-density electrical resistivity tomography (EDT) specifically includes: The data acquisition and preprocessing module is used to acquire raw electrical resistivity data and environmental parameters, and to process the raw electrical resistivity data. The data fusion module is used to fuse the processed raw electrical resistivity data and environmental parameters to generate a multi-parameter data cube for the dam. The physical constraint inversion module is used to input the multi-parameter data cube of the dam into the U-shaped network model trained with physical constraint loss for inversion, and obtain a high-resolution three-dimensional resistivity distribution model. The intelligent diagnosis and risk classification module is used to make decisions based on a high-resolution three-dimensional resistivity distribution model. It integrates the hazard features automatically extracted by the three-dimensional convolutional neural network with environmental parameters weighted by the attention mechanism, and uses an integrated classifier to realize intelligent hazard identification and risk classification, and finally generates a diagnostic report.

[0012] Thirdly, an electronic device includes a processor, a memory, and a bus system, wherein the processor and the memory are connected via the bus system, the memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to realize the aforementioned artificial intelligence-based dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography.

[0013] Fourthly, a computer storage medium stores a computer software product, the computer software product including several instructions for causing a computer device to execute the aforementioned artificial intelligence-based dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography.

[0014] Compared with existing technologies, the beneficial effects of this invention are summarized as follows: (1) The multi-electrode distributed layout and intelligent switching technology are adopted to support the dynamic optimization of various electrode arrangement modes (such as Wenner, dipole, etc.). Combined with three-dimensional grid scanning and tomographic imaging algorithms, the spatial resolution and detection depth of the dam hidden danger body are significantly improved, overcoming the defects of incomplete information in traditional two-dimensional electrical methods.

[0015] (2) This invention uses a three-dimensional convolutional neural network to automatically extract multi-dimensional features of hidden dangers, and integrates environmental parameters through an attention mechanism. It utilizes an integrated learning strategy combined with extreme gradient boosting and a graph neural network classifier to achieve intelligent identification and risk classification of hidden danger types such as leakage, piping, and cracks. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the artificial intelligence-based dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the three-dimensional multi-electrode array layout and environmental sensor layout provided in the embodiments of the present invention. Detailed Implementation

[0018] like Figure 1 As shown, embodiments of the present invention propose an artificial intelligence-based method for diagnosing potential hazards in dams based on three-dimensional high-density electrical resistivity tomography, comprising the following steps: Step S1: Deploy a three-dimensional multi-electrode array along the dam body axis and key areas, and simultaneously deploy environmental sensors. By intelligently switching the electrode arrangement mode, synchronously collect resistivity, polarizability and natural potential data to obtain raw electrical resistivity data. Step S2: Perform terrain correction, noise filtering and bad pixel removal on the original electrical resistivity data to obtain the processed original electrical resistivity data. Then, use a spatiotemporal interpolation algorithm to fuse the processed original electrical resistivity data and environmental parameters collected by environmental sensors into a unified three-dimensional grid coordinate system to generate a multi-parameter data cube of the dam. Step S3: Construct a U-shaped network model by inputting the multi-parameter data cube of the dam into the U-shaped network model and outputting a high-resolution three-dimensional resistivity distribution model; wherein, the U-shaped network model is obtained by training using a total loss function that includes physical constraint loss; Step S4: Based on the high-resolution three-dimensional resistivity distribution model, a three-dimensional convolutional neural network is used to automatically extract the spatial morphological features, connectivity features, and electrical anomaly features of the hazard body; the spatial morphological features, connectivity features, and electrical anomaly features are fused with environmental parameters weighted by an attention mechanism, and an ensemble classifier is used to achieve intelligent identification and risk rating of the hazard body type, and finally generate a hazard diagnosis report.

[0019] Furthermore, in step S1, a three-dimensional multi-electrode array is deployed along the dam body axis and key areas, while environmental sensors are also deployed. By intelligently switching the electrode arrangement mode, resistivity, polarizability, and natural potential data are collected synchronously to obtain raw electrical resistivity data. The specific steps are as follows: Step S11: Install stainless steel silver-plated electrodes in a grid pattern along the axis of the dam body, with a longitudinal spacing of 5-20 meters and a transverse spacing of 2-5 meters, to form a three-dimensional observation network; Step S12: Densify the deployment of stainless steel silver-plated electrodes in key areas and dynamically adjust the spacing of the stainless steel silver-plated electrodes according to the detection depth requirements. The selection of the spacing of the stainless steel silver-plated electrodes meets the following empirical criteria: the effective detection depth of the stainless steel silver-plated electrodes is 1 / 6 to 1 / 8 of the spacing of the stainless steel silver-plated electrodes, and the spacing of the stainless steel silver-plated electrodes is not greater than the minimum design target size to be detected. The minimum target size mentioned above refers to the minimum spatial size of the potential hazard that needs to be identified according to the relevant specifications for dam safety monitoring. Step S13: After the three-dimensional multi-electrode array is deployed, environmental sensors are simultaneously deployed at key nodes and key areas of the three-dimensional multi-electrode array. The environmental sensors include temperature sensors, humidity sensors and pore water pressure sensors. Step S14: Simultaneously collect resistivity, polarizability, and natural potential data, which is achieved through intelligent switching of the arrangement mode; wherein, intelligent switching includes automatically selecting and switching between the Wenner device arrangement mode and the dipole device arrangement mode; wherein, the natural potential data is directly measured.

[0020] 3D multi-electrode array deployment and environmental sensor deployment, such as Figure 2 As shown, Figure 2 In the diagram, 1 is a high-density electrode, 2 is a temperature sensor, 3 is a humidity sensor, 4 is a pore water pressure sensor, 5 is a dam body, 6 is a high-density host, 7 is a data processing terminal, 8 is a temperature sensor receiver, 9 is a humidity sensor receiver, and 10 is a pore water pressure sensor receiver; the environmental sensors include a temperature sensor, a humidity sensor, and a pore water pressure sensor.

[0021] Specifically, high-density electrodes 1 are arranged in a three-dimensional grid along the axis and key areas on the surface of the dam body 5, and all high-density electrodes 1 are uniformly connected to a high-density host 6. The high-density host 6 collects resistivity, polarizability, and natural potential data, i.e., raw electrical resistivity data.

[0022] Meanwhile, temperature sensor 2, humidity sensor 3, and pore water pressure sensor 4 are deployed on the surface of the dam body 5 and near the three-dimensional multi-electrode array to collect environmental parameters. Temperature sensor 2 is connected to temperature sensor receiver 8 via a dedicated cable, humidity sensor 3 is connected to humidity sensor receiver 9 via a dedicated cable, and pore water pressure sensor 4 is connected to pore water pressure sensor receiver 10 via a dedicated cable.

[0023] The raw electrical resistivity data collected by the high-density host 6, along with the temperature, humidity, and pore water pressure data collected by the sensor receiver, are processed and then transmitted to the data processing terminal 7 via a 5G network or wireless local area network.

[0024] The formula for resistivity is as follows: ; in, Resistivity For device coefficients, Potential difference, unit: millivolt. Injected current, unit: milliampere; The formula for polarizability is as follows: ; in, Polarizability The potential difference is the secondary field. This is the primary field potential difference.

[0025] Further, in step S2, the original electrical resistivity data undergoes terrain correction, noise filtering, and bad pixel removal to obtain processed original electrical resistivity data. Then, a spatiotemporal interpolation algorithm is used to fuse the processed original electrical resistivity data and environmental parameters collected by environmental sensors into a unified three-dimensional grid coordinate system, generating a multi-parameter data cube for the dam. The specific steps are as follows: Step S21: Perform terrain correction on the resistivity to obtain the corrected resistivity. Use the Gauss-Seidel iterative method to eliminate the influence of terrain undulations, as shown in the following formula: ; in, The corrected resistivity is expressed in ohm-meters. This refers to the actual terrain path length, in meters. The horizontal reference path length is expressed in meters. Step S22: The corrected resistivity is filtered. Noise filtering is performed using wavelet thresholding, with wavelet coefficients obtained through a 5-level decomposition using wavelet basis functions, and then a hard thresholding function is applied. The formula for processing wavelet coefficients is as follows: ; in, These are wavelet coefficients. The threshold value is [value]. , The standard deviation of noise. This represents the total number of data points. Step S23: Based on the Laida criterion, perform defect removal on the filtered resistivity, eliminating outliers exceeding ±3 times the standard deviation, to obtain the processed resistivity, as shown in the following formula: ; in, The resistivity after treatment. The average resistivity after processing. Standard deviation; Step S24: Perform the same filtering and bad pixel removal processes as in steps S22 and S23 on the polarizability and natural potential data to obtain the processed polarizability and natural potential data. Step S25: Establish a three-dimensional grid coordinate system with a resolution of 0.5m × 0.5m × 0.2m. Use the Kriging spatiotemporal interpolation algorithm to interpolate the processed resistivity, processed polarizability, processed natural potential data, and temperature, humidity, and pore water pressure data collected by environmental sensors onto the grid of the unified three-dimensional grid coordinate system to generate a multi-parameter data cube of the dam.

[0026] The Kriging spatiotemporal interpolation algorithm is used to interpolate temperature, humidity, and pore water pressure data onto a unified three-dimensional grid coordinate system, as shown in the following formula: ; in, The values ​​of the points to be interpolated are the spatial and temporal values ​​of those points. The number of observation points participating in the interpolation. For the index of the observation point, These are the weighting coefficients. These are the observation point values, that is, the actual measured temperature, humidity, and pore water pressure values ​​at different observation points and at different times.

[0027] Further, in step S3, a U-shaped network model is constructed. The multi-parameter data cube of the dam is input into the U-shaped network model, and a high-resolution three-dimensional resistivity distribution model is output. The U-shaped network model is trained using a total loss function that includes physical constraint losses. The specific steps are as follows: Step S31: The U-shaped network model consists of an encoder and a decoder; the encoder uses 4 three-dimensional convolutional layers, i.e., the kernel size is 3×3×3 and the stride is 2 for downsampling; the decoder uses 4 transposed convolutional layers for upsampling. Step S32: The total loss function of the U-shaped network model is composed of a weighted sum of the data fitting loss and the physical constraint loss; the data fitting loss is calculated using the mean square error to determine the difference between the predicted and measured values; the physical constraint loss is calculated based on Maxwell's equations; a regularization parameter is used to balance the weights of the data fitting loss and the physical constraint loss, and the value of the regularization parameter ranges from 0.1 to 0.5. Physical constraint loss The formula is as follows: ; Where M is the total number of grid points, and ∇ is the gradient operator. Here is the conductivity value at the i-th grid point, in Siemens per meter. The potential value at the i-th grid point, in volts; Step S33: The constructed U-shaped network model is trained using the Adam optimizer with an initial learning rate of 0.001 and an exponential decay strategy. The training batch size is set to 16, and the number of training rounds is 200. After training is completed, the multi-parameter data cube of the dam is input into the trained U-shaped network model, and a high-resolution three-dimensional resistivity distribution model is output. Based on the high-resolution three-dimensional resistivity distribution model, three-dimensional visualization processing is performed to generate a three-dimensional image of the resistivity distribution inside the dam.

[0028] To ensure effective training of the U-shaped network model, this invention constructs a comprehensive training dataset, the specific construction method of which is as follows: The training samples are from the following three parts: (1) Physical model test samples. The physical model test training samples are from the standardized dam physical model test platform of Jiangxi Academy of Water Resources and Poyang Lake Model Test Research Base. Based on the typical cross-section design of homogeneous earth dam and clay core wall dam in the "Design Code for Dike Engineering" (GB50286-2013), the proportionally scaled down physical dam test models constructed using the principles of geometric similarity, mechanical similarity and seepage similarity are collected, totaling 600 sets; (2) Numerical simulation test samples. Three-dimensional dam geoelectric models with different geological conditions and different types of hidden dangers are established using finite element software. After forward simulation, they are processed by the method described in step S2, totaling 2000 sets; (3) Field measured samples. The samples are from the measured data of dams in the middle and lower reaches of the Yangtze River in Jiangxi (Jiangzhou Town, Jiujiang, and Poyang Lake shoreline) where hidden danger verification has been completed, totaling 400 sets; All samples are confirmed by borehole sampling, excavation verification or historical disease records to confirm the type and spatial location of hidden dangers.

[0029] The training set contains 3000 samples, randomly divided into a training set of 2400 samples, a validation set of 300 samples, and a test set of 300 samples in an 8:1:1 ratio. By fusing samples from physical model experiments, numerical simulation experiments, and field measurements, the model covers different geological conditions, different types of hazards, and different environmental conditions, avoiding overfitting of the model to a single data source. To improve the robustness of the U-shaped network model to complex field environments, data augmentation operations such as random noise addition, terrain disturbance, parameter scaling, and spatial rotation were performed on the training samples during training.

[0030] Furthermore, in step S4, based on a high-resolution three-dimensional resistivity distribution model, a three-dimensional convolutional neural network is used to automatically extract the spatial morphological features, connectivity features, and electrical anomaly features of the hazard. These features are then fused with environmental parameters weighted through an attention mechanism, and an ensemble classifier is used to achieve intelligent identification and risk rating of the hazard type, ultimately generating a hazard diagnosis report. The specific steps are as follows: Step S41: Based on the high-resolution three-dimensional resistivity distribution model, a three-dimensional convolutional neural network is used to automatically extract the hidden danger features. The three-dimensional convolutional neural network contains 4 convolutional layers with a kernel size of 3×3×3 and a stride of 1. The pooling layer uses max pooling with a pooling window of 2×2×2. A three-dimensional convolutional neural network extracts spatial morphological features, connectivity features, and electrical anomaly features, respectively. Spatial morphological features include the volume, surface area, and flattening of the hazard body. Connectivity features include Euler number and pore-throat ratio. Electrical anomaly features include resistivity anomaly degree and resistivity contrast. Step S42: Environmental parameters are weighted and fused using an attention mechanism. These parameters include temperature, humidity, and pore water pressure. The formula is as follows: ; in, For the first The attention weights for each environmental parameter, where exp is an exponential function. Weight vector The transpose of , where tanh is the hyperbolic tangent activation function. This is the weight matrix. For the first Feature vectors of environmental parameters For bias vectors, The number of environmental parameters; Step S43: The spatial morphological features, connectivity features, and electrical anomaly features are fused with environmental parameters weighted by the attention mechanism, and the ensemble classifier is used to output the hazard type and confidence probability. Step S44: Based on the confidence probability, the hidden danger risk is divided into three risk levels: low risk, medium risk, and high risk; at the same time, a pre-set hidden danger-disposal rule knowledge base is established, with hidden danger type and risk level as joint input keys. Step S45: Based on the combination of the hazard type and risk level, match and obtain corresponding targeted handling suggestions from the hazard-handling rule knowledge base; automatically generate a hazard diagnosis report, which includes the hazard type, hazard spatial location, risk level, and handling suggestions; wherein, the spatial location of the hazard in the hazard diagnosis report is determined based on the spatial coordinate system of the high-resolution three-dimensional resistivity distribution model.

[0031] Specifically, a confidence probability less than 0.3 is considered low risk, a confidence probability greater than or equal to 0.3 and less than 0.7 is considered medium risk, and a confidence probability greater than or equal to 0.7 is considered high risk.

[0032] The rules in the hazard-handling rule knowledge base are all derived from the "Technical Specification for Management of Dike Engineering" (SL260-2014) and the "Technical Specification for Hazard Detection of Dike Engineering" (SL / T793-2020). The knowledge base uses hazard type + risk level as the joint input key to store corresponding graded handling measures. The hazard types, their electrical characteristics, and graded handling rules covered by the hazard-handling rule knowledge base are as follows: (1) Leakage: This manifests as continuous seepage channels within the dam soil, with a significant decrease in local resistivity, exceeding 50% of the background value, mostly distributed at the junction of the dam body and foundation. Treatment rules: In low-risk situations, record the location and parameters of potential hazards in the daily inspection log, and strengthen post-rain inspections to observe for seepage; in medium-risk situations, use the splitting grouting method for seepage prevention, controlling the grouting pressure at 0.2MPa-0.3MPa, and conduct quality testing after grouting; in high-risk situations, immediately activate the emergency plan to evacuate downstream residents, employ a combination of pressure-seepage-covering and anti-seepage curtain solutions for emergency repairs, and conduct comprehensive reinforcement after the emergency repairs are completed.

[0033] (2) Piping: This is characterized by the loss of soil particles carried by seepage, forming concentrated seepage channels. It presents as point-like or linear low resistivity anomalies and is often accompanied by a sudden increase in pore water pressure. Treatment rules: In the case of low risk, a filter layer (sand and gravel + geotextile) is set at the piping outlet, and the integrity of the filter layer is checked regularly; in the case of medium risk, the seepage control method is adopted by using a cofferdam, with the cofferdam height 0.5m higher than the seepage water level, and the piping channel is sealed by high-pressure jet grouting; in the case of high risk, the upstream water level is immediately lowered to reduce the head difference, and a combination of steel sheet piles and anti-seepage walls is used to seal the piping channel, with dedicated personnel on duty 24 hours a day.

[0034] (3) Cracks: These are cracks in the dam soil caused by uneven settlement or shrinkage, appearing as strip-shaped high resistivity anomalies. They are classified as transverse cracks and longitudinal cracks according to their direction. Treatment rules: In low-risk cases, the cracks are excavated and backfilled, with the backfill soil compaction degree not less than 95%, and a geomembrane is laid on the surface to prevent rainwater infiltration; in medium-risk cases, transverse cracks are treated with grouting and geomembrane seepage prevention, and longitudinal cracks are treated with excavation, backfilling and reinforced soil reinforcement, and settlement observation points are set up; in high-risk cases, the dam is immediately restricted from traffic and warning signs are set up, and a concrete anti-seepage wall and anchor reinforcement scheme is adopted, and the development trend of cracks is monitored in real time.

[0035] (4) Animal burrows: These are cavities dug by animals such as badgers and rats, appearing as isolated high resistivity anomalies, typically ranging in volume from 0.1 m³ to 10 m³. Treatment rules: For low-risk situations, excavation and backfilling are used, with backfill soil compacted in layers. Weeds around the burrows are removed and animal-proof fences are installed. For medium-risk situations, grouting is used for burrows deeper than 2 m, with cement-soil mixture as the grouting material. A burrow distribution survey is conducted, and surrounding animal nests are removed. For high-risk situations, the affected area is immediately sealed off to prevent personnel from approaching. A combination of excavation, backfilling, and grouting is used, and the overall stability of the dam is assessed.

[0036] (5) Termite nests: These are characterized by termite tunnels and nest cavities formed by termite activity, exhibiting a honeycomb-like structure with abnormally high resistivity, often accompanied by loose soil and abnormal moisture levels. Treatment guidelines: In low-risk situations, use termite baits for chemical control and remove termite-preyed plants around the dam; in medium-risk situations, use grouting to fill termite tunnels and nest cavities, employing a mixture of chlorpyrifos emulsion and cement slurry, and conduct comprehensive termite eradication treatment within a 50m radius; in high-risk situations, immediately carry out comprehensive termite eradication work, using fumigation to kill the termites, excavating and backfilling the nest area for reinforcement, and assessing the structural safety of the dam.

[0037] (6) No hidden dangers: The resistivity, polarizability and natural potential data are all within the normal background value range, with no obvious abnormal characteristics. No special treatment is required. Continue routine inspection and monitoring.

[0038] Furthermore, the ensemble classifier integrates the outputs of the extreme gradient boosting classifier and the graph neural network classifier through weighted voting.

[0039] The extreme gradient boosting classifier uses 100 decision trees with a maximum depth of 6; the graph neural network classifier uses a 2-layer graph convolutional network.

[0040] Embodiments of this application also provide an artificial intelligence-based dam hazard diagnosis system based on three-dimensional high-density electrical resistivity tomography, specifically including: The data acquisition and preprocessing module is used to acquire raw electrical resistivity data and environmental parameters, and to process the raw electrical resistivity data. The data fusion module is used to fuse the processed raw electrical resistivity data and environmental parameters to generate a multi-parameter data cube for the dam. The physical constraint inversion module is used to input the multi-parameter data cube of the dam into the U-shaped network model trained with physical constraint loss for inversion, and obtain a high-resolution three-dimensional resistivity distribution model. The intelligent diagnosis and risk classification module is used to make decisions based on a high-resolution three-dimensional resistivity distribution model. It integrates the hazard features automatically extracted by the three-dimensional convolutional neural network with environmental parameters weighted by the attention mechanism, and uses an integrated classifier to realize intelligent hazard identification and risk classification, and finally generates a diagnostic report.

[0041] An embodiment of this application also provides an electronic device, which includes a processor, a memory, and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the artificial intelligence dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography.

[0042] An embodiment of this application also provides a computer storage medium storing a computer software product, the computer software product including several instructions to cause a computer device to execute the aforementioned artificial intelligence dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based method for diagnosing potential hazards in dams based on three-dimensional high-density electrical resistivity tomography, characterized in that: Includes the following steps: Step S1: Deploy a three-dimensional multi-electrode array along the dam body axis and key areas, and simultaneously deploy environmental sensors. By intelligently switching the electrode arrangement mode, synchronously collect resistivity, polarizability and natural potential data to obtain raw electrical resistivity data. Step S2: Perform terrain correction, noise filtering and bad pixel removal on the original electrical resistivity data to obtain the processed original electrical resistivity data. Then, use a spatiotemporal interpolation algorithm to fuse the processed original electrical resistivity data and environmental parameters collected by environmental sensors into a unified three-dimensional grid coordinate system to generate a multi-parameter data cube of the dam. Step S3: Construct a U-shaped network model. Input the multi-parameter data cube of the dam into the U-shaped network model and output a high-resolution three-dimensional resistivity distribution model. The U-shaped network model is trained using a total loss function that includes physical constraint losses. Specific steps are as follows: Step S31: The U-shaped network model consists of an encoder and a decoder; the encoder uses 4 three-dimensional convolutional layers, i.e., the kernel size is 3×3×3 and the stride is 2 for downsampling; the decoder uses 4 transposed convolutional layers for upsampling. Step S32: The total loss function of the U-shaped network model is composed of a weighted sum of the data fitting loss and the physical constraint loss; the data fitting loss is calculated using the mean square error to determine the difference between the predicted and measured values; the physical constraint loss is calculated based on Maxwell's equations; a regularization parameter is used to balance the weights of the data fitting loss and the physical constraint loss, and the value of the regularization parameter ranges from 0.1 to 0.

5. Physical constraint loss The formula is as follows: ； Where M is the total number of grid points, For gradient operators, Here is the conductivity value at the i-th grid point, in Siemens per meter. The potential value at the i-th grid point, in volts; Step S33: The constructed U-shaped network model is trained using the Adam optimizer with an initial learning rate of 0.001 and an exponential decay strategy. The training batch size is set to 16, and the number of training rounds is 200. After training is completed, the multi-parameter data cube of the dam is input into the trained U-shaped network model, and a high-resolution three-dimensional resistivity distribution model is output. Based on the high-resolution three-dimensional resistivity distribution model, three-dimensional visualization processing is performed to generate a three-dimensional image of the resistivity distribution inside the dam. Step S4: Based on the high-resolution three-dimensional resistivity distribution model, a three-dimensional convolutional neural network is used to automatically extract the spatial morphological features, connectivity features, and electrical anomaly features of the hazard body; the spatial morphological features, connectivity features, and electrical anomaly features are fused with environmental parameters weighted by an attention mechanism, and an ensemble classifier is used to achieve intelligent identification and risk rating of the hazard body type, and finally generate a hazard diagnosis report.

2. The artificial intelligence-based dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography (EDT) according to claim 1, characterized in that, In step S1, a three-dimensional multi-electrode array is deployed along the dam's axis and in key areas, while environmental sensors are also deployed. By intelligently switching the electrode arrangement mode, resistivity, polarizability, and natural potential data are collected simultaneously to obtain raw electrical resistivity data. The specific steps are as follows: Step S11: Install stainless steel silver-plated electrodes in a grid pattern along the axis of the dam body, with a longitudinal spacing of 5-20 meters and a transverse spacing of 2-5 meters, to form a three-dimensional observation network; Step S12: Densify the deployment of stainless steel silver-plated electrodes in key areas and dynamically adjust the spacing of the stainless steel silver-plated electrodes according to the detection depth requirements. The selection of the spacing of the stainless steel silver-plated electrodes meets the following empirical criteria: the effective detection depth of the stainless steel silver-plated electrodes is 1 / 6 to 1 / 8 of the spacing of the stainless steel silver-plated electrodes, and the spacing of the stainless steel silver-plated electrodes is not greater than the minimum design target size to be detected. Step S13: After the three-dimensional multi-electrode array is deployed, environmental sensors are simultaneously deployed at key nodes and key areas of the three-dimensional multi-electrode array. The environmental sensors include temperature sensors, humidity sensors and pore water pressure sensors. Step S14: Simultaneously collect resistivity, polarizability, and natural potential data, which is achieved through intelligent switching of the arrangement mode; wherein, intelligent switching includes automatically selecting and switching between the Wenner device arrangement mode and the dipole device arrangement mode; wherein, the natural potential data is directly measured.

3. The artificial intelligence-based dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography (EDT) according to claim 2, characterized in that, In step S2, the raw electrical resistivity data undergoes terrain correction, noise filtering, and bad pixel removal to obtain processed raw electrical resistivity data. Then, a spatiotemporal interpolation algorithm is used to fuse the processed raw electrical resistivity data with environmental parameters collected by environmental sensors into a unified three-dimensional grid coordinate system, generating a multi-parameter data cube for the dam. The specific steps are as follows: Step S21: Perform terrain correction on the resistivity to obtain the corrected resistivity, and use the Gauss-Seidel iterative method to eliminate the influence of terrain undulations. Step S22: The corrected resistivity is filtered. The noise filtering process adopts the wavelet threshold denoising method. The wavelet coefficients are obtained by decomposing the wavelet basis function into 5 levels and then processing the wavelet coefficients using the hard threshold function. Step S23: Based on the Laida criterion, perform defect removal on the filtered resistivity, removing outliers exceeding ±3 times the standard deviation range to obtain the processed resistivity. Step S24: Perform the same filtering and bad pixel removal processes as in steps S22 and S23 on the polarizability and natural potential data to obtain the processed polarizability and natural potential data. Step S25: Establish a three-dimensional grid coordinate system with a resolution of 0.5m × 0.5m × 0.2m. Use the Kriging spatiotemporal interpolation algorithm to interpolate the processed resistivity, processed polarizability, processed natural potential data, and temperature, humidity, and pore water pressure data collected by environmental sensors onto the grid of the unified three-dimensional grid coordinate system to generate a multi-parameter data cube of the dam.

4. The artificial intelligence-based dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography (EDT) according to claim 3, characterized in that, In step S4, based on the high-resolution three-dimensional resistivity distribution model, a three-dimensional convolutional neural network is used to automatically extract the spatial morphological features, connectivity features, and electrical anomaly features of the hidden danger body. The system integrates spatial morphological features, connectivity features, and electrical anomaly features with environmental parameters weighted by an attention mechanism, and utilizes an ensemble classifier to achieve intelligent identification and risk rating of hazard types, ultimately generating a hazard diagnosis report. The specific steps are as follows: Step S41: Based on the high-resolution three-dimensional resistivity distribution model, a three-dimensional convolutional neural network is used to automatically extract the hidden danger features. The three-dimensional convolutional neural network contains 4 convolutional layers with a kernel size of 3×3×3 and a stride of 1. The pooling layer uses max pooling with a pooling window of 2×2×2. Three-dimensional convolutional neural networks extract spatial morphological features, connectivity features, and electrical anomaly features, respectively. Spatial morphological characteristics: including the volume, surface area, and flattening of the potential hazard; connectivity characteristics: including Euler number and porosity-throat ratio; electrical anomaly characteristics: including resistivity anomaly and resistivity contrast. Step S42: Environmental parameters are weighted and fused using an attention mechanism. These parameters include temperature, humidity, and pore water pressure. The formula is as follows: ； in, For the first The attention weights for each environmental parameter, where exp is an exponential function. Weight vector The transpose of , where tanh is the hyperbolic tangent activation function. This is the weight matrix. For the first Feature vectors of environmental parameters For bias vectors, The number of environmental parameters; Step S43: The spatial morphological features, connectivity features, and electrical anomaly features are fused with environmental parameters weighted by the attention mechanism, and the ensemble classifier is used to output the hazard type and confidence probability. Step S44: Based on the confidence probability, the hidden danger risk is divided into three risk levels: low risk, medium risk, and high risk; at the same time, a pre-set hidden danger-disposal rule knowledge base is established, with hidden danger type and risk level as joint input keys. Step S45: Based on the combination of the hazard type and risk level, match and obtain corresponding targeted handling suggestions from the hazard-handling rule knowledge base; automatically generate a hazard diagnosis report, which includes the hazard type, hazard spatial location, risk level, and handling suggestions; wherein, the spatial location of the hazard in the hazard diagnosis report is determined based on the spatial coordinate system of the high-resolution three-dimensional resistivity distribution model.

5. The artificial intelligence-based dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography (EDT) according to claim 4, characterized in that, The ensemble classifier integrates the outputs of the extreme gradient boosting classifier and the graph neural network classifier through weighted voting.

6. An AI-based dam hazard diagnosis system based on three-dimensional high-density electrical resistivity tomography, characterized in that, The system is used to implement the artificial intelligence-based dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography as described in any one of claims 1 to 5, specifically including: The data acquisition and preprocessing module is used to acquire raw electrical resistivity data and environmental parameters, and to process the raw electrical resistivity data. The data fusion module is used to fuse the processed raw electrical resistivity data and environmental parameters to generate a multi-parameter data cube for the dam. The physical constraint inversion module is used to input the multi-parameter data cube of the dam into the U-shaped network model trained with physical constraint loss for inversion, and obtain a high-resolution three-dimensional resistivity distribution model. The intelligent diagnosis and risk classification module is used to make decisions based on a high-resolution three-dimensional resistivity distribution model. It integrates the hazard features automatically extracted by the three-dimensional convolutional neural network with environmental parameters weighted by the attention mechanism, and uses an integrated classifier to realize intelligent hazard identification and risk classification, and finally generates a diagnostic report.

7. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a bus system. The processor and the memory are connected through the bus system. The memory is used to store instructions, and the processor is used to execute the instructions stored in the memory to implement the artificial intelligence dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography as described in any one of claims 1 to 5.

8. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, which includes several instructions to cause a computer device to execute the artificial intelligence dam hazard diagnosis method based on three-dimensional high-density electrical resistivity tomography as described in any one of claims 1 to 5.

Images