Geomembrane defect detection system and method under soil and rock medium covering condition

Abstract

The present application relates to the field of engineering technology, and more particularly to a geomembrane defect detection system and method under soil and rock medium covering conditions, comprising a multi-modal adaptive sensing module, an edge computing module, a data fusion module, a digital twin module and a decision and application module; the present application overcomes the shortcomings of single resistivity or geological radar detection in the prior art, which is susceptible to medium conductivity, metal impurity interference and insufficient deep resolution, and the present scheme adopts multi-modal sensor network cooperative work combined with dynamic density layout and time period acquisition strategy; through multi-modal complementation of electrode array, ultra-wideband radar, distributed optical fiber and MEMS vibration sensor, the dominant sensing mode is automatically switched in the conductive clay area, temperature gradient interference is avoided at night, multi-band radar fusion improves the penetration depth to 15 meters and suppresses metal interference; the detection accuracy and anti-interference ability in complex medium environment are significantly improved, and the problems of high false alarm rate and large missed detection rate of traditional methods are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of engineering technology, and particularly relates to a geomembrane defect detection system and method under soil and rock medium covering conditions. BACKGROUND

[0002] In the field of civil engineering seepage prevention, geomembranes (such as HDPE, LDPE, etc.) are widely used in landfill sites, tailings ponds, water conservancy projects, etc. due to their excellent impermeability and corrosion resistance. However, in the working conditions where the soil and rock medium covering layer is thick (usually 5-15 meters) and the medium composition is complex (such as a mixture of clay, gravel, and metal impurities), geomembrane defect detection faces severe challenges.

[0003] Traditional detection methods such as electrical resistance tomography (ERT) and geological radar (GPR) can partially identify leakage points, but have significant limitations: ERT is easily disturbed by the medium background in conductive clay or high humidity areas, with a resistivity measurement error of more than 30%, resulting in a high false positive rate; GPR has severe electromagnetic wave scattering in metal impurity-rich areas, with a penetration depth that drops from the theoretical 10 meters to less than 3 meters, and insufficient resolution for millimeter-sized holes. In addition, existing systems mostly use fixed-density electrodes or sensors, which are difficult to adapt to soil heterogeneity - for example, excessive deployment in uniform sandstone areas results in resource waste, while insufficient density in joint or slope mutation areas leads to missed detection. In one case, a tailings pond missed 3 cracks due to uneven sensor coverage, ultimately causing a leakage accident. Manual inspection can supplement some defects, but is inefficient and cannot provide real-time monitoring, especially in deep coverage or dangerous slope areas, where labor costs and safety risks increase sharply.

[0004] In existing technologies, data processing relies on cloud computing, but field engineering often faces unstable networks or bandwidth limitations, resulting in a detection result delay of up to several hours, which cannot meet the real-time response needs of leakage emergencies. For example, a landfill site failed to locate the leakage point after a heavy rain due to data transmission delay, resulting in the spread of pollutants to the surrounding soil. In recent years, some research has attempted to introduce distributed optical fiber sensing or unmanned aerial vehicle infrared thermal imaging technology, but the multi-modal data fusion capability is insufficient - for example, the spatial and temporal alignment error between optical fiber vibration signals and resistivity data is more than 0.5 meters, and there is a lack of correlation analysis for small defects (<1 centimeter), with an actual detection rate of only 70%-80%. In addition, existing systems still rely on static empirical models for geomembrane aging assessment, without considering the real-time influence of temperature, humidity, dynamic load, and other environmental stresses, resulting in a life prediction error of more than 30%. For example, a water conservancy project failed to predict material softening caused by high temperature, and the geomembrane cracked on a large scale after 3 years of service, 50% shorter than the predicted life.

[0005] Therefore, there is an urgent need for a full-process solution that integrates multi-modal sensing, edge intelligence compression, graph neural network fusion, and digital twin life prediction to break through the multiple bottlenecks of precision, efficiency, and reliability. SUMMARY

[0006] In order to overcome the problems presented in the above background art, the present application proposes a geomembrane defect detection system and method under soil-rock medium coverage conditions.

[0007] The technical solution of the present application is: a geomembrane defect detection system under soil-rock medium coverage conditions, comprising:

[0008] A multi-modal adaptive sensing module for real-time acquisition of electrical, acoustic, and electromagnetic wave signals in the soil-rock medium through a multi-modal sensor network;

[0009] An edge computing module for real-time data compression and triggering of high-precision algorithms through edge computing technology;

[0010] A data fusion module for correlating multi-modal data using graph neural networks;

[0011] A digital twin module for simulating the aging process of the geomembrane, dynamically optimizing the alarm threshold and life prediction;

[0012] A decision and application module for driving sound and light alarms, coordinate positioning, and automated repair equipment based on data analysis results, achieving full-process closed-loop control from defect detection to repair execution.

[0013] As a preferred embodiment, the multi-modal adaptive sensing module specifically includes:

[0014] A11: an electrode array module for dynamic density layout through flexible electrodes, locating leakage points by detecting resistivity changes;

[0015] A12: an electromagnetic wave detection module for penetrating deep coverage using ultra-wideband radar to identify dielectric constant abnormal areas;

[0016] A13: a vibration module for capturing vibration signals through optical fiber and MEMS sensors to distinguish noise from real defects;

[0017] A14: an optical auxiliary module for combining mobile phone infrared and laser scanning to screen for shallow temperature abnormalities and surface deformation.

[0018] A15: a salt solution tracing unit for verifying the leakage position by conductivity changes after injecting salt water.

[0019] As a preferred embodiment, the electrode array module adopts a dynamic density strategy for layout, and the specific operation steps include:

[0020] S11: Preliminary scan. First, metal electrodes are laid out according to the initial electrode spacing of 5m. At the same time, the ground-penetrating radar is started to conduct a large-scale scan. The laid-out metal electrodes and ground-penetrating radar are used to obtain impedance and electromagnetic wave reflection data, and generate a preliminary resistivity distribution map and reflection anomaly thermal map.

[0021] S12: Abnormal region identification. The ERT low-resistivity region and the GPR reflection amplitude abrupt change region are marked as suspected defects. The abnormal points are clustered by the DBSCAN algorithm to remove isolated noise and determine the range of the region to be encrypted.

[0022] S13: Dynamic encryption deployment, inserting moving electrodes at 0.5m intervals within the boundaries of clustered abnormal regions;

[0023] S14: Real-time data fusion, combining newly added electrode data to update ERT imaging, iterating until the detection resolution meets the requirements;

[0024] S15: Resource recycling and reuse. When the encrypted area is confirmed to be free of defects, the mobile electrode is recycled to other areas for reuse.

[0025] Preferably, the electrode array module collects data from 20:00 to 6:00 every day during operation, outputs a three-dimensional resistivity distribution map, and marks low-resistivity abnormal areas.

[0026] Preferably, when the edge computing module compresses data in real time and triggers high-precision algorithms using edge computing technology, it specifically includes:

[0027] S21: Data compression. The original sensor data is compressed using wavelet transform, reducing its volume to 10% of its original size.

[0028] S22: Anomaly detection: Set a dynamic threshold based on the initial data collected by the multimodal adaptive sensing module, and detect the data from the original sensor based on the set dynamic threshold.

[0029] S23: Model inference, using the lightweight CNN built into the edge computing module to analyze abnormal data, analyze the specific defect type and preliminary coordinates.

[0030] Preferably, when the edge computing module compresses the original sensor data using wavelet transform, reducing the volume of the original sensor data to 10% of its original size, it specifically includes:

[0031] S31: The ERT voltage signal is decomposed into low-frequency approximate components and high-frequency detail components through the Daubechies wavelet basis.

[0032] S32: hard threshold filtering is performed on the high-frequency detail component, low-frequency energy is retained, and low-amplitude noise coefficients are discarded;

[0033] S33: based on the retained approximation and the threshold-processed detail component, the compressed signal is restored through inverse wavelet transform;

[0034] S34: FFT calculation is performed on the vibration sensor waveform to generate a frequency spectrum energy distribution;

[0035] S35: the effective frequency band of 0-500Hz is intercepted, and high-frequency noise is discarded;

[0036] S36: the amplitude is normalized to 8-bit integer, the phase information is ignored, and the non-zero coefficients are stored.

[0037] As preferred, when analyzing abnormal data, analyzing specific defect types and preliminary coordinates by using a lightweight CNN built in an edge computing module, the structure of the lightweight CNN includes:

[0038] A21: an input layer including a resistivity image channel and a vibration spectrum feature channel, the resistivity image channel is used to input a resistivity single-channel grayscale image with a size of 64x64x1, and the vibration spectrum feature is used to input a 64-dimensional vector of normalized energy values after FFT processing of vibration data;

[0039] A22: an image processing branch including a convolutional layer one, a convolutional layer two and a convolutional layer three, the convolutional layer one is a 3x3 convolution with 8 filters, the convolutional layer two is a 3x3 convolution with 16 filters, and the convolutional layer three is a 3x3 convolution with 32 filters;

[0040] A23: a concatenation layer used to concatenate a 2048-dimensional vector after flattening the image with a 64-dimensional vector of vibration spectrum, to obtain a 2112-dimensional feature;

[0041] A24: a fully connected layer including a fully connected layer one and a fully connected layer two, the fully connected layer one is used to reduce the 2112-dimensional feature to 124 dimensions, and the fully connected layer two is used to reduce the 124-dimensional feature to 3 dimensions.

[0042] As preferred, when correlating multi-modal data by using a graph neural network, the data fusion module specifically includes:

[0043] S41: time-space alignment, time synchronization is performed on all sensor data by using a GPS clock label, and spatial registration is performed on the sensor data;

[0044] S42: graph neural network analysis, the data after time-space alignment is analyzed by using a graph neural network to obtain a prediction point and a defect probability of the geomembrane defect;

[0045] S43: Three-dimensional defect map generation, dividing the space into 0.1m x 0.1m x 0.1m voxels, assigning each voxel a defect probability value, and visualizing the defect probability value by color.

[0046] As preferred, when analyzing the spatio-temporally aligned data using a graph neural network to obtain the predicted points and defect probabilities of the geomembrane defects, the structure of the graph neural network is positioned as follows: each sensor is a node containing a feature vector, an edge is established between nodes with a physical distance of less than one meter, graph convolution layers are used to aggregate the information of adjacent nodes, and an attention mechanism is introduced to dynamically allocate the correlation weight between nodes, and finally the defect probability of each node is output through the Sigmoid function.

[0047] As preferred, the digital twin module simulates the aging process of the geomembrane and dynamically optimizes the alarm threshold and life prediction, specifically including:

[0048] S51: Material attenuation equation modeling, establishing a tensile strength attenuation equation based on the characteristics of HDPE material to quantify the effects of time, temperature, and load on aging, wherein the principle formula of the tensile strength attenuation equation is:

[0049] ;

[0050] wherein, is the initial tensile strength, is the time attenuation coefficient, is the first stress sensitivity coefficient, and , is the second stress sensitivity coefficient, and , is the cumulative load stress;

[0051] S52: Real-time data input, real-time collection of environmental parameters including temperature, stress, and humidity through fiber-optic temperature measurement, MEMS pressure sensor, and humidity sensor;

[0052] S53: Dynamic threshold adjustment, combining rule engine and random forest model prediction to optimize alarm triggering conditions;

[0053] S54: Residual life calculation, using finite element simulation to simulate the aging rate of stress concentration areas and generate a residual life heat map;

[0054] S55: Risk classification, dividing risk levels according to defect probability and residual life to guide repair priority;

[0055] S56: Maintenance suggestion generation, automatically prompting geomembrane replacement or reinforcement scheme for areas with a life less than a set threshold and a defect probability greater than 70%;

[0056] S57: Threshold adaptive adjustment, dynamically adjust sensor sensitivity according to environmental changes;

[0057] S58: Visualization output, show life prediction, risk distribution and maintenance work order through WebGL interactive interface or CAD compatible format.

[0058] The geomembrane defect detection method under the condition of soil and rock medium covering comprises the following steps:

[0059] S61: Multi-modal data acquisition, dynamically deploy electrode array and ultra-wideband radar scanning, real-time acquisition of resistivity, electromagnetic wave, vibration and optical signal, and mark abnormal area;

[0060] S62: Edge intelligent compression and detection, wavelet transform compression ERT data, FFT truncated vibration spectrum, lightweight CNN analyzes abnormal type and generates preliminary coordinates;

[0061] S63: Multi-source data fusion, time and space alignment sensor data, graph neural network fusion of electrical, acoustic and temperature features, and generation of three-dimensional defect probability map;

[0062] S64: Digital twin and life prediction, based on material aging model and real-time environmental data, dynamically adjust alarm threshold, predict remaining life and divide risk level.

[0063] The beneficial effects of the present application are:

[0064] 1. Compared with the prior art of single resistivity or geological radar detection, which is easy to be disturbed by medium conductivity and metal impurities and has insufficient deep resolution, the present scheme adopts multi-modal sensor network cooperative work, combined with dynamic density layout and time period acquisition strategy; through the multi-modal complementation of electrode array, ultra-wideband radar, distributed optical fiber and MEMS vibration sensor, the dominant sensing mode is automatically switched in the conductive clay area, the temperature gradient interference is avoided at night, the multi-band radar fusion improves the penetration depth to 15 meters and suppresses the metal interference; this scheme significantly improves the detection accuracy and anti-interference ability in complex medium environment, the deep defect positioning error is less than 0.3 meters, and the millimeter level hole recognition rate is improved to 95%, effectively solving the problems of high false alarm rate and large missed detection rate of traditional methods;

[0065] 2. Compared with the prior art which relies on cloud computing and causes high delay and large bandwidth pressure, the present scheme realizes real-time data compression and local analysis through the edge computing module; wavelet transform and FFT spectrum truncation technology are adopted to compress the original data to 10%, combined with a lightweight CNN model (parameter quantity 1MB) to realize millisecond level defect classification and coordinate positioning; this scheme shortens the response time to less than 0.1 seconds, reduces the data transmission energy consumption by 90%, supports real-time detection and decision making in no network environment, and significantly improves the adaptability and system reliability in complex field conditions.

[0066] 3、In view of the defects of independent processing of multi-sensor data, large space-time alignment error of prior art, the scheme constructs a sensor node topology graph based on a graph neural network (GNN), dynamically correlates multi-modal features such as resistivity, vibration and temperature through a space-time alignment algorithm and an attention mechanism; the scheme realizes deep fusion of multi-source data, improves the defect detection rate below 1 cm to 95%, the three-dimensional imaging positioning accuracy is ±0.1 meters, the false alarm rate is reduced by 50%, and seamless docking with engineering modeling software (BIM / CAD) is supported, providing visual decision support for accurate construction and repair;

[0067] 4、In view of the defects of static threshold alarm and artificial experience repair of the prior art, the scheme integrates a material aging model and environmental data through a digital twin platform, dynamically optimizes the alarm threshold and predicts the remaining life (error <10%), and simultaneously links robots and unmanned aerial vehicles to realize full-process closed-loop control; based on the HDPE tensile strength decay equation and real-time temperature and humidity and load data, the system self-adaptively adjusts the detection sensitivity and repair strategy, improves the repair efficiency to 2 places / hour, reduces material waste by 60%, shortens the maintenance response time in high-risk areas to within 30 minutes, and significantly improves the engineering safety and the whole life cycle management efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0068] Fig. 1 A schematic diagram of a geomembrane defect detection system under soil and rock medium coverage conditions is shown.

[0069] Fig. 2 A work flow diagram of a geomembrane defect detection method under soil and rock medium coverage conditions is shown. DETAILED DESCRIPTION

[0070] The present application will be further described below in conjunction with the drawings and examples.

[0071] Please refer to Figs. 1-2 The present application provides an example: a geomembrane defect detection system under soil and rock medium coverage conditions, comprising:

[0072] A multi-modal adaptive sensing module is used to collect electrical, acoustic and electromagnetic wave signals in the soil and rock medium in real time through a multi-modal sensor network;

[0073] An edge computing module is used to compress data in real time through edge computing technology and trigger high-precision algorithms;

[0074] A data fusion module is used to correlate multi-modal data using a graph neural network;

[0075] A digital twin module is used to simulate the aging process of the geomembrane, dynamically optimize the alarm threshold and life prediction;

[0076] A decision and application module for driving sound and light alarms, coordinate positioning and automatic repair equipment based on data analysis results, realizing full-process closed-loop control from defect detection to repair execution.

[0077] As described above, the multi-modal adaptive sensing module in the application can collect multiple signals in real time and comprehensively obtain information in soil and rock media; the edge computing module compresses data in real time and triggers high-precision algorithms to improve data processing efficiency and accuracy; the data fusion module uses graph neural networks to associate multi-modal data and deeply excavate data value; the digital twin module simulates the aging process of geomembranes, dynamically optimizes alarm thresholds and life prediction, and improves the accuracy and foresight of detection; the decision and application module realizes full-process closed-loop control from defect detection to repair execution, can timely drive sound and light alarms, coordinate positioning and start automatic repair equipment, ensures the safe and stable operation of geomembranes, and reduces maintenance costs and risks.

[0078] Further, the multi-modal adaptive sensing module specifically includes:

[0079] A11: an electrode array module for locating leakage points by detecting changes in resistivity through flexible electrode dynamic density layout;

[0080] A12: an electromagnetic wave detection module for identifying dielectric constant abnormal areas by using ultra-wideband radar to penetrate deep coverage;

[0081] A13: a vibration module for distinguishing noise from real defects by capturing vibration signals through optical fiber and MEMS sensors;

[0082] A14: an optical auxiliary module for screening shallow temperature abnormalities and surface deformation by combining mobile phone infrared and laser scanning.

[0083] A15: a salt solution tracing unit for verifying the leakage position by changes in electrical conductivity after injecting salt water.

[0084] As described above, the electrode array module in the application locates leakage points by detecting changes in resistivity through flexible electrode dynamic density layout, which can accurately lock the leakage position; the electromagnetic wave detection module identifies dielectric constant abnormal areas by using ultra-wideband radar to penetrate deep coverage, which can effectively detect potential problems in deep soil and rock media; the vibration module captures vibration signals through optical fiber and MEMS sensors and distinguishes noise from real defects, improving the accuracy of defect identification; the optical auxiliary module screens shallow temperature abnormalities and surface deformation by combining mobile phone infrared and laser scanning, which can quickly find shallow abnormal conditions; the salt solution tracing unit verifies the leakage position by changes in electrical conductivity after injecting salt water, providing a reliable supplementary means for confirming the leakage position. Multiple modules work together to achieve comprehensive and multi-level accurate detection of geomembrane defects.

[0085] Further, the electrode array module is laid out using a dynamic density strategy, and the specific operation steps include:

[0086] S11: preliminary scanning, first, lay out metal electrodes with an initial electrode spacing of 5m, simultaneously start a large-scale scan using a geological radar, use the laid-out metal electrodes and the geological radar to obtain resistance and impedance and electromagnetic wave reflection data, and generate a preliminary resistivity distribution map and a reflection anomaly heat map;

[0087] S12: abnormal area identification, mark the ERT low-resistance area and the GPR reflection amplitude mutation area as suspected defects, and cluster the abnormal points through the DBSCAN algorithm to eliminate isolated noise and determine the range of the area that needs to be encrypted.

[0088] S13: dynamic encryption deployment, insert mobile electrodes in the 0.5m spacing within the boundary of the clustered abnormal area;

[0089] S14: real-time data fusion, update the ERT imaging in combination with the new electrode data, and iterate until the detection resolution meets the requirements;

[0090] S15: resource recycling and reuse, when the encrypted area detection is confirmed to have no defects, recycle the mobile electrodes to other areas for reuse.

[0091] As described above, the present application lays out metal electrodes with a spacing of 5m through preliminary scanning and cooperates with a geological radar to obtain data and generate related images, providing a basis for subsequent analysis; abnormal area identification marks suspected defects and uses an algorithm to cluster and eliminate noise, accurately determines the area that needs to be encrypted, and avoids blind encryption; dynamic encryption deployment inserts mobile electrodes in the 0.5m spacing within the boundary of the abnormal area, significantly improving the detection accuracy of the defect area; real-time data fusion updates the imaging in combination with the new electrode data and iterates to ensure that the detection resolution meets the requirements, ensuring the reliability of the detection results; the resource recycling and reuse mechanism recycles the mobile electrodes for use in other areas when the encrypted area has no defects, effectively improving the utilization rate of electrode resources and reducing the detection cost.

[0092] Further, the electrode array module performs data acquisition from 20:00 to 6:00 every day and outputs a resistivity three-dimensional distribution map and marks the low-resistance abnormal area.

[0093] As described above, the electrode array module in the application works in a specific time period (20:00-6:00) and outputs the resistivity three-dimensional distribution map and marks the low-resistance abnormal area, which has remarkable effects: selecting the night period for data collection can effectively avoid complex electromagnetic interference (such as industrial power, high-frequency radiation of electronic equipment) and human activity noise (vibration of construction machinery, disturbance of personnel walking), greatly improving the signal purity and data accuracy; the output resistivity three-dimensional distribution map presents the spatial distribution characteristics of the earth and rock medium resistivity in a stereoscopic and intuitive form, helping the detection personnel quickly understand the electrical property change law of the medium around the geomembrane; accurately marking the low-resistance abnormal area provides clear guidance for quickly locking potential leakage, damage and other defects, significantly shortens the defect investigation time, improves the detection efficiency, ensures the long-term safe operation of the geomembrane, and reduces the risk of engineering accidents and maintenance costs caused by defects.

[0094] Further, when the edge computing module compresses data in real time through edge computing technology and triggers high-precision algorithms, it specifically includes:

[0095] S21: Data compression, the original sensor data is compressed through wavelet transform, and the volume of the original sensor data is compressed to 10% of the original;

[0096] S22: Abnormality detection, setting a dynamic threshold according to the initial data collected by the multi-modal adaptive sensing module, and detecting the original sensor data according to the set dynamic threshold;

[0097] S23: Model inference, analyzing the abnormal data by using the built-in lightweight CNN of the edge computing module to analyze the specific defect type and preliminary coordinates.

[0098] As described above, the application compresses the original sensor data to 10% of the original volume through wavelet transform in the data compression link, greatly reducing the data transmission amount, reducing the network bandwidth occupation, improving the data transmission efficiency, and saving storage resources for subsequent processing; in the abnormality detection link, the dynamic threshold is set according to the initial data collected by the multi-modal adaptive sensing module, and the original sensor data is detected accordingly, which can more accurately capture abnormal changes in the data, improve the sensitivity and accuracy of abnormality detection; in the model inference link, the built-in lightweight CNN is used to analyze the abnormal data, which can quickly and accurately determine the specific defect type and preliminary coordinates, providing strong support for subsequent accurate positioning and repair, and realizing efficient and accurate data processing and defect analysis as a whole.

[0099] Further, when the edge computing module compresses the original sensor data through wavelet transform, the volume of the original sensor data is compressed to 10% of the original, specifically including:

[0100] S31: multi-level decomposition of the ERT voltage signal by Daubechies wavelet basis, separating low-frequency approximation components and high-frequency detail components;

[0101] S32: hard threshold filtering of the high-frequency detail components, retaining low-frequency energy and discarding low-amplitude noise coefficients;

[0102] S33: based on the retained approximation and threshold-processed detail components, restoring the compressed signal by inverse wavelet transform;

[0103] S34: FFT calculation of the vibration sensor waveform to generate a frequency spectrum energy distribution;

[0104] S35: intercepting the 0-500Hz effective frequency band and discarding high-frequency noise;

[0105] S36: amplitude normalization to 8-bit integer, ignoring phase information, and storing non-zero coefficients.

[0106] As described above, the present application effectively retains the low-frequency energy, which is the key information, and discards the low-amplitude noise coefficients, reducing redundant data, by multi-level decomposition of the ERT voltage signal by Daubechies wavelet basis and processing of the high-frequency detail components; based on the retained approximation and processed detail components, the signal is restored by inverse wavelet transform, ensuring that the main features of the signal are retained; the data is further purified by performing FFT calculation on the vibration sensor waveform and intercepting the effective frequency band and discarding high-frequency noise; the amplitude is normalized to 8-bit integer and the phase information is ignored, and the non-zero coefficients are stored, which significantly reduces the data volume to 10% of the original under the premise of ensuring data availability, significantly reduces the data storage and transmission cost, improves the data processing efficiency, and provides strong support for subsequent real-time analysis and decision-making.

[0107] Further, when analyzing abnormal data using the lightweight CNN built-in the edge computing module to analyze the specific defect type and preliminary coordinates, the structure of the lightweight CNN includes:

[0108] A21: an input layer including a resistivity image channel and a vibration spectrum feature channel, the resistivity image channel being used to input a resistivity single-channel grayscale image with a size of 64x64x1, and the vibration spectrum feature being used to input a 64-dimensional vector of normalized energy values after FFT processing of vibration data;

[0109] A22: an image processing branch including convolution layer one, convolution layer two and convolution layer three, the convolution layer one being a 3x3 convolution with 8 filters, the convolution layer two being a 3x3 convolution with 16 filters, and the convolution layer three being a 3x3 convolution with 32 filters;

[0110] A23: concatenation layer, used for concatenating the 2048-dimensional vector after flattening the image and the 64-dimensional vector of vibration spectrum, to obtain a 2112-dimensional feature;

[0111] A24: fully connected layer, including fully connected layer one and fully connected layer two, the fully connected layer one is used for reducing the dimension of the 2112-dimensional feature to 124-dimensional, and the fully connected layer two is used for reducing the dimension of the 124-dimensional feature to 3-dimensional.

[0112] As described above, the present application sets up resistivity image channel and vibration spectrum feature channel through input layer, respectively inputs resistivity single-channel gray image and 64-dimensional vector of normalized energy value, integrates multi-source data information, provides rich features for defect analysis; image processing branch extracts deep-level features of resistivity image through three convolution layers with different filter numbers, enhances the capture ability of image features; the concatenation layer concatenates image flattened vector and vibration spectrum vector, fuses image and spectrum features, so that the model can comprehensively consider different types of data information; the fully connected layer gradually compresses the feature dimension through two times of dimension reduction, finally obtains 3-dimensional feature, which not only reduces the model calculation amount, but also accurately extracts key features for defect type and preliminary coordinate analysis, ensures the analysis accuracy, improves the operation efficiency of the model, and makes it more suitable for edge computing scene.

[0113] Further, the data fusion module specifically includes the following steps when using the graph neural network to associate multi-modal data:

[0114] S41: time-space alignment, time synchronization is performed on all sensor data through GPS clock label, and spatial registration is performed on sensor data;

[0115] S42: graph neural network analysis, using graph neural network to analyze the data after time-space alignment, to obtain the prediction point and defect probability of geomembrane defect;

[0116] S43: three-dimensional defect map generation, dividing the space into 0.1m×0.1m×0.1m voxels, assigning a defect probability value to each voxel, and visualizing the defect probability value through color output.

[0117] As described above, the data fusion module in the application has obvious technical advantages in associating multi-modal data using a graph neural network: the time and space alignment process uses GPS clock labels to achieve time synchronization and spatial registration of all sensor data, eliminating the deviations of different sensors in time and space dimensions, and providing a unified and accurate data basis for subsequent analysis; the graph neural network analysis process uses its powerful relationship modeling capability to deeply mine the data after time and space alignment, accurately obtain the prediction points and defect probabilities of the geomembrane defects, and improve the accuracy and reliability of defect prediction; the three-dimensional defect map generation process divides the space into fine voxels and assigns defect probability values, visualizes the output through color, and intuitively presents the spatial distribution of the geomembrane defects, facilitating the detection personnel to quickly understand the defect location and severity, providing a clear guide for subsequent repair work, and effectively improving the efficiency and quality of geomembrane defect detection and processing.

[0118] Further, when the data after time and space alignment is analyzed using a graph neural network to obtain the prediction points and defect probabilities of the geomembrane defects, the structure of the graph neural network is positioned as follows: each sensor is a node with a feature vector, an edge is established between nodes with a physical distance of less than one meter, graph convolution layers are used to aggregate the information of adjacent nodes, an attention mechanism is introduced to dynamically allocate the correlation weights between nodes, and finally the defect probability of each node is output through a Sigmoid function.

[0119] As described above, the application takes each sensor as a node and assigns a feature vector, constructing a clear data representation that can fully exploit the characteristics of the sensor itself; an edge is established between nodes based on a physical distance of less than one meter, reasonably constructing the connection relationship between nodes, which conforms to the actual physical scenario; graph convolution layers are used to aggregate the information of adjacent nodes, effectively integrating the data within the local region and enhancing the model's learning ability for local features; an attention mechanism is introduced to dynamically allocate the correlation weights between nodes, allowing the model to automatically adjust the weights according to the actual correlation between different nodes, improving the model's ability to model complex relationships; finally, the defect probability of each node is output through a Sigmoid function, which intuitively reflects the possibility of defects at the location of each sensor, providing a reliable basis for accurately locating the prediction points of geomembrane defects, and helping to improve the accuracy and efficiency of defect detection.

[0120] Further, the digital twin module simulates the aging process of the geomembrane and dynamically optimizes the alarm threshold and life prediction, specifically including:

[0121] S51: Material attenuation equation modeling, a tensile strength attenuation equation is established based on the characteristics of HDPE material to quantify the effects of time, temperature, and load on aging, wherein the principle formula of the tensile strength attenuation equation is:

[0122] ;

[0123] wherein, is the initial tensile strength, is the time decay coefficient, is the first stress sensitivity coefficient, and , is the second stress sensitivity coefficient, and , is the cumulative load stress;

[0124] S52: Real-time data input, real-time collection of environmental parameters including temperature, stress and humidity through optical fiber temperature measurement, MEMS pressure sensor and humidity sensor;

[0125] S53: Dynamic threshold adjustment, combining rule engine and random forest model prediction to optimize alarm triggering conditions;

[0126] S54: Residual life calculation, using finite element simulation to simulate stress concentration area aging rate to generate residual life thermal map;

[0127] S55: Risk classification, according to defect probability and residual life to divide risk level, guide repair priority;

[0128] S56: Maintenance suggestion generation, for areas with life less than a set threshold and defect probability greater than 70%, automatically prompt geomembrane replacement or reinforcement scheme;

[0129] S57: Threshold adaptive adjustment, dynamically adjust sensor sensitivity according to environmental changes;

[0130] S58: Visualization output, through WebGL interactive interface or CAD compatible format, show life prediction, risk distribution and maintenance work order.

[0131] As described above, the material attenuation equation modeling in the present application quantifies the influence of time, temperature, and load on aging based on the characteristics of HDPE materials, providing a scientific basis for accurately simulating the aging process of geomembranes; real-time data input is collected through various sensors to capture environmental parameters, ensuring the timeliness and accuracy of the data, and providing reliable data support for subsequent analysis; dynamic threshold adjustment combined with rule engine and random forest model prediction optimizes the alarm trigger conditions, improving the accuracy and adaptability of the alarm; residual life calculation uses finite element simulation to generate a heat map, which intuitively presents the aging condition and residual life of each region of the geomembrane; risk classification divides the grades according to the defect probability and residual life, which helps to reasonably plan the repair work and improve the resource utilization efficiency; maintenance suggestions automatically prompt replacement or reinforcement schemes for specific areas, providing specific guidance for actual maintenance work; threshold self-adaptive adjustment dynamically adjusts the sensitivity of the sensor according to environmental changes, enhancing the stability and reliability of the system; visual output displays key information through various interface formats, making it easy for managers to intuitively understand the status of the geomembrane and providing strong support for decision-making, which overall improves the intelligent level of geomembrane aging monitoring, life prediction, and maintenance management.

[0132] The geomembrane defect detection method under the condition of soil and rock medium covering includes the following steps:

[0133] S61: Multi-modal data acquisition, dynamically deploying electrode array and ultra-wideband radar scanning, real-time acquisition of resistivity, electromagnetic wave, vibration, and optical signal, and marking abnormal areas;

[0134] S62: Edge intelligent compression and detection, wavelet transform compression ERT data, FFT truncated vibration spectrum, lightweight CNN analysis of abnormal types and generation of preliminary coordinates;

[0135] S63: Multi-source data fusion, spatio-temporal alignment of sensor data, graph neural network fusion of electrical, acoustic, and temperature features, generation of a three-dimensional defect probability map;

[0136] S64: Digital twin and life prediction, based on material aging model and real-time environmental data, dynamic adjustment of alarm threshold, prediction of residual life and division of risk grade.

[0137] As described above, the multi-modal data acquisition in the application collects various signals in real time and marks abnormal areas by dynamically deploying various devices, comprehensively obtaining information of geomembranes and their surrounding environment, and providing a rich data basis for accurate detection; the edge intelligent compression and detection link uses wavelet transform, FFT truncation and lightweight CNN to effectively compress data, analyze abnormal types and generate preliminary coordinates, which reduces data transmission and storage costs, quickly locates the defect position, and improves the detection efficiency; multi-source data fusion generates a three-dimensional defect probability map by aligning time and space and fusing multi-source features through a graph neural network, which intuitively presents the defect distribution and enhances the accuracy and reliability of defect positioning; digital twinning and life prediction dynamically adjust the alarm threshold, predict the remaining life and divide the risk level based on the material aging model and real-time environmental data, which realizes dynamic monitoring and scientific evaluation of the state of geomembranes, provides a strong basis for timely repair and maintenance, helps to prolong the service life of geomembranes, and ensures the safe and stable operation of the project.

[0138] Example one: geomembrane defect detection and repair in a landfill

[0139] Technical solution application overview:

[0140] Under the condition of the landfill cover layer (thickness 8m), deploy multi-modal adaptive sensing modules:

[0141] Dynamic deployment of electrode array: deploy flexible electrodes at an interval of 5m, collect data at night (20:00-6:00), generate a resistivity three-dimensional distribution map in combination with ultra-wideband radar scanning, and mark low-resistance abnormal areas (resistivity <80Ω·m); simultaneously capture leakage vibration signals through distributed optical fibers and MEMS sensors, and exclude mechanical noise interference.

[0142] Data compression and edge analysis:

[0143] Daubechies wavelet transform is used to compress ERT data to 10%, 0-500Hz frequency band of vibration signals is intercepted, and a lightweight CNN model (input 64×64 resistivity image+64-dimensional vibration spectrum) classifies 2 hole defects (confidence>90%).

[0144] Multi-source data fusion:

[0145] After time and space alignment, a graph neural network (nodes with a distance of <1m are connected, and resistivity, vibration, and temperature features are fused) generates a three-dimensional defect map, and locates 3 joint leakage points (probability>95%).

[0146] Digital twinning optimization:

[0147] Based on the HDPE aging model and real-time load / humidity data, the remaining life of the high-risk area is predicted to be <1 year, and the resistivity threshold is dynamically adjusted to 70Ω·m during the rainy season.

[0148] Closed-loop repair execution:

[0149] UWB navigation robots drill grouting for defects >5 cm (epoxy 30 minutes curing), and drones spray polyurea material (3 mm impermeable layer) for small holes. After repair, the leakage volume is reduced by 98%.

[0150] Technical effects:

[0151] Detection accuracy: ±0.1m positioning, 1cm hole detection rate >95%. Efficiency improvement: deployment time reduced by 50%, repair efficiency up to 3 / hour. Cost optimization: electrode usage reduced by 40%, operation and maintenance cost reduced by 35%.

[0152] Example two: Geomembrane health monitoring and maintenance of mine tailings

[0153] Application overview of technical solutions:

[0154] In the scenario of high-conductivity clay-covered tailings, the system performs the following operations:

[0155] Multi-modal sensing:

[0156] Disable electrical impedance method, prefer acoustic sensing (fiber vibration + acoustic emission) and UWB radar (low frequency 0.1GHz penetration 10m cover layer), combined with salt solution tracing to verify leakage points.

[0157] Edge computing optimization:

[0158] Vibration signal FFT compression transmission to edge node, lightweight CNN model identifies microcracks (confidence 85%), triggers high-density ERT encryption detection (0.5m interval).

[0159] Data fusion and simulation:

[0160] GNN analyzes the correlation between vibration energy and dielectric constant anomalies, and digital twin predicts the remaining life of high-stress areas to be 6 months, dynamically adjusts the vibration threshold to 0.3m / s².

[0161] Automatic maintenance:

[0162] Repair robots grout and reinforce the crack area according to the three-dimensional defect map (accuracy ±0.1m), and drones spray anticorrosive coating simultaneously, extending the service life of geomembrane by more than 2 years.

[0163] Technical effects:

[0164] Complex medium adaptation: Conductive clay area detection accuracy > 90%, false positive rate reduced by 50%. Life extension: Digital twin model error < 8%, high-risk area maintenance response time < 4 hours. Safety enhancement: Hazardous environment (deep coverage area, high steep slope) realizes unmanned repair, and manual intervention is reduced by 80%.

[0165] The embodiments of the application are described in detail above with reference to the accompanying drawings, but the application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A geomembrane defect detection system under soil and rock cover conditions; characterized in that: include: The multimodal adaptive sensing module is used to acquire electrical, acoustic, and electromagnetic wave signals in soil and rock media in real time through a multimodal sensor network; The edge computing module is used to compress data in real time and trigger high-precision algorithms using edge computing technology. The data fusion module is used to connect multimodal data using graph neural networks; The digital twin module is used to simulate the aging process of geomembranes and dynamically optimize alarm thresholds and lifespan predictions. The decision and application module is used to drive audible and visual alarms, coordinate positioning, and automated repair equipment based on data analysis results, realizing closed-loop control of the entire process from defect detection to repair execution; The multimodal adaptive sensing module specifically includes: A11: Electrode array module, used to locate leakage points by detecting changes in resistivity using a flexible electrode dynamic density deployment strategy. When the electrode array module is working, it collects data from 20:00 to 6:00 every day, outputs a three-dimensional resistivity distribution map, and marks low-resistivity abnormal areas. A12: Electromagnetic wave detection module, used to use ultra-wideband radar to penetrate deep coverage and identify regions with abnormal dielectric constants; A13: Vibration module, used to capture vibration signals through fiber optic and MEMS sensors to distinguish noise from real defects; A14: Optical auxiliary module, used to combine mobile phone infrared and laser scanning to screen for shallow temperature anomalies and surface deformation; A15: Salt solution tracer unit, used to verify the location of leakage by detecting changes in conductivity after salt water is injected; The electrode array module is deployed using a dynamic density strategy. The specific operation steps include: S11: Preliminary scan. First, metal electrodes are laid out according to the initial electrode spacing of 5m. At the same time, the ground-penetrating radar is started to conduct a large-scale scan. The laid-out metal electrodes and ground-penetrating radar are used to obtain impedance and electromagnetic wave reflection data, and generate a preliminary resistivity distribution map and reflection anomaly thermal map. S12: Abnormal area identification. The ERT low-resistivity area and the GPR reflection amplitude change area are marked as suspected defects. The abnormal points are clustered by the DBSCAN algorithm to remove isolated noise and determine the range of the area to be encrypted. S13: Dynamic encryption deployment, inserting moving electrodes at 0.5m intervals within the boundaries of clustered abnormal regions; S14: Real-time data fusion, combining newly added electrode data to update ERT imaging, iterating until the detection resolution meets the requirements; S15: Resource recycling and reuse. When the encrypted area is confirmed to be free of defects, the mobile electrode is recycled to other areas for reuse.

2. The geomembrane defect detection system under soil and rock cover conditions according to claim 1, characterized in that: When the edge computing module compresses data in real time and triggers high-precision algorithms using edge computing technology, it specifically includes: S21: Data compression. The original sensor data is compressed using wavelet transform, reducing its volume to 10% of its original size. S22: Anomaly detection: Set a dynamic threshold based on the initial data collected by the multimodal adaptive sensing module, and detect the data from the original sensor based on the set dynamic threshold. S23: Model inference, using the lightweight CNN built into the edge computing module to analyze abnormal data, analyze the specific defect type and preliminary coordinates.

3. The geomembrane defect detection system under soil and rock cover conditions according to claim 2, characterized in that: When the edge computing module compresses the original sensor data using wavelet transform, reducing the volume of the original sensor data to 10% of its original size, it specifically includes: S31: The ERT voltage signal is decomposed into low-frequency approximate components and high-frequency detail components through the Daubechies wavelet basis. S32: Performs hard threshold filtering on high-frequency detail components, retaining low-frequency energy and discarding low-amplitude noise figures; S33: Based on the preserved approximation and thresholded detail components, the compressed signal is recovered by inverse wavelet transform; S34: Perform FFT calculation on the vibration sensor waveform to generate the spectral energy distribution; S35: Extracts the effective frequency band of 0-500Hz and discards high-frequency noise; S36: Normalize the amplitude to an 8-bit integer, ignore phase information, and store non-zero coefficients.

4. The geomembrane defect detection system under soil and rock cover conditions according to claim 3, characterized in that: When using the lightweight CNN built into the edge computing module to analyze abnormal data, specifically identifying defect types and preliminary coordinates, the structure of the lightweight CNN includes: A21: Input layer, including resistivity image channel and vibration spectrum feature channel. The resistivity image channel is used to input a single-channel grayscale image of resistivity with a size of 64×64×1. The vibration spectrum feature is used to input a 64-dimensional vector of energy values ​​normalized after FFT processing of vibration data. A22: Image processing branch, including convolutional layer 1, convolutional layer 2 and convolutional layer 3. Convolutional layer 1 is a 3×3 convolution with 8 filters, convolutional layer 2 is a 3×3 convolution with 16 filters, and convolutional layer 3 is a 3×3 convolution with 32 filters. A23: The stitching layer is used to stitch the 2048-dimensional vector of the flattened image with the 64-dimensional vector of the vibration spectrum to obtain 2112-dimensional features; A24: Fully connected layer, including fully connected layer 1 and fully connected layer 2. Fully connected layer 1 is used to reduce the dimensionality of 2112-dimensional features to 124-dimensional features, and fully connected layer 2 is used to reduce the dimensionality of 124-dimensional features to 3-dimensional features.

5. The geomembrane defect detection system under soil and rock cover conditions according to claim 4, characterized in that: When using graph neural networks to correlate multimodal data, the data fusion module specifically includes: S41: Spatiotemporal alignment, which synchronizes all sensor data in time using GPS clock tags and performs spatial registration of sensor data; S42: Graph Neural Network Analysis. Graph neural networks are used to analyze spatiotemporally aligned data to obtain prediction points and probability of defects in geomembranes. S43: 3D defect map generation, which divides the space into 0.1m×0.1m×0.1m voxels, assigns a defect probability value to each voxel, and visualizes the defect probability value through color.

6. The geomembrane defect detection system under soil and rock cover conditions according to claim 5, characterized in that: When using graph neural networks to analyze spatiotemporally aligned data to obtain predicted points and probability of defects in geomembranes, the structure of the graph neural network is defined as follows: each sensor is a node containing a feature vector; edges are established when the physical distance between nodes is less than one meter; information of neighboring nodes is aggregated using graph convolutional layers; and an attention mechanism is introduced to dynamically allocate the association weights between nodes. Finally, the defect probability of each node is output through the Sigmoid function.

7. The geomembrane defect detection system under soil and rock cover conditions according to claim 6, characterized in that: The digital twin module, in simulating the aging process of geomembranes and dynamically optimizing alarm thresholds and lifespan predictions, specifically includes: S51: Material attenuation equation modeling, establishing a tensile strength attenuation equation based on HDPE material properties, and quantifying the effects of time, temperature and load on aging; S52: Real-time data input, through fiber optic temperature measurement, MEMS pressure sensor and humidity sensor, to collect environmental parameters including temperature, stress and humidity in real time; S53: Dynamic threshold adjustment, combining rule engine and random forest model prediction to optimize alarm triggering conditions; S54: Remaining life calculation, using finite element simulation to simulate the aging rate of stress concentration areas and generate a remaining life thermal map; S55: Risk classification, which classifies risk levels based on defect probability and remaining lifetime to guide repair priorities; S56: Maintenance suggestion generation. For areas where the lifespan is less than the set threshold and the defect probability is greater than 70%, the system will automatically suggest geomembrane replacement or reinforcement solutions. S57: Threshold adaptive adjustment, dynamically adjusting sensor sensitivity according to environmental changes; S58: Visual output, displaying life predictions, risk distributions, and maintenance work orders through a WebGL interactive interface or CAD-compatible format.

8. A method for detecting defects in geomembranes under soil and rock cover conditions, applied to the system described in any one of claims 1-7, characterized in that: Includes the following steps: S61: Multimodal data acquisition, dynamic deployment of electrode array and ultra-wideband radar scanning, real-time acquisition of resistivity, electromagnetic waves, vibration and optical signals, and marking of abnormal areas; S62: Edge intelligent compression and detection, wavelet transform to compress ERT data, FFT to truncate vibration spectrum, lightweight CNN to analyze anomaly types and generate preliminary coordinates; S63: Multi-source data fusion, spatiotemporal alignment of sensor data, graph neural network fusion of electrical, acoustic and temperature features to generate a three-dimensional defect probability map; S64: Digital Twin and Life Prediction, based on material aging models and real-time environmental data, dynamically adjusts alarm thresholds, predicts remaining life, and classifies risk levels.

Images