A buried pool body concealed leakage point detection method based on low-frequency electromagnetic waves

Abstract

The application discloses a buried pool body concealed leakage point detection method based on low-frequency electromagnetic waves, and belongs to the technical field of buried pool body leakage detection. The method emits low-frequency electromagnetic waves in the buried pool body, collects electromagnetic field intensity signals of the ground outside the pool body, performs time-space filtering processing on the collected data, extracts reflection characteristics and spatial characteristics from the denoised data, constructs a multi-feature data set, constructs a domain adaptive neural network comprising a feature extractor, an adaptive fusion layer, a leakage predictor and a domain discriminator, performs adversarial training on the basis of domain labels of different detection scenes as supervision signals, inputs a to-be-detected sample into the trained network, and outputs leakage point position and leakage intensity discrimination results. The application adopts low-frequency electromagnetic waves in combination with a domain adaptive neural network, does not need to empty the pool body or damage the ground, can adapt to different geological conditions and pool body structures, and realizes rapid, nondestructive and high-precision detection of the concealed leakage points of the buried pool body.

Description

Technical Field

[0001] This invention relates to the field of buried pool leakage detection technology, and more specifically, to a method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves. Background Technology

[0002] Buried pools in production enterprises come in various types, including raw material storage pools, temporary storage pools for production wastewater, emergency response pools, and final treatment pools. Among these, temporary storage pools for production wastewater have the highest risk of leakage due to the complexity of pollutants and their long retention time. In terms of structural materials, concrete pools are prone to aging and cracking, HDPE membrane geomembrane pools are susceptible to mechanical damage, and fiberglass pools are easily degraded in strong acid and alkaline environments. Once a leak occurs, it will severely pollute the soil and groundwater; therefore, it is urgent to detect leaks as early and as minimally as possible.

[0003] However, buried ponds are located below the surface and are usually filled with sewage, presenting significant limitations to existing detection technologies. Traditional methods require emptying the sewage, severely impacting normal production; tracer methods require drilling observation wells near the pond, taking several days and damaging factory roads. Other physical detection methods, such as ground-penetrating radar and high-density electrical resistivity tomography (EDT), suffer from shallow detection depth, susceptibility to environmental interference, and insufficient resolution of minute leaks. Acoustic detection struggles to identify weak leak signals in noisy factory environments; fiber optic thermometry requires excavation to bury optical fibers, damaging the ground and failing to meet non-destructive testing requirements.

[0004] Existing solutions are mostly focused on leakage detection in areas such as landfill geomembranes, water conservancy projects, and pipelines. For example, methods for detecting geomembrane leakage by deploying electrode arrays are only applicable to newly constructed sites; detection methods based on seepage-thermal coupling rely on temperature anomalies; tracer methods are difficult to apply to sites lacking monitoring wells; ground-penetrating radar methods have limited detection depth and are prone to misjudgments; and fiber optic thermometry requires ground damage. Therefore, there is an urgent need for a method for detecting hidden leaks in buried ponds that does not affect normal business operations, is pollution-free, has high detection efficiency, and has high resolution for minute leaks. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a method for detecting hidden leaks in buried ponds based on low-frequency electromagnetic waves. This method combines low-frequency electromagnetic waves with a domain-adaptive neural network, eliminating the need to empty the pond or damage the surface. It can adapt to different geological conditions and pond structures, enabling rapid, non-destructive, and high-precision detection of hidden leaks in buried ponds.

[0006] This invention provides a method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves, comprising the following steps: Step 1: Place a transmitting antenna in the liquid inside the buried pool to transmit low-frequency electromagnetic waves. Collect electromagnetic field intensity signals at each measuring point on the ground surface outside the pool along the direction away from the pool wall, and record the location and time information of the measuring points. Step 2: Perform spatiotemporal filtering on the collected electromagnetic field data to suppress large-scale background field interference and environmental noise, and obtain the denoised electromagnetic field data. Step 3: Extract reflection features and spatial features from the denoised electromagnetic field data, and combine the extracted features with the corresponding domain labels and leakage labels to construct a multi-feature dataset; Step 4: Construct a domain adaptive neural network, including a feature extractor, an adaptive fusion layer, a leakage predictor, and a domain discriminator. Train the network using the multi-feature dataset, wherein the domain discriminator is trained adversarially using domain labels of different detection scenarios as supervision signals. Step 5: Input the sample to be detected into the trained network, and perform weighted fusion of reflection features and spatial features through the adaptive fusion layer to obtain the fused feature vector. Input the fused feature vector into the leakage predictor and output the leakage point location and leakage intensity discrimination results.

[0007] In this solution, step 1 specifically includes: The low-frequency electromagnetic wave transmitting antenna is immersed below the liquid surface of the underground pool through the maintenance port at the top of the pool. The antenna is immersed to a depth of half the depth of the pool, and the operating frequency of the antenna is adaptively adjusted according to the burial depth of the pool and the environmental noise characteristics. Using the pool wall as a reference, survey lines are laid out outward in a direction perpendicular to the pool wall. The length of the survey lines is determined based on the burial depth of the pool, the maximum emission angle of electromagnetic waves in the soil medium, and the estimated range of leakage impact. A non-uniform adaptive encryption strategy was adopted to deploy measurement points, and the boundary positions of each area were adaptively adjusted according to the preliminary on-site scanning results. Parallel reference survey lines are set up at symmetrical positions on both sides of the main survey line. The reference survey lines adopt the same point layout strategy as the main survey line to deduct the background electromagnetic field interference of the site. A low-frequency electromagnetic field receiver was used to collect signals at each measuring point. The axis of the receiver's ferrite rod antenna was set to be perpendicular to the pool wall. Signals were collected at each measuring point for a preset duration, and the arithmetic average was taken after repeated collection. The spatial coordinates, start and end times, and power frequency phase information of each measuring point are recorded synchronously to form raw electromagnetic field data with spatiotemporal tags.

[0008] In this scheme, the spatiotemporal filtering in step 2 specifically includes: A spatial sequence of electromagnetic field intensity data for each measurement line is constructed. Non-uniform measurement points are resampled into a uniform spacing sequence through interpolation and then spatial spectrum analysis is performed. The cutoff wavenumber of the spatial high-pass filter is determined based on the cumulative energy curve of the spatial spectrum, and the cutoff wavenumber is corrected by introducing site empirical coefficients. Each measurement point is assigned a weight coefficient proportional to the length of the corresponding representative interval. In the filtering convolution operation, the weighted sequence is used instead of the original sequence for spatial filtering. Subtract the average of the filtering results of the left and right reference lines from the filtering result of the main line to obtain the data after deducting the reference background. Short-time energy is calculated for the data after removing the reference background and compared with the statistical threshold. The historical average is used to replace the data intervals that are determined to be transient pulses. Synchronous detection technology is used for time filtering. A reference signal is obtained from the transmitting end, and an orthogonal reference signal with the same frequency and phase as the transmitted signal is generated. The received signal is coherently demodulated, and the output amplitude of the detector is extracted after low-pass filtering. When the reference signal cannot be obtained, an adaptive time bandpass filter is used. The passband center frequency and bandwidth are set according to the power spectral density of the received signal, and a notch filter is inserted.

[0009] In this scheme, the spatiotemporal filtering cascade strategy further includes: First, the raw data undergoes initial spatial filtering. The time mean of the time domain data at each measurement point is calculated to form a spatial sequence for further filtering. Then, the time series data after the first round of spatial filtering is subjected to time filtering, and the time-filtered data is reorganized into a spatial sequence for the second round of spatial filtering. Calculate the change in signal-to-noise ratio before and after filtering. If the change in signal-to-noise ratio is lower than a preset threshold, adjust the filter parameters and perform filtering again to obtain the denoised electromagnetic field data.

[0010] In this scheme, the reflection features in step 3 specifically include: Short-time Fourier transform is performed on the denoised time series data of each measurement point to extract the spectral amplitude at the transmission frequency and odd harmonics. The reflection coefficient spectrum is calculated with the spectral amplitude of the measurement point farthest from the pool wall as the reference background. The fundamental frequency reflection coefficient, harmonic ratio, reflection coefficient bandwidth and reflection phase are extracted from the reflection coefficient spectrum. Two orthogonal ferro rod antennas are set up at the receiving end to collect electromagnetic field components perpendicular to the pool wall and parallel to the pool wall, respectively, and the polarization ratio, polarization ellipse tilt angle and ellipticity are calculated. The extracted fundamental frequency reflection coefficient, harmonic ratio, reflection coefficient bandwidth, reflection phase and polarization ratio, polarization ellipse tilt angle, and ellipticity are combined to form the reflection characteristics.

[0011] In this scheme, the spatial features in step 3 specifically include: For each measurement line, the global maximum point of the denoised electromagnetic field intensity data is found as the candidate leak point projection position. The full width at half maximum, attenuation index and left-right symmetry coefficient are calculated. A sliding window peak search is used to detect multiple local maxima. Geometric consistency optimization is performed on the peak positions of multiple measuring lines. A cost function is constructed and gradient descent optimization is performed to achieve the best match between the peak positions of each measuring line and the projected positions of the seepage points. The intersection location results, location residuals, and the number of measuring lines participating in the intersection are extracted. Calculate the extreme points and zero-crossing points of the first and second spatial gradients at each measuring point; The extracted full width at half maximum (FWHM), attenuation index, symmetry coefficient, local maximum point location, intersection location results, location residual, number of survey lines participating in the intersection, and spatial gradient extreme points and zero-crossing points are combined into spatial features.

[0012] In this scheme, step 4, training the domain adaptive neural network model, specifically includes: The multi-feature dataset is input into the feature extractor. After the adaptive feature recalibration module performs weighted processing on each dimension of the feature extractor, the output is a feature representation containing reflection feature components and spatial feature components. The output feature representation is input into the adaptive fusion layer, and the reflection feature component and the spatial feature component are weighted and fused to output a fused feature vector. The fused feature vector is simultaneously input into the leakage predictor and the neighborhood discriminator. In the leakage predictor, the predicted values ​​of the leakage point location and leakage intensity are calculated and output. In the neighborhood discriminator, after being processed by the gradient inversion layer, the probability distribution of the sample in each detection scene neighborhood category is output. The mean squared error and mean absolute error between the leakage prediction value and the true label are calculated as the leakage prediction loss, and the cross-entropy between the domain probability distribution and the true domain label is calculated as the domain discrimination loss. The two losses are weighted and summed to obtain the total loss. The current accuracy is calculated based on the probability distribution output by the neighborhood discriminator, and the strength parameters of the gradient inversion layer are dynamically adjusted. Perform backpropagation to update the parameters of each sub-network. After three stages of warm-up training, adversarial training, and fine-tuning training, output the trained domain adaptive neural network model.

[0013] In this scheme, the adaptive fusion layer performs weighted fusion of reflection feature components and spatial feature components, specifically including: Calculate the quality statistics of the reflection feature components, which include the signal-to-noise ratio, variance of each dimension, and Mahalanobis distance from the historical average of the reflection feature components. Input the reflection feature component and the spatial feature component, and calculate the signal-to-noise ratio, variance of each dimension and Mahalanobis distance of the reflection feature component and the signal-to-noise ratio, variance of each dimension and goodness of fit of the spherical wave propagation law of the spatial feature component respectively. Input the signal-to-noise ratio, variance of each dimension, and Mahalanobis distance of the historical average of the reflection feature components and spatial feature components into the weight generation module, and output the basic weight coefficients. Receive the weight correction factor of the domain category to which the current sample belongs, multiply the base weight coefficient by the weight correction factor and normalize it, and output the corrected weight coefficient; The corrected weight coefficients are subjected to an exponential moving average with the smoothed weights of the previous time step to output the final fused weights. The reflection feature component and the spatial feature component are multiplied by their respective final fusion weights and then concatenated into vectors to output the fused feature vector.

[0014] In this solution, the construction of the domain tags specifically includes: Soil samples were collected from each testing site and the soil particle size distribution was determined. The sites were classified into clay sites, silt sites, sandy sites, or gravelly sites according to the soil texture classification standard. Obtain the design drawings and construction records of the buried pool to be tested, extract the pool wall material information and pool burial depth information, divide the site into concrete pool site, HDPE membrane pool site or fiberglass pool site according to the pool wall material, and divide the site into shallow buried pool site or deep buried pool site according to the pool burial depth. Soil type, pool wall material type, and pool burial depth type are combined and coded to generate composite domain labels, with each composite domain label corresponding to a detection scenario category.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention uses a low-frequency electromagnetic wave detection method, which eliminates the need to drain sewage from the pool, excavate and damage the ground surface, or drill observation wells. The detection process does not affect the normal production and operation of the enterprise, realizing non-destructive and rapid operation for detecting leakage in buried pools.

[0016] This invention suppresses background field interference and environmental noise through spatiotemporal filtering, and combines the extraction of reflection and spatial features to effectively identify millimeter-level micro-leakage signals with high detection sensitivity and accurate positioning.

[0017] This invention introduces a domain-adaptive neural network. Through adversarial training between the domain discriminator and the feature extractor, it reduces the impact of differences in site conditions such as soil type, pool wall material, and burial depth on the detection results. Only one model is needed to adapt to multiple detection scenarios, and it has strong generalization ability.

[0018] This invention performs quality-driven weighted fusion of reflection and spatial features through an adaptive fusion layer. The fusion weights are dynamically adjusted based on indicators such as signal-to-noise ratio, feature variance, and goodness of fit of spherical waves. This allows the invention to maintain stable detection performance and high robustness even under low signal-to-noise ratio or strong interference conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0020] Figure 1 A flowchart of a method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves is shown. Figure 2 A schematic diagram of leakage in the buried pool is shown; Figure 3 A flowchart illustrating the construction and training of a domain-adaptive neural network model is shown. Figure 4 The flowchart for calculating the fused feature vector in the adaptive fusion layer is shown. Detailed Implementation

[0021] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0023] like Figure 1 As shown, this embodiment provides a method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves, including: Step 1: Place a transmitting antenna in the liquid inside the buried pool to emit low-frequency electromagnetic waves. Collect electromagnetic field intensity signals at each measuring point on the ground surface outside the pool along the direction away from the pool wall, and record the location and time information of the measuring points.

[0024] It should be noted that the low-frequency electromagnetic wave transmitting antenna is immersed below the liquid surface of the buried pool through the inspection port or reserved observation hole at the top of the pool. The operating frequency range of the antenna is 10 Hz to 1 kHz, and the specific frequency is adjusted according to the burial depth of the pool and the electromagnetic noise characteristics of the surrounding environment. The greater the burial depth and the stronger the electromagnetic interference of the site, the lower the frequency should be selected. The transmitting antenna adopts a dipole antenna structure with a figure-eight radiation pattern to ensure that the electromagnetic waves can be radiated uniformly in all directions of the pool. The antenna is deployed at half the depth of the pool below the liquid surface, so that the electromagnetic waves propagate in an approximately spherical wave form. If there are multiple inspection ports in the pool, the transmitting antenna is deployed sequentially at each inspection port for independent detection, covering the entire area of ​​the pool.

[0025] Using the pool wall as a reference, survey lines are laid outwards in a direction perpendicular to the pool wall. The length of the survey lines is determined based on the pool's burial depth, the maximum emission angle of electromagnetic waves in the soil medium, and the estimated leakage impact range. Represented as: ,in For the depth of the pool body, This is the maximum exit angle of the electromagnetic wave in the soil medium. To estimate the extent of leakage, for rectangular pools, main measuring lines are laid out on the outer sides of each of the four pool walls; for circular pools, main measuring lines are laid out radially at 30-degree intervals.

[0026] A non-uniform adaptive encryption strategy was adopted to deploy measurement points, and the boundary positions of each area were adaptively adjusted based on the preliminary on-site scanning results. Preferably, the boundary positions were located at a distance from the pool wall. Within the range, denser measurement points are used, with a spacing between adjacent measurement points. It precisely captures the peak electromagnetic field intensity in the near-field zone of the leak point; at a distance from the pool wall Within the range, conventional point layout is used, with the distance between adjacent measuring points... At a distance from the pool wall Within the range, sparse measurement points are used, with a spacing between adjacent measurement points. This is used to verify the trend of electromagnetic field strength decreasing with distance; the dividing point. and The algorithm is adaptively adjusted based on the initial rapid scan results on site. If a significant peak is detected in the encrypted area, the original scheme is maintained; if the peak appears in the regular or sparse area, the encrypted area is expanded to the regular or sparse area.

[0027] To eliminate background electromagnetic interference, parallel reference survey lines were symmetrically positioned on both sides of the main survey line. During data acquisition, the electromagnetic field strength of both the main survey line and the two reference survey lines was recorded simultaneously. In subsequent data processing, the measured values ​​of the reference survey lines were used to subtract the background field of the site environment. The effective signal at each measuring point on the main survey line = original signal of the main survey line - (signal of the left reference survey line + signal of the right reference survey line) / 2. A low-frequency electromagnetic field receiver was used to collect signals at each measuring point. The axis of the receiver's ferrite rod antenna was set perpendicular to the pool wall to receive the electromagnetic field component perpendicular to the pool wall. Signals were collected at each measuring point for a preset duration, and the arithmetic mean was calculated after multiple collections to eliminate random errors. The spatial coordinates, start and end times of the collection, and power frequency phase information of each measuring point were recorded simultaneously to form raw electromagnetic field data with spatiotemporal tags. In subsequent spatiotemporal filtering, timestamps were used to help identify and remove transient anomalies caused by environmental electromagnetic fluctuations.

[0028] Step 2: Perform spatiotemporal filtering on the collected electromagnetic field data to suppress large-scale background field interference and environmental noise, and obtain the denoised electromagnetic field data.

[0029] It should be noted that a processing architecture combining spatial and temporal filtering is adopted. Spatial filtering is performed first to suppress large-scale background fields, such as the soil background electromagnetic field and slowly varying fields generated by surrounding fixed interference sources. Temporal filtering is then performed to suppress random environmental noise, such as power frequency harmonics, transient pulses generated by equipment start-up and shutdown, and broadband noise generated by lightning activity.

[0030] In spatial filtering, for the electromagnetic field intensity data of each measurement line, a spatial sequence is constructed with the measurement point location as the abscissa and the field intensity as the ordinate. For non-uniform measurement points, interpolation and resampling are used to convert the data into a uniformly spaced sequence before spatial spectrum analysis is performed to prevent spectral leakage caused by directly using Fourier transform. The cumulative energy curve of the spatial spectrum is calculated, and the inflection point where the cumulative energy changes from a rapid increase to a gradual change is found. The wavenumber corresponding to the inflection point is used as the initial cutoff wavenumber. Site empirical coefficients are introduced to correct the initial cutoff wavenumber, thus determining the cutoff wavenumber of the spatial high-pass filter. If there are multiple pool walls on the measurement line, the cutoff wavenumber is calculated separately for each measurement line.

[0031] Due to significant differences in the density of measuring points across different regions, to prevent excessive differences in contributions from different areas, each measuring point is assigned a weight coefficient proportional to the length of the corresponding representative interval. For example, the weight coefficient assigned to each measuring point is proportional to the length of the interval it represents: Let the measuring point... The previous measuring point position is The location of the next measuring point is Then the length of the representative interval of the measuring point is Weight The weighted sum of all measuring points is equal to the total length of the measuring line. In the filtering convolution operation, the weighted sequence is used instead of the original sequence for spatial filtering. After filtering, the weighted sequence is restored to the original measuring point sequence.

[0032] In the spatial filtering process, reference survey line data is introduced for accurate background field subtraction. The electromagnetic field data of the two reference survey lines to the left and right of the main survey line are subjected to the same spatial filtering process to obtain the filtered reference signal. The estimated value of the site background field at the location of the main survey line is calculated. The estimated value of the background field is subtracted from the filtering result of the main survey line to obtain the anomalous signal after background subtraction. Short-time energy is calculated for the data after background subtraction and compared with a statistical threshold. Data intervals identified as transient pulses are replaced with historical averages. The reference survey lines are located at equidistant positions on both sides of the main survey line. Their response to leakage anomaly signals is much smaller than that of the main survey line, but their response to the large-scale background field is approximately equal to that of the main survey line. Therefore, after subtraction, the background field is effectively suppressed, while the anomalous signal is preserved.

[0033] Synchronous detection technology is used for time filtering. The core idea of ​​synchronous detection is to use a reference signal that is in phase and frequency with the transmitted signal to coherently demodulate the received signal and extract the useful components. The reference signal is obtained from the transmitting end, generating an orthogonal reference signal that is in phase and frequency with the transmitted signal. ,in For frequency, As a reference phase; for the received signal Perform coherent demodulation, respectively with and Multiply to obtain the in-phase components. and orthogonal components The amplitude of the detector output is extracted after low-pass filtering. ,in and This represents the DC component after low-pass filtering. Synchronous detection has an extremely narrow equivalent bandwidth, effectively suppressing all interference signals at frequencies different from the transmission frequency, including power frequency and harmonics. When a reference signal is unavailable, an adaptive time bandpass filter is used, setting the passband center frequency and bandwidth based on the power spectral density of the received signal. The passband center frequency is set to the transmission frequency. passband bandwidth Determined based on transmission frequency stability and leakage signal bandwidth. , For the relative stability of the transmission frequency, To mitigate modulation broadening caused by leakage, a notch filter is inserted within the passband to suppress power frequency harmonics falling into the passband.

[0034] For each measuring point, the time-series data is first averaged to obtain representative amplitudes for each point, generating a spatial sequence. Spatial filtering is then applied to remove large-scale background fields. After the first round of spatial filtering, synchronous detection or adaptive bandpass filtering is performed on the time-series data to suppress environmental noise. The time-filtered data is then reorganized into a spatial sequence for a second round of spatial filtering to further extract local spatial anomalies. These two rounds of spatial filtering are cascaded to form a structure of spatial coarse screening and temporal fine filtering. The spatiotemporally filtered data is compared with the original data to calculate the signal-to-noise ratio (SNR) improvement. If the SNR improvement is lower than a preset threshold, the filter parameters are adaptively adjusted, such as reducing the spatial filter cutoff wavenumber or narrowing the temporal filter bandwidth, and the filtering process is repeated until the requirements are met, ultimately obtaining the denoised electromagnetic field data.

[0035] Step 3: Extract reflection and spatial features from the denoised electromagnetic field data, and combine the extracted features with the corresponding domain labels and leakage labels to construct a multi-feature dataset.

[0036] It should be noted that the reflection characteristics reflect the reflection response of the leakage channel to low-frequency electromagnetic waves, and are closely related to the dielectric properties of the leakage point, the size and shape of the leakage channel. A short-time Fourier transform was performed on the denoised time-series data at each measuring point to extract the spectral amplitude at the emission frequency and odd harmonics. The spectral amplitude of the measuring point farthest from the pool wall was used as the reference background. Calculate the reflection coefficient spectrum , For measuring points In frequency The fundamental frequency reflection coefficient is extracted from the reflection coefficient spectrum based on the spectral amplitude at that location. Harmonic ratios (the ratio of the third harmonic to the fundamental frequency, and the ratio of the fifth harmonic to the fundamental frequency), and reflection coefficient bandwidth (defined as the bandwidth from which the reflection coefficient drops to its peak value). The frequency range at times (multiple times) and the reflection phase (the ratio of in-phase to quadrature components obtained through synchronous detection).

[0037] Two orthogonal ferroelectric antennas are set up at the receiving end to collect electromagnetic field components perpendicular to and parallel to the pool wall, respectively, denoted as . At each measuring point, the electromagnetic field intensity of two orthogonal components is simultaneously collected, and the polarizability, polarization ellipse tilt angle, and ellipticity are calculated. The polarizability is expressed as: In the non-seepage zone, due to the isotropic nature of the soil medium, the polarizability ratio is close to 1; above the seepage point, the seepage channel exhibits anisotropy, and the polarizability ratio deviates significantly from 1. The calculated polarization ellipse inclination angle is expressed as... , The phase difference between the two components is given. The extracted fundamental frequency reflection coefficient, harmonic ratio, reflection coefficient bandwidth, reflection phase and polarization ratio, polarization ellipse tilt angle, and ellipticity are combined to form the reflection characteristics.

[0038] Spatial characteristics reflect the spatial distribution of the electromagnetic field generated at the leakage point, particularly the spatial variation of field intensity based on spherical wave propagation theory. For example... Figure 2 As shown, according to the Huygens-Fresnel principle and the laws of spherical wave propagation, the leakage point acts as a secondary sound source (corresponding to a secondary radiation source in the electromagnetic wave scenario). The electromagnetic wave spreads outward as a spherical wave with the leakage point as the origin. The field strength is the greatest at the projection point directly above the leakage point, and the field strength decreases gradually along the direction away from the projection point.

[0039] Find the global maximum point of the denoised electromagnetic field intensity data for each survey line. As candidate leak point projection locations, As candidate leak point projection locations, we examine whether they conform to the laws of spherical wave propagation and calculate the attenuation curve on the left. ,in To calculate the offset distance, calculate the attenuation curve on the right: Ideal spherical wave attenuation should meet the following requirements: , The medium absorption coefficient, This is the distance correction factor. The amplitude coefficient is related to the transmit power and the leak size. The eigenvectors of spherical wave propagation are extracted: full width at half maximum (FWHM), attenuation index, and left-right symmetry coefficient. For the field strength to decrease to The corresponding space width, For peak amplitude, the decay exponent is... and Perform linear fitting, with the slope being the decay exponent and the left-right symmetry coefficient. In several The mean value is calculated. A sliding window peak search is used to detect multiple local maxima. The spherical wave features described above are extracted for each local maximum, and each maximum corresponds to a candidate leakage point.

[0040] For a rectangular pool, there are measuring lines on the outer sides of all four pool walls; for a circular pool, there are multiple radial measuring lines. The actual location of the leak point should satisfy the following condition: on each measuring line, the distance between the projected position of the leak point and the peak position of the measuring line conforms to a geometric relationship. Assume the coordinates of the leak point in the horizontal plane of the pool are... Given the location of the pool wall, the projection location of the leakage point on a survey line is... This represents the perpendicular distance from the leakage point to the measuring line. Peak positions are extracted from the spherical wave characteristics of each measuring line. , Assign survey line numbers. Perform geometric consistency optimization on the peak positions of multiple survey lines and construct a cost function. ,right Gradient descent optimization is performed to achieve the best match between the peak positions of each measuring line and the projected positions of the leakage points. The optimization result is the leakage point location determined by the intersection of multiple measuring lines. The intersection location result, location residual, and number of measuring lines involved in the intersection are extracted. The spatial gradient change of the electromagnetic field intensity around the leakage point is also an important spatial feature. The extreme points and zero-crossing points of the first-order and second-order spatial gradients at each measuring point are calculated. The extracted full width at half maximum (FWHM), attenuation exponent, symmetry coefficient, local maximum point locations, intersection location results, location residuals, number of measuring lines involved in the intersection, and spatial gradient extreme points and zero-crossing points (points where the first-order gradient changes from positive to negative, corresponding to peak positions) are combined to form spatial features.

[0041] Furthermore, the location of the leak point is initially determined using the propagation law of spherical waves. After the domain adaptive neural network outputs the location of the leak point, the result is compared with the result determined by the spherical wave propagation law to achieve cross-validation. Additionally, the initial determination is used as a pseudo-label for model training in semi-supervised learning. For unlabeled detection data, pseudo-labels are generated using the spherical wave law determination result. For each measurement line, if the denoised electromagnetic field intensity data has a unique global maximum point... If multiple separable local maxima exist, the attenuation process from the maximum point to both sides is monotonous and without abnormal fluctuations, the correlation coefficient between the attenuation curve and the theoretical spherical wave attenuation model is greater than a preset threshold, and the peak amplitude exceeds the average value of the measurement line plus twice the standard deviation, then leakage is determined to exist on the outer side of the pool wall corresponding to that measurement line. If peak values ​​are detected on measurement lines of adjacent or opposite pool walls, then spatial intersection is performed based on the peak positions of each measurement line. If the intersection point is located inside the pool or near the pool wall, then the intersection point is determined to be the location of the leakage point. A list of candidate leakage point locations is output, and the leakage intensity indication is determined by inversely calculating the leakage point size based on the peak amplitude and attenuation index.

[0042] Domain labels identify the detection scenario type to which a sample belongs (e.g., soil type, pool depth, pool wall material, etc.) and are used for training the domain discriminator of the domain adaptive neural network. The construction of these domain labels includes: collecting soil samples from each detection site and determining the soil particle size distribution; classifying the site into clay, silt, sand, or gravel sites according to soil texture classification standards; obtaining design drawings and construction records of the buried pools to be tested, extracting pool wall material information and pool depth information; classifying the site into concrete pools, HDPE membrane pools, or fiberglass pools based on pool wall material, and into shallow or deep pools based on pool depth; and combining and encoding the soil type, pool wall material type, and pool depth type to generate composite domain labels, each corresponding to a detection scenario category. Leakage labels are used as ground truth labels for supervised learning, including the location of the leakage point and the leakage intensity level, and can be obtained through confirmatory excavation or tracer verification. The aforementioned reflection and spatial features are integrated and standardized. Mutual information ranking and correlation coefficient analysis are used for feature selection and dimensionality reduction. The processed features are then combined with the corresponding domain labels and leakage labels to form a multi-feature dataset.

[0043] Step 4: Construct a domain-adaptive neural network, including a feature extractor, an adaptive fusion layer, a leakage predictor, and a domain discriminator. Train the network using the multi-feature dataset, wherein the domain discriminator is trained adversarially using domain labels of different detection scenarios as supervision signals.

[0044] It should be noted that, as Figure 3 As shown, the domain-adaptive neural network enables the model to transfer learning across different detection scenarios. Through adversarial competition between the domain discriminator and the feature extractor, it learns leakage feature representations that are insensitive to scene changes. The domain-adaptive neural network consists of a feature extractor, an adaptive fusion layer, a leakage predictor, and a domain discriminator. The feature extractor adopts a three-layer fully connected network structure. An adaptive feature recalibration module is introduced after the first hidden layer. The recalibration weights of each feature dimension are calculated through two fully connected layers and multiplied element-wise with the original features to achieve adaptive weighting of the feature dimensions. The weighted features are input to the second hidden layer (feature bottleneck layer), which is used to output the compressed feature representation. The feature extractor receives a multi-feature dataset as input and outputs a feature representation containing reflection feature components and spatial feature components. The feature representation is transmitted in two paths: one path is input to the adaptive fusion layer, which performs weighted fusion of the reflection and spatial feature components and outputs a fused feature vector to the leakage predictor; the other path is directly input to the domain discriminator. The neighborhood discriminator employs a two-layer fully connected network structure. A gradient inversion layer is placed after the input layer to transmit the input signal identically during forward propagation and to multiply the gradient by a negative constant during backward propagation. ,in To adapt the intensity parameter to the domain, the output layer uses the Softmax activation function, with the number of output neurons equal to the number of domain categories. The output layer displays the probability distribution of samples across different domain categories in each detection scenario. The leakage predictor employs a two-layer fully connected network structure, with two output units for predicting leakage location and one output unit for predicting leakage intensity, forming a multi-task learning structure that outputs the predicted values ​​for leakage location and leakage intensity.

[0045] The mean squared error and mean absolute error (MAE) between the predicted leakage values ​​and the true labels are calculated as the leakage prediction loss, measuring the difference between the leakage predictor output and the true leakage labels. For location prediction (regression task), the mean squared error loss is used, while for intensity prediction (regression task), the mean absolute error loss is used, which is more robust to outliers. The cross-entropy between the neighborhood probability distribution and the true neighborhood labels is calculated as the neighborhood discrimination loss, measuring the difference between the neighborhood discriminator output and the true neighborhood labels. The two losses are weighted and summed to obtain the total loss. The current accuracy is calculated based on the probability distribution output by the neighborhood discriminator, and the intensity parameters of the gradient inversion layer are dynamically adjusted. In the early stages of training, Set to a small value (e.g., 0.1) to allow the feature extractor to prioritize learning leakage detection capabilities. Gradually increase the value as the leakage prediction loss decreases to a preset threshold. Enhance domain obfuscation strength, strength parameter The adjustment is based on the accuracy of the neighborhood discriminator: if the neighborhood discriminator accuracy exceeds a preset high threshold, it indicates that the neighborhood information is too strong, and the intensity parameter should be increased. If the accuracy of the neighborhood discriminator is lower than the preset low threshold, it indicates that the features have been completely confused but leakage information may be lost. In this case, the intensity parameter should be reduced. Strength parameters The adjustment formula is expressed as: , Based on strength, Adjustment coefficient, This represents the current accuracy of the neighborhood discriminator. The optimization objective of the neighborhood discriminator is to minimize the neighborhood discrimination loss. Due to the existence of the gradient reversal layer, the gradients back to the feature extractor are in the opposite direction, causing the feature extractor to update its parameters in the direction of increasing the neighborhood discrimination loss. Ultimately, the feature representation extracted by the feature extractor contains leakage-related information but minimizes the inclusion of neighborhood-related information.

[0046] After three stages—warm-up training, adversarial training, and fine-tuning training—the trained domain-adaptive neural network model is output. Warm-up training only trains the feature extractor, adaptive fusion layer, and leakage predictor, omitting the domain discriminator. Adversarial training introduces the domain discriminator, enables a gradient inversion layer, and performs joint training. Forward propagation calculates the leakage prediction loss and the domain discriminator loss, while backpropagation inverts the gradient of the domain discriminator through the gradient inversion layer before passing it to the feature extractor. The Adam optimizer is then used to update the parameters of each module. Fine-tuning training, after the adversarial training converges, fixes the feature extractor parameters and jointly fine-tunes the adaptive fusion layer and the leakage predictor. By setting up an adversarial training mechanism between the domain discriminator and the feature extractor, the feature extractor outputs domain-invariant feature representations, severing the correlation between features and site background information. This allows the trained model to be directly applied to detection scenarios with different soil types, different burial depths, and different pool wall materials, without requiring retraining for new sites or extensive sample collection, significantly improving the cross-scenario generalization ability and engineering practicality of the buried pool leakage detection method.

[0047] like Figure 4 As shown, the adaptive fusion layer is located between the output of the feature extractor and the input of the leakage predictor. Its input is the original feature vector output by the feature extractor, which is composed of reflection feature sub-vectors and spatial feature sub-vectors. The output of the adaptive fusion layer is a weighted fused feature vector for use by the leakage predictor. The reflection feature component and the spatial feature component are input, and the quality statistics of the reflection feature component are calculated. These quality statistics include the signal-to-noise ratio (SNR), variance of each dimension, and Mahalanobis distance from the historical average. A larger variance indicates a more significant difference in the reflection feature within the current sample, and the Mahalanobis distance from the historical average reflects whether the reflection feature is within a normal distribution range. The quality statistics of the spatial feature component are also calculated, including the SNR, variance of each dimension, and Mahalanobis distance from the historical average. Preferably, the quality statistics of the spatial feature component also include the goodness of fit of the spherical wave propagation law, used to evaluate the degree of agreement between the measured electromagnetic field intensity attenuation curve and the theoretical spherical wave attenuation model, obtaining the ideal model fit value for spherical wave attenuation, and calculating the Pearson correlation coefficient between the measured attenuation sequence and the theoretical model fit value.

[0048] The signal-to-noise ratio, variance of each dimension, and Mahalanobis distance of the historical average values ​​of the reflection and spatial feature components are input into the weight generation module, which outputs the basic weight coefficients. The module queries the domain category of the current sample and retrieves the domain weight correction factor from the domain weight correction table, which is obtained through backpropagation during model training and initialized to 1. The basic weight coefficients are multiplied by the weight correction factor and then normalized to output the corrected weight coefficients. To avoid drastic fluctuations in weights between adjacent samples that could lead to prediction instability, the corrected weight coefficients are exponentially moved averaged with the smoothed weights from the previous time step to output the final fused weights. The reflection and spatial feature components are multiplied by their respective final fused weights and then concatenated to output the fused feature vector.

[0049] Step 5: Input the sample to be detected into the trained network, and perform weighted fusion of reflection features and spatial features through the adaptive fusion layer to obtain the fused feature vector. Input the fused feature vector into the leakage predictor and output the leakage point location and leakage intensity discrimination results.

[0050] It should be noted that the multi-feature data of the sample to be detected is input into the trained feature extractor, which outputs a feature representation containing reflection and spatial feature components. This feature representation is then input into the trained adaptive fusion layer, where the quality statistics of the reflection and spatial feature components are calculated. Fusion weights are generated based on these quality statistics, and the reflection and spatial feature components are weighted and fused to output a fused feature vector. This fused feature vector is then input into the trained leakage predictor, which calculates and outputs the discrimination results for the leakage point location and leakage intensity. This embodiment of the method requires no tracer, no drilling of observation wells, and no on-site data interpretation by specialized geophysical personnel. After model training is completed, only on-site collection of electromagnetic field data is needed to input into the model to obtain the discrimination results, significantly reducing the labor and material costs of a single detection and facilitating its widespread use in numerous buried ponds.

[0051] In another preferred embodiment of the present invention, after the spatiotemporal filtering process in step 2, the large-scale site background field and environmental random noise in the electromagnetic field data have been initially suppressed, but residual background clutter may still affect the identification of weak leakage signals. To improve the detection capability of small leakage points, this embodiment introduces a clutter background modeling and constant false alarm rate target detection method based on generalized extreme value distribution after spatiotemporal filtering and before multi-feature extraction.

[0052] For the denoised electromagnetic field data output in step 2, a reference area without suspected leakage is selected, such as the area of ​​the measuring point farthest from the pool wall or a measuring line segment that has been preliminarily judged to be without abnormalities. The electromagnetic field intensity data of each measuring point in the reference area is extracted as the observation sample of background clutter. The statistical characteristics are modeled using the generalized extreme value distribution, and the optimal location parameters of the generalized extreme value distribution are estimated by fitting the reference area data using the maximum likelihood estimation method. Scale parameters Shape parameters A statistical model describing the background clutter characteristics of the current detection site is obtained. Based on the generalized extreme value distribution background model, a constant false alarm rate (CFAR) detection algorithm is used to set an adaptive detection threshold, given a preset false alarm probability. According to the inverse function of the cumulative distribution function of the generalized extreme value distribution Calculate the detection threshold In areas with fluctuating backgrounds, the threshold is automatically raised to reduce false alarms; in areas with stable backgrounds, the threshold is automatically lowered to improve detection sensitivity. The denoised electromagnetic field strength value at each measuring point is compared with the detection threshold: if the field strength value at a measuring point exceeds the threshold, it is determined that an abnormal electromagnetic response signal generated by a leak exists at that point; if it is below the threshold, it is determined to be background clutter. All abnormal detection points and their corresponding measuring point location information are recorded. An abnormal detection point list is output, including the location coordinates and electromagnetic field strength values ​​of each abnormal point. The above results are used to replace or supplement the original denoised data in step 3, serving as input for subsequent reflection and spatial feature extraction or directly for leak point location analysis based on spherical wave propagation laws.

[0053] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0054] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves, characterized in that, Includes the following steps: Step 1: Place a transmitting antenna in the liquid inside the buried pool to transmit low-frequency electromagnetic waves. Collect electromagnetic field intensity signals at each measuring point on the ground surface outside the pool along the direction away from the pool wall, and record the location and time information of the measuring points. Step 2: Perform spatiotemporal filtering on the collected electromagnetic field data to suppress large-scale background field interference and environmental noise, and obtain the denoised electromagnetic field data. Step 3: Extract reflection features and spatial features from the denoised electromagnetic field data, and combine the extracted features with the corresponding domain labels and leakage labels to construct a multi-feature dataset; Step 4: Construct a domain adaptive neural network, including a feature extractor, an adaptive fusion layer, a leakage predictor, and a domain discriminator. Train the network using the multi-feature dataset, wherein the domain discriminator is trained adversarially using domain labels of different detection scenarios as supervision signals. Step 5: Input the sample to be detected into the trained network, and perform weighted fusion of reflection features and spatial features through the adaptive fusion layer to obtain the fused feature vector. Input the fused feature vector into the leakage predictor and output the leakage point location and leakage intensity discrimination results. The reflection features in step 3 specifically include: Short-time Fourier transform is performed on the denoised time series data of each measurement point to extract the spectral amplitude at the transmission frequency and odd harmonics. The reflection coefficient spectrum is calculated with the spectral amplitude of the measurement point farthest from the pool wall as the reference background. The fundamental frequency reflection coefficient, harmonic ratio, reflection coefficient bandwidth and reflection phase are extracted from the reflection coefficient spectrum. Two orthogonal ferro rod antennas are set up at the receiving end to collect electromagnetic field components perpendicular to the pool wall and parallel to the pool wall, respectively, and the polarization ratio, polarization ellipse tilt angle and ellipticity are calculated. The extracted fundamental frequency reflection coefficient, harmonic ratio, reflection coefficient bandwidth, reflection phase and polarization ratio, polarization ellipse tilt angle, and ellipticity are combined to form the reflection characteristics. The spatial features in step 3 specifically include: For each measurement line, the global maximum point of the denoised electromagnetic field intensity data is found as the candidate leak point projection position. The full width at half maximum, attenuation index and left-right symmetry coefficient are calculated. A sliding window peak search is used to detect multiple local maxima. Geometric consistency optimization is performed on the peak positions of multiple measuring lines. A cost function is constructed and gradient descent optimization is performed to achieve the best match between the peak positions of each measuring line and the projected positions of the seepage points. The intersection location results, location residuals, and the number of measuring lines participating in the intersection are extracted. Calculate the extreme points and zero-crossing points of the first and second spatial gradients at each measuring point; The extracted full width at half maximum (FWHM), attenuation index, symmetry coefficient, local maximum point location, intersection location results, location residual, number of survey lines participating in the intersection, and spatial gradient extreme points and zero-crossing points are combined into spatial features.

2. The method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves according to claim 1, characterized in that, Step 1 specifically includes: The low-frequency electromagnetic wave transmitting antenna is immersed below the liquid surface of the underground pool through the maintenance port at the top of the pool. The antenna is immersed to a depth of half the depth of the pool, and the operating frequency of the antenna is adaptively adjusted according to the burial depth of the pool and the environmental noise characteristics. Using the pool wall as a reference, survey lines are laid out outward in a direction perpendicular to the pool wall. The length of the survey lines is determined based on the burial depth of the pool, the maximum emission angle of electromagnetic waves in the soil medium, and the estimated range of leakage impact. A non-uniform adaptive encryption strategy was adopted to deploy measurement points, and the boundary positions of each area were adaptively adjusted according to the preliminary on-site scanning results. Parallel reference survey lines are set up at symmetrical positions on both sides of the main survey line. The reference survey lines adopt the same point layout strategy as the main survey line to deduct the background electromagnetic field interference of the site. A low-frequency electromagnetic field receiver was used to collect signals at each measuring point. The axis of the receiver's ferrite rod antenna was set to be perpendicular to the pool wall. Signals were collected at each measuring point for a preset duration, and the arithmetic average was taken after repeated collection. The spatial coordinates, start and end times, and power frequency phase information of each measuring point are recorded synchronously to form raw electromagnetic field data with spatiotemporal tags.

3. The method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves according to claim 1, characterized in that, The spatiotemporal filtering in step 2 specifically includes: A spatial sequence of electromagnetic field intensity data for each measurement line is constructed. Non-uniform measurement points are resampled into a uniform spacing sequence through interpolation and then spatial spectrum analysis is performed. The cutoff wavenumber of the spatial high-pass filter is determined based on the cumulative energy curve of the spatial spectrum, and the cutoff wavenumber is corrected by introducing site empirical coefficients. Each measurement point is assigned a weight coefficient proportional to the length of the corresponding representative interval. In the filtering convolution operation, the weighted sequence is used instead of the original sequence for spatial filtering. Subtract the mean of the filtering results of the left and right reference lines from the filtering results of the main line to obtain the data after deducting the reference background. Calculate the short-time energy of the data after deducting the reference background and compare it with the statistical threshold. Replace the data intervals that are determined to be transient pulses with the historical mean. Synchronous detection technology is used for time filtering. A reference signal is obtained from the transmitting end, and an orthogonal reference signal with the same frequency and phase as the transmitted signal is generated. The received signal is coherently demodulated, and the output amplitude of the detector is extracted after low-pass filtering. When the reference signal cannot be obtained, an adaptive time bandpass filter is used. The passband center frequency and bandwidth are set according to the power spectral density of the received signal, and a notch filter is inserted.

4. The method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves according to claim 3, characterized in that, The spatiotemporal filtering cascade strategy further includes: First, the raw data undergoes initial spatial filtering. The time mean of the time domain data at each measurement point is calculated to form a spatial sequence for further filtering. Then, the time series data after the first round of spatial filtering is subjected to time filtering, and the time-filtered data is reorganized into a spatial sequence for the second round of spatial filtering. Calculate the change in signal-to-noise ratio before and after filtering. If the change in signal-to-noise ratio is lower than a preset threshold, adjust the filter parameters and perform filtering again to obtain the denoised electromagnetic field data.

5. The method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves according to claim 1, characterized in that, Step 4, training the domain adaptive neural network model, specifically includes: The multi-feature dataset is input into the feature extractor. After the adaptive feature recalibration module performs weighted processing on each dimension of the feature extractor, the output is a feature representation containing reflection feature components and spatial feature components. The output feature representation is input into the adaptive fusion layer, and the reflection feature component and the spatial feature component are weighted and fused to output a fused feature vector. The fused feature vector is simultaneously input into the leakage predictor and the neighborhood discriminator. In the leakage predictor, the predicted values ​​of the leakage point location and leakage intensity are calculated and output. In the neighborhood discriminator, after being processed by the gradient inversion layer, the probability distribution of the sample in each detection scene neighborhood category is output. The mean squared error and mean absolute error between the leakage prediction value and the true label are calculated as the leakage prediction loss, and the cross-entropy between the domain probability distribution and the true domain label is calculated as the domain discrimination loss. The two losses are weighted and summed to obtain the total loss. The current accuracy is calculated based on the probability distribution output by the neighborhood discriminator, and the strength parameters of the gradient inversion layer are dynamically adjusted. Perform backpropagation to update the parameters of each sub-network. After three stages of warm-up training, adversarial training, and fine-tuning training, output the trained domain adaptive neural network model.

6. The method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves according to claim 5, characterized in that, The adaptive fusion layer performs weighted fusion of reflection feature components and spatial feature components, specifically including: Calculate the quality statistics of the reflection feature components, which include the signal-to-noise ratio, variance of each dimension, and Mahalanobis distance from the historical average of the reflection feature components. Input the reflection feature component and the spatial feature component, and calculate the signal-to-noise ratio, variance of each dimension and Mahalanobis distance of the reflection feature component and the signal-to-noise ratio, variance of each dimension and goodness of fit of the spherical wave propagation law of the spatial feature component respectively. Input the signal-to-noise ratio, variance of each dimension, and Mahalanobis distance of the historical average of the reflection feature components and spatial feature components into the weight generation module, and output the basic weight coefficients. Receive the weight correction factor of the domain category to which the current sample belongs, multiply the base weight coefficient by the weight correction factor and normalize it, and output the corrected weight coefficient; The corrected weight coefficients are subjected to an exponential moving average with the smoothed weights of the previous time step to output the final fused weights. The reflection feature component and the spatial feature component are multiplied by their respective final fusion weights and then concatenated into vectors to output the fused feature vector.

7. The method for detecting hidden leaks in buried pools based on low-frequency electromagnetic waves according to claim 5, characterized in that, The construction of the domain labels specifically includes: Soil samples were collected from each testing site and the soil particle size distribution was determined. The sites were classified into clay sites, silt sites, sandy sites, or gravelly sites according to the soil texture classification standard. Obtain the design drawings and construction records of the buried pool to be tested, extract the pool wall material information and pool burial depth information, divide the site into concrete pool site, HDPE membrane pool site or fiberglass pool site according to the pool wall material, and divide the site into shallow buried pool site or deep buried pool site according to the pool burial depth. Soil type, pool wall material type, and pool burial depth type are combined and coded to generate composite domain labels, with each composite domain label corresponding to a detection scenario category.

Images