Concealed fault positioning method based on radon multi-peak anomaly intelligent identification

Abstract

The application provides a concealed fault positioning method based on radon multi-peak anomaly intelligent identification, which can accurately depict the local background level and natural fluctuation characteristics of radon concentration in different spatial sections, construct a sensitivity function by combining the dimensionless local variation coefficient, and determine the sensitivity function, which can adaptively adapt to the differentiated fluctuation law of the radon background in different regions along the survey line, avoiding the defect that the fixed threshold cannot take into account the background difference of the region; at the same time, through the spatial connected domain analysis of the anomaly points, the automatic merging of continuous anomaly points is realized to form an initial anomaly section, and then combined with the physical length threshold, the scale screening is completed, which can effectively eliminate the scattered small anomaly sections without geological significance, accurately retain the candidate anomaly sections strongly related to the concealed fault structure, on the basis of ensuring the comprehensiveness of anomaly identification, reduce the data amount and invalid calculation overhead of subsequent model operation, and improve the operation efficiency of the overall identification process and the accuracy of concealed fault anomaly identification.

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration technology, and in particular to a method for locating hidden faults based on intelligent identification of radon multi-peak anomalies. Background Technology

[0002] Concealed faults refer to active faults that do not expose at the surface and are covered by loose Quaternary sediments. They are characterized by their sudden onset and the difficulty in surface identification and detection, making them a key focus for urban earthquake disaster risk prevention and control. Traditional detection methods, such as shallow seismic exploration, generally suffer from high costs and low efficiency, and their effectiveness is limited in noisy urban environments. Soil radon, as a naturally occurring radioactive gas, easily migrates upwards along fault fracture zones and accumulates, forming observable geochemical anomalies at the surface. It has become an important indicator for the geochemical detection of concealed faults.

[0003] Existing radon detection methods largely rely on manual experience to interpret overall high values ​​or single-peak anomalies, making it difficult to effectively analyze multi-peak superimposed anomalies formed under complex geological conditions, such as those caused by a main fault plus secondary ruptures or diffusion-convection coupling effects. Current deep learning-based radon anomaly identification methods mostly employ steady-state physical constraints or purely data-driven models, failing to effectively utilize the temporal dynamic characteristics of radon transport. Furthermore, they cannot solve the inverse problem of retrieving underground fault source terms from surface observation concentrations, resulting in an inability to quantitatively evaluate fault activity intensity (radon release flux) and distinguish the independent fault sources corresponding to superimposed peaks. Such multi-peak structures contain crucial information about fault geometry, such as fracture zone width, fault dip, and branching characteristics, but due to the lack of intelligent identification and intelligent inverse source retrieval methods guided by non-steady-state physics, they have long been ignored or misjudged, leading to low positioning accuracy and poor repeatability of detection results, thus hindering the application of radon measurement methods in the detailed exploration of active urban faults. Therefore, there is an urgent need to develop an intelligent method that can automatically identify radon multi-peak anomalies and quantitatively invert hidden fault parameters, and an intelligent method that can quantitatively extract fault source distribution and activity parameters through unsteady-state inverse source inversion, so as to improve the accuracy, efficiency and reliability of hidden fault detection. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for locating hidden faults based on intelligent identification of radon multi-peak anomalies. This method solves the problems existing in current soil radon detection technologies, such as significant environmental interference, difficulty in analyzing multi-peak anomalies, inability to quantitatively invert the distribution and activity flux of radon release sources on faults, and reliance on human experience for inverting hidden fault parameters.

[0005] To achieve the above objectives, the technical solution of this invention is as follows: In a first aspect, the present invention provides a method for locating concealed faults based on intelligent identification of radon multi-peak anomalies, the method comprising: The original soil radon concentration time series data and synchronously collected environmental parameters are obtained from multiple sampling points along a preset survey line in the area to be tested. The original soil radon concentration time series data are preprocessed to obtain a standardized radon concentration time series. Based on the environmental parameters, a multivariate collaborative environmental disturbance correction model is constructed. The standardized radon concentration time series is dynamically corrected using the multivariate collaborative environmental disturbance correction model to obtain the corrected radon concentration time series. The sliding window dynamic threshold method is used to perform anomaly screening on the corrected radon concentration time series. Spatially continuous anomaly points are merged into candidate anomaly segments, and the time series concentration subsequence corresponding to each candidate anomaly segment is extracted. A physical information neural network model is constructed that integrates the physical constraints of unsteady radon diffusion-convection-source term coupling. The time-series concentration subsequences corresponding to each candidate anomaly segment are input into the physical information neural network model, and the concentration probability mask and the intensity distribution of fault radon emission source term are output in parallel. The concentration probability mask is post-processed to extract peak features, and a global peak feature set is obtained. Based on a preset multi-peak-fault mapping rule library, the fault center coordinates are determined by the peak position of the intensity distribution of the fault radon emission source term, the width of the fracture zone is determined by the width of the non-zero region of the source term, the fault radon emission flux is determined by the integrated flux of the source term, and the fault dip is inferred by combining the peak characteristics to generate hidden fault parameters. By combining the hidden fault parameters with the geographic coordinates of the survey line, standardized fault location results, including fault location maps, source term distribution maps, and confidence interval heat maps, are generated and output.

[0006] In some embodiments, the step of constructing a multivariate collaborative environmental interference correction model based on the environmental parameters, and using the multivariate collaborative environmental interference correction model to dynamically correct the standardized radon concentration time series to obtain the corrected radon concentration time series, includes: inputting temperature, humidity, and atmospheric pressure as input covariates into the pre-trained multivariate collaborative environmental interference correction model, and outputting an environmental bias; subtracting the environmental bias from the standardized radon concentration time series to obtain the corrected radon concentration time series; wherein the multivariate collaborative environmental interference correction model is trained by a lightweight multilayer perceptron network, the lightweight multilayer perceptron network including: an input layer with 3 neurons, two hidden layers each with 8 neurons and using the ReLU activation function, and a single-neuron output layer for outputting the prediction bias.

[0007] In some embodiments, the step of using a sliding window dynamic threshold method to perform anomaly screening on the corrected radon concentration time series, merging spatially continuous anomalies into candidate anomaly segments, includes: selecting sampling points within a neighborhood range satisfying a preset physical length along the measurement line direction, centered on each sampling point in the corrected radon concentration time series, to form a local window; calculating the local mean and local standard deviation of the radon concentration within the local window; calculating the local coefficient of variation based on the local mean and the local standard deviation, and determining a sensitivity function based on the local coefficient of variation; comparing the corrected radon concentration value of the current sampling point with a dynamic threshold determined based on the local mean, the local standard deviation, and the sensitivity function; if the corrected radon concentration value is greater than the dynamic threshold, then marking the current sampling point as an anomaly; performing spatial connectivity analysis on all marked anomalies, merging continuous anomalies into independent initial anomaly segments, and filtering the candidate anomaly segments according to a set physical length threshold.

[0008] In some embodiments, the physical information neural network model uses a one-dimensional U-Net as the backbone network, including an encoder, a bottleneck layer, and a decoder, with a concentration prediction head and a source term inversion head set in parallel at the end of the decoder; wherein, the input of the physical information neural network model is a three-dimensional tensor formed by concatenating normalized spatial coordinates, time coordinates, and corrected radon concentration values; the concentration prediction head outputs the concentration probability mask through a Sigmoid activation function, and the source term inversion head outputs the fault radon emission source term intensity distribution through a ReLU activation function; the total loss function of the physical information neural network model is composed of a weighted sum of data fitting loss, unsteady-state physical residual loss, and source term sparse regularization loss; wherein, the data fitting loss is used to measure the difference between the predicted concentration probability mask and the expert-annotated true mask, the unsteady-state physical residual loss is constructed based on the unsteady-state radon diffusion-convection-source term equation, and the source term sparse regularization loss is the L1 norm of the source term intensity distribution output by the source term inversion head.

[0009] In some embodiments, post-processing is performed on the concentration probability mask to extract peak features and obtain a global peak feature set, including: setting a preset threshold for binarization of the concentration probability mask, extracting continuous high-response regions as potential peak regions; for each potential peak region: using the corrected radon concentration value of each sampling point in the potential peak region as a weight, calculating the weighted average of the spatial coordinates as the peak center position; taking the maximum corrected radon concentration value in the potential peak region as the peak height; determining the length of the interval where the corrected radon concentration is greater than or equal to the half-peak height by interpolation as the half-width; calculating the average value of the concentration probability mask in the potential peak region as the identification confidence of the corresponding peak; calculating the ratio of the cube of the third central moment to the standard deviation of the corrected radon concentration value in the potential peak region as the peak shape skewness, used to characterize the peak shape symmetry and assist in inferring the fault tendency; summarizing the peak features extracted from all candidate anomaly segments to form the global peak feature set.

[0010] In some embodiments, the formula for calculating the data fitting loss is: In the formula, The loss value is used to fit the data; The number of sampling points contained in the k-th candidate anomaly segment; Provide the expert-labeled true mask for the i-th sampling point in the k-th candidate anomaly segment; It is the concentration probability mask for the i-th sampling point in the k-th candidate anomaly segment.

[0011] In some embodiments, the formula for calculating the unsteady physical residual loss is: In the formula, This represents the unsteady physical residual loss value. The number of sampling points contained in the k-th candidate anomaly segment; The timing of the time sequence sampling for each sampling point; For spacetime points The physical residual at the location is ideally zero; among which, ; ; ; In the formula, Let be the first partial derivative of concentration with respect to time, representing the rate of change of radon concentration over time; Let be the first partial derivative of concentration with respect to space, representing the rate of change of radon concentration along the measurement line; To predict the radon concentration field, it is obtained by multiplying the concentration probability mask point by point with the input concentration curve; Let be the second partial derivative of concentration with respect to space, representing the degree of diffusion of concentration in space; The diffusion coefficient is denoted as . This refers to the velocity of underground airflow. The intensity distribution of radon gas release source terms in the fault zone.

[0012] In some embodiments, the formula for calculating the source term sparsity regularization loss is: In the formula, The source term sparsity regularization loss value; The number of sampling points contained in the k-th candidate anomaly segment; The intensity of the fault radon gas release source term at the i-th sampling point; This is the absolute value of the source term strength.

[0013] In some embodiments, the preset multi-peak-fault mapping rule base is defined at least as the following mapping relationships: Single-peak mode: containing only one peak with a confidence level not less than a preset threshold, mapped to a simple main fault, with the fault center determined by the position of the source term peak and the width of the non-zero region of the source term determining the width of the fracture zone; Bimodal symmetrical mode: containing two peaks with a spacing within the preset fracture zone width, both peaks having an absolute skewness value less than a preset threshold or the product of the two peak skewness values ​​being greater than zero and approximately equal, mapped to a fault fracture zone with a certain width, and the dip is determined to be symmetrical and without dip; Bimodal asymmetrical mode: containing two peaks with a spacing between them, and both peaks having an absolute skewness value less than a preset threshold or the product of the two peak skewness values ​​being greater than zero and approximately equal, mapped to a fault fracture zone with a certain width, and the dip is determined to be symmetrical and without dip; Peak skewness with opposite signs and absolute values ​​both greater than a preset threshold is mapped to a dipping fault, with the fault dip direction determined by the skewness sign; Three-peak M-type pattern: containing 3 peaks, with the amplitude of the middle peak significantly lower than a certain proportion of the average values ​​of the two side peaks, and the two side peaks being roughly symmetrical, is mapped to a main fault plus secondary ruptures on both sides or a graben structure; Multi-peak dense cluster pattern: containing no less than 4 peaks and the distance between adjacent peaks being less than a preset density threshold, is mapped to a complex strong fracture zone or a fault intersection area; If the conditions of the multi-peak dense cluster pattern and the bi-peak asymmetric pattern are met simultaneously, the bi-peak asymmetric pattern rule is applied preferentially.

[0014] The concealed fault location method based on intelligent identification of radon multi-peak anomalies provided by this invention constructs local windows distributed along the survey line, centered on each sampling point and combined with a preset physical length. Based on these local windows, the local mean and local standard deviation are calculated, accurately characterizing the local background level and natural fluctuation characteristics of radon concentration in different spatial segments. A sensitivity function is constructed and determined using dimensionless local coefficients of variation, which can adaptively adapt to the differentiated fluctuation patterns of radon background in different regions along the survey line, avoiding the shortcomings of fixed thresholds that cannot account for regional background differences. Simultaneously, by conducting spatial connectivity analysis on anomaly points, continuous anomaly points are automatically merged to form initial anomaly segments. Then, scale screening is completed using physical length thresholds, effectively eliminating scattered and minor anomaly segments without geological significance, and accurately retaining candidate anomaly segments strongly correlated with concealed fault structures. While ensuring the comprehensiveness of anomaly identification, this method reduces the amount of data and unnecessary computational overhead in subsequent model calculations, improving the overall efficiency of the identification process and the accuracy of concealed fault anomaly identification. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the method for locating hidden faults based on intelligent identification of radon multi-peak anomalies provided in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] In the following description, references to "some embodiments" refer to a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. Unless otherwise defined, all technical and scientific terms used in the embodiments of the invention have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of the invention pertain. The terminology used in the embodiments of the invention is for the purpose of describing the embodiments of the invention only and is not intended to limit the invention.

[0018] The following describes an exemplary application of the concealed fault location device based on intelligent identification of radon multi-peak anomalies according to embodiments of the present invention. This device can be implemented as a terminal or a server. In one implementation, the device can be implemented as a terminal such as a laptop, tablet, desktop computer, or mobile device. In another implementation, it can also be implemented as a server. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited in this embodiment. The following will illustrate an exemplary application of the hidden fault location device based on radon multi-peak anomaly intelligent identification when it is used as a server.

[0019] This invention provides a method for locating hidden faults based on intelligent identification of radon multi-peak anomalies. (See also...) Figure 1 , Figure 1This is a flowchart illustrating a method for locating hidden faults based on intelligent identification of radon multi-peak anomalies provided in an embodiment of the present invention. Figure 1 The steps shown are explained.

[0020] S110: Obtain the original soil radon concentration time series data and synchronously collected environmental parameters from multiple sampling points along the preset survey line in the area to be tested, and preprocess the original soil radon concentration time series data to obtain a standardized radon concentration time series.

[0021] Here, the area to be tested refers to a specific geographical area where a hidden fault is suspected to exist, and the location of the hidden fault needs to be determined by detecting the radon concentration in the soil. The scope of this area can be pre-determined based on geological survey needs, regional geological background, and other factors, and is usually an area where faults may be distributed or a geologically anomalous area.

[0022] Here, the pre-planned survey line refers to a straight line or broken line that is pre-planned and laid out in the area to be tested, based on the inferred strike of the hidden fault, geological structural features and sampling efficiency requirements, for the purpose of arranging sampling points. The sampling points are evenly or as needed along the survey line. The purpose is to achieve comprehensive detection of the area to be tested through the coverage of the survey line and to capture the spatial distribution characteristics of radon gas concentration anomalies.

[0023] Here, raw soil radon concentration time-series data refers to the collection of unprocessed soil radon concentration data continuously collected at each sampling point using radon detection equipment such as a soil radon meter. These data are arranged in chronological order of collection to form a time-series sequence. This data directly reflects the change in soil radon concentration at the sampling point over time, including various information such as natural fluctuations in radon itself and environmental disturbances.

[0024] Here, environmental parameters refer to environmental factors that affect radon migration and diffusion and detection results, collected simultaneously with the acquisition of raw soil radon concentration time-series data. In practical applications, environmental parameters include, but are not limited to, soil temperature, soil moisture, atmospheric pressure, surface wind speed, atmospheric temperature, and precipitation. In this invention, these environmental parameters can be used to construct an environmental interference correction model to eliminate the interference of environmental factors on radon concentration detection results.

[0025] Here, preprocessing refers to a series of preliminary processing operations performed on the raw soil radon concentration time series data. The purpose is to eliminate abnormal noise, missing values, and systematic errors in the raw data, providing a data foundation for environmental interference correction and anomaly identification. Preprocessing includes, but is not limited to, missing value imputation, outlier removal, data smoothing, normalization, and other processing operations to improve data accuracy.

[0026] For example, in the original radon concentration time series data of a certain sampling point, three data points were missing due to equipment power failure, and were filled by linear interpolation of adjacent data; one data point was significantly outside the normal concentration range, such as more than 5 times the normal maximum value, and was judged as an outlier and removed; the processed data were smoothed by moving average to eliminate random noise, and finally the normalized data before standardization was obtained.

[0027] Here, the standardized radon concentration time series refers to the original soil radon concentration time series data that has been preprocessed to eliminate missing values, outlier noise, and systematic errors, and has been normalized or standardized to make radon concentration data from different sampling points and time periods comparable. This series preserves the trend of radon concentration over time while eliminating the effects of data dimension and baseline differences.

[0028] S120, Based on the environmental parameters, a multivariate collaborative environmental interference correction model is constructed, and the standardized radon concentration time series is dynamically corrected using the multivariate collaborative environmental interference correction model to obtain the corrected radon concentration time series.

[0029] Here, the multivariate collaborative environmental disturbance correction model refers to a mathematical model constructed using multiple synchronously collected environmental parameters as input variables, and employing machine learning methods such as multiple linear regression, random forests, neural networks, or statistical analysis methods to quantify the impact of environmental factors on radon concentration. This model can dynamically correct for environmental disturbances, eliminate spurious anomalies in radon concentration caused by fluctuations in environmental factors, and restore the true changes in radon concentration, i.e., changes caused by geological factors such as hidden faults.

[0030] Here, dynamic correction refers to the process of using a multivariate collaborative environmental disturbance correction model to perform targeted correction on each data point in the standardized radon concentration time series based on changes in environmental parameters at different times.

[0031] Here, the corrected radon concentration time series refers to the time series obtained after the standardized radon concentration time series has been dynamically corrected by a multivariate collaborative environmental disturbance correction model. This time series eliminates environmental disturbances and can truly reflect the background fluctuations of soil radon concentration at the sampling point and the influence of geological factors, such as the release of radon from hidden faults.

[0032] S130, the sliding window dynamic threshold method is used to perform anomaly screening on the corrected radon concentration time series, and spatially continuous anomaly points are merged into candidate anomaly segments, and the time series concentration subsequence corresponding to each candidate anomaly segment is extracted.

[0033] Here, the sliding window dynamic threshold method refers to an algorithm used for initial screening of outliers from a corrected radon concentration time series. It involves setting a sliding time window and dynamically calculating an anomaly threshold based on the radon concentration data within the window. The window is then slid along the time series, checking each data point one by one to see if it exceeds the dynamic threshold; if it does, it is considered an anomaly. This method can adapt to dynamic fluctuations in radon concentration, avoiding missed or false positives caused by fixed thresholds.

[0034] For example, the sliding window size is set to 6 data points (corresponding to 1 hour of data collection), and the sliding step size is 1 data point. For the 6 corrected radon concentration data in each window, the mean μ and standard deviation σ are calculated, and the threshold is set to μ+3σ. When the window slides to a certain position, if the concentration value of the current data point is greater than μ+3σ, the data point is determined to be an outlier. The window is slid in sequence to complete the initial screening of all data points for outliers.

[0035] Here, an anomaly refers to a data point in the corrected radon concentration time series that exceeds the abnormal threshold set by the sliding window dynamic threshold method. The radon concentration of this data point is significantly higher than the background concentration during the same period, which may be caused by the release of radon from a hidden fault or by interference that has not been completely eliminated.

[0036] Here, a candidate anomaly segment refers to a line segment region formed by multiple consecutive anomaly points distributed along a preset survey line in space. It is formed by merging consecutive anomaly points among adjacent sampling points along the survey line direction to form one or more continuous anomaly regions. Each candidate anomaly segment corresponds to a spatial range on the survey line, within which the radon concentration is abnormal.

[0037] For example, if anomalies are detected at sampling points 10-15 along a pre-defined survey line, and these sampling points are continuously distributed along the survey line, then the areas corresponding to sampling points 10-15 are merged into a candidate anomaly segment. The spatial range of this candidate anomaly segment is the area between sampling points 10-15 on the survey line.

[0038] Here, the time-series concentration subsequence refers to the corrected radon concentration time-series of all sampling points corresponding to each candidate anomaly segment. That is, the time-series data of all sampling points belonging to the candidate anomaly segment are extracted from the radon concentration time-series data of the entire survey line to form an independent subsequence, which is then input into the physical information neural network model for analysis.

[0039] S140, construct a physical information neural network model that integrates the physical constraints of unsteady radon diffusion-convection-source term coupling, input the time-series concentration subsequences corresponding to each candidate anomaly segment into the physical information neural network model, and output the concentration probability mask and the intensity distribution of fault radon release source term in parallel; perform post-processing on the concentration probability mask to extract peak features and obtain a global peak feature set.

[0040] Here, the unsteady-state radon diffusion-convection-source term coupling physical constraint refers to the physical laws constructed based on the physical mechanism of radon migration and diffusion in soil media, used to constrain the training of the physical information neural network model. It includes three core coupling processes: radon diffusion: the molecular diffusion movement of radon gas in soil pores due to concentration differences; radon convection: the directional flow of radon gas along with airflow in soil pores (such as airflow caused by changes in atmospheric pressure or groundwater flow); and source term release: the process by which concealed fault fracture zones act as radon release sources, releasing radon gas into the surrounding soil. These three processes are mutually coupled and jointly determine the spatiotemporal distribution of radon concentration. This physical constraint ensures that the output of the neural network model conforms to the actual physical laws of radon migration, preventing the model output from deviating from the actual geological scenario.

[0041] Here, the physical information neural network model refers to a neural network model that embeds the physical constraints of the unsteady radon diffusion-convection-source term coupling into the loss function of the neural network, realizing a combination of data-driven and physical mechanism-driven neural network models. This model can not only learn from the data features of time-series concentration subsequences, but also ensure, under the constraints of physical constraints, that the output concentration probability mask and source term intensity distribution conform to the actual physical laws of radon migration, improving the accuracy and reliability of the model output and avoiding the problems of overfitting or outputs that do not conform to physical common sense that may occur in purely data-driven models.

[0042] Here, the concentration probability mask refers to a two-dimensional mask matrix output by the physical information neural network model after analyzing the time-series concentration subsequence. The dimension is time × space, with the space corresponding to the sampling points of candidate anomaly segments. The value of each element in the matrix represents the probability that the radon concentration anomaly at that moment and sampling point is caused by a fault. The closer the probability is to 1, the more likely the radon concentration anomaly at that moment and sampling point is caused by radon release from a hidden fault; the closer it is to 0, the more likely it is residual interference or natural fluctuation.

[0043] Here, the intensity distribution of fault radon release source terms refers to the spatial distribution data output by the physical information neural network model, which describes the intensity of radon release from the concealed fault fracture zone to the surrounding soil. Its dimension corresponds to the spatial range of the candidate anomaly segment. The value corresponding to each sampling point represents the intensity of radon release from the fault at that location. This distribution data can directly reflect the spatial location and release capacity of the fault fracture zone.

[0044] Here, post-processing includes, but is not limited to, probability threshold filtering, peak detection, and peak shape feature extraction.

[0045] Here, the global peak feature set refers to the set formed by integrating all peak features extracted after analyzing and post-processing the time-series concentration subsequences corresponding to all candidate anomaly segments through a physical information neural network model.

[0046] S150, based on the preset multi-peak-fault mapping rule library, the fault center coordinates are determined by the peak position of the intensity distribution of the fault radon gas release source term, the width of the fracture zone is determined by the width of the non-zero region of the source term, the fault radon gas release flux is determined by the source term integral flux, and the fault dip is inferred by combining the peak characteristics to generate hidden fault parameters.

[0047] Here, the multi-peak-fault mapping rule base refers to a pre-defined set of rules used to establish a correspondence between the global peak feature set, the intensity distribution of fault radon emission source terms, and hidden fault parameters. This rule base is constructed based on a large amount of geological measurement data and radon migration physical simulation results. It includes a quantitative mapping relationship between peak features, source term intensity distribution, and fault parameters, and can directly infer the corresponding fault parameters based on the input peak features and source term distribution.

[0048] Here, the fault center coordinates refer to the geographic coordinates corresponding to the center of the concealed fault's projection on the Earth's surface. These coordinates are determined by the peak position of the radon emission source term intensity distribution, i.e., the geographic coordinates of the sampling point where the peak value in the source term intensity distribution is located. The fracture zone width refers to the lateral width of the concealed fault fracture zone projected onto the Earth's surface. This width is determined by the width of the non-zero region in the radon emission source term intensity distribution, i.e., the length of the survey line corresponding to the region where the source term intensity value is greater than 0.

[0049] Here, fault radon release flux refers to the total amount of radon released from a concealed fault fracture zone to the surrounding soil per unit time and per unit area. It is determined by the source term integral flux of the source term intensity distribution of fault radon release, that is, by integrating the source term intensity distribution over the entire non-zero region, and the integral value obtained is the fault radon release flux, which is used to reflect the strength of the fault's ability to release radon.

[0050] Here, the fault dip refers to the tilt direction of the concealed fault, that is, the direction perpendicular to the intersection of the fault plane and the horizontal plane, which is inferred from the peak shape features in the global peak feature set.

[0051] Here, the parameters of a concealed fault refer to a set of parameters used to fully describe the structural features and radon release characteristics of a concealed fault, which may include the fault center coordinates, the width of the fracture zone, the radon release flux of the fault, and the fault dip.

[0052] S160, combine the hidden fault parameters with the geographic coordinate information of the survey line to generate and output standardized fault location results including fault location map, source term distribution map and confidence interval heat map.

[0053] Here, a fault location map refers to a graphic drawn based on the geographic coordinates of the survey lines and fault parameters, which is used to intuitively show the spatial location of a hidden fault in the area to be measured. The map clearly marks the fault center location, the range of the fracture zone, the distribution of survey lines, the location of sampling points, and other information, and is an intuitive presentation of the fault location results.

[0054] Here, the source term distribution map refers to a graphic drawn based on the geographic coordinate information of the survey line and the intensity distribution of radon release source terms in the fault, which is used to show the spatial distribution of radon release intensity in the concealed fault fracture zone. It is usually presented in the form of a heat map. The darker the color, the greater the source term intensity, which can intuitively reflect the distribution characteristics of radon release capacity in the fault fracture zone.

[0055] Here, the confidence interval heatmap refers to a graph drawn based on the output of a physical information neural network model to show the confidence level of the inferred results of fault parameters (such as fault center coordinates and fracture zone width), presented in the form of a heatmap.

[0056] Here, standardized fault location results refer to complete fault location results generated by integrating fault parameters, fault location maps, source term distribution maps, and confidence interval heat maps, and in accordance with a preset standard format (such as a format that meets the requirements of a geological exploration report).

[0057] This invention uses a multivariate collaborative environmental interference correction model to dynamically correct time-series data of soil radon concentration. This effectively eliminates spurious anomalies in radon concentration caused by multiple environmental factors such as temperature, humidity, and air pressure, reduces positioning errors caused by environmental interference, and restores the true radon release characteristics of the formation. This solves the problems of traditional radon detection methods being susceptible to fluctuations in the field environment and having poor reliability in anomaly identification. Simultaneously, a sliding window dynamic threshold is used to perform initial anomaly screening and merge candidate anomaly segments. A physical information neural network integrating the physical constraints of unsteady radon diffusion-convection-source term coupling is used for parallel inversion. This approach balances the advantages of data-driven intelligent identification with the true physical mechanism of radon migration. It also outputs a concentration probability mask and the intensity distribution of fault radon release source terms, accurately extracting multi-peak anomaly features. This overcomes the shortcomings of traditional single-peak identification, which easily misses hidden faults and cannot invert source term parameters, achieving accurate mapping between multi-peak anomalies and hidden fault structures. Furthermore, based on the multi-peak-fault mapping rule base, complete fault parameters such as fault center coordinates, fracture zone width, radon release flux, and fault dip can be quantitatively inverted, enabling multi-dimensional comprehensive interpretation of the location, development scale, release intensity, and structural attitude of concealed faults; ultimately, fault parameters, fault location maps, source term distribution maps, and confidence interval thermodynamics can be integrated. Figure 1 The data is integrated and standardized in accordance with the geological exploration industry reporting standards to form standardized fault location results. The results are complete, the illustrations are intuitive, the data is standardized, and the credibility is quantifiable. They can be directly applied to engineering scenarios such as field geological exploration, geological disaster assessment, and deep structural detection without the need for secondary data processing and format conversion, which significantly improves the practicality and engineering reusability of the hidden fault detection results.

[0058] In some embodiments, the above-described S120 can also be implemented by the following: Temperature, humidity, and atmospheric pressure are input as covariates into a pre-trained multivariate collaborative environmental disturbance correction model, which outputs an environmental bias. The environmental bias is then subtracted from the standardized radon concentration time series to obtain the corrected radon concentration time series. The multivariate collaborative environmental disturbance correction model is trained using a lightweight multilayer perceptron network, which includes an input layer with 3 neurons, two hidden layers each with 8 neurons using the ReLU activation function, and a single-neuron output layer for outputting the prediction bias.

[0059] This embodiment explicitly selects temperature, humidity, and atmospheric pressure as fixed input covariates, effectively avoiding redundant parameters that introduce additional noise and significantly improving the pertinence and overall computational efficiency of environmental interference correction. At the same time, it adopts a direct correction method that combines model prediction of environmental deviation with difference subtraction correction. The calculation logic is simple, clear, and highly interpretable. It can dynamically deduct environmental interference components point by point from time series data, accurately remove spurious fluctuations in radon concentration caused by environmental changes, and effectively improve the data authenticity and reliability of the corrected radon concentration time series. Furthermore, this embodiment employs a well-defined, lightweight multilayer perceptron network to construct the environmental interference correction model. It precisely limits the network layers, the number of neurons, and the ReLU activation function, resulting in a streamlined overall network structure with fewer parameters and lower computational load. This allows for computation without relying on high-performance computing hardware, making it suitable for real-time solutions from portable field detection devices and edge terminals, effectively reducing the difficulty and cost of field deployment. Moreover, the standardized network architecture, with unified constraints on three-dimensional input, two hidden layers, and single-dimensional output, ensures a regular and consistent model structure, reducing the difficulty of model training and replication, and providing excellent scene generalization capabilities. This makes it adaptable to environmental interference correction operations under different geographical and geological conditions.

[0060] In some embodiments, the above-described S130 can also be implemented by the following: Taking each sampling point in the corrected radon concentration time series as the center, sampling points within a neighborhood range satisfying a preset physical length are selected along the measurement line direction to form a local window; the local mean and local standard deviation of radon concentration within the local window are calculated; the local coefficient of variation is calculated based on the local mean and the local standard deviation, and a sensitivity function is determined based on the local coefficient of variation; the corrected radon concentration value of the current sampling point is compared with a dynamic threshold determined based on the local mean, the local standard deviation, and the sensitivity function; if the corrected radon concentration value is greater than the dynamic threshold, the current sampling point is marked as an anomaly; spatial connectivity analysis is performed on all marked anomalies, consecutive anomalies are merged into independent initial anomaly segments, and candidate anomaly segments are selected according to the set physical length threshold.

[0061] Here, a local window refers to a set of neighboring sampling points defined by a preset physical length along the spatial extension direction of a preset measurement line, centered on a single sampling point in the corrected radon concentration time series, providing a limited spatial range for calculating local statistical features; the local window has spatial locality, only representing the radon concentration distribution characteristics within a limited range around the current point.

[0062] Here, the preset physical length refers to a fixed spatial distance threshold set in advance based on the accuracy of geological exploration, the spatial scale of radon diffusion, and the sampling interval of the survey line. It is used to constrain the spatial coverage of the local window and ensure that the window scale conforms to the actual spatial development scale of soil radon anomalies.

[0063] Here, the local mean specifically refers to the arithmetic mean of the corrected radon concentration values ​​corresponding to all sampling points within a single local window. It is used to characterize the background level of radon concentration within the window area and reflect the overall steady-state characteristics of radon concentration in the local area.

[0064] Here, local standard deviation refers to the statistical measure of the dispersion of radon concentration values ​​at each sampling point within a local window relative to the local mean. It is used to quantify the natural fluctuation range of radon concentration in a local area and to reflect the fluctuation differences in normal background data.

[0065] Here, the sensitivity function is a dynamic adjustment function obtained by quantifying and fitting the local coefficient of variation, which is used to adaptively adjust the sensitivity of anomaly detection; the greater the relative fluctuation of radon concentration, the corresponding adjustment coefficient of the sensitivity function changes accordingly to adapt to the background fluctuation differences in different sections.

[0066] Here, spatial connectivity analysis refers to the analysis method that identifies the continuity and adjacency of discretely marked outliers along the spatial dimension of a preset survey line. It is used to determine whether outliers are continuously distributed in space and to distinguish between isolated discrete outliers and continuous outlier areas.

[0067] Here, the initial anomaly segment refers to the continuous anomaly segment formed by merging multiple spatially adjacent anomaly points after spatial connectivity analysis. It is the original anomaly interval that has not undergone scale screening and includes the entire spatially continuous radon anomaly region.

[0068] Here, the physical length threshold refers to the spatial length screening standard set in advance, which combines the minimum development scale of the concealed fault fracture zone and the requirement of geological exploration to eliminate invalid micro-anomalies. It is used to filter short-distance scattered anomaly segments that have no geological significance.

[0069] This embodiment constructs local windows along the survey line, centered on each sampling point and combined with a preset physical length. Based on these local windows, the local mean and local standard deviation are calculated, which can accurately characterize the local background level and natural fluctuation characteristics of radon concentration in different spatial segments. By combining the dimensionless local coefficient of variation to construct and determine the sensitivity function, it can adaptively adapt to the differentiated fluctuation patterns of radon background in different areas along the survey line, avoiding the shortcomings of fixed thresholds that cannot take into account regional background differences. At the same time, by conducting spatial connectivity analysis on anomaly points, continuous anomaly points are automatically merged to form initial anomaly segments. Then, by combining the physical length threshold to complete scale screening, scattered and small anomaly segments without geological significance can be effectively eliminated, and candidate anomaly segments strongly correlated with concealed fault structures can be accurately retained. While ensuring the comprehensiveness of anomaly identification, the amount of data and invalid computational overhead of subsequent model calculations are reduced, improving the overall efficiency of the identification process and the accuracy of concealed fault anomaly identification.

[0070] In some embodiments, the physical information neural network model uses a one-dimensional U-Net as the backbone network, including an encoder, a bottleneck layer, and a decoder, with a concentration prediction head and a source term inversion head set in parallel at the end of the decoder; wherein, the input of the physical information neural network model is a three-dimensional tensor formed by concatenating normalized spatial coordinates, time coordinates, and corrected radon concentration values; the concentration prediction head outputs the concentration probability mask through a Sigmoid activation function, and the source term inversion head outputs the fault radon emission source term intensity distribution through a ReLU activation function; the total loss function of the physical information neural network model is composed of a weighted sum of data fitting loss, unsteady-state physical residual loss, and source term sparse regularization loss; wherein, the data fitting loss is used to measure the difference between the predicted concentration probability mask and the expert-annotated true mask, the unsteady-state physical residual loss is constructed based on the unsteady-state radon diffusion-convection-source term equation, and the source term sparse regularization loss is the L1 norm of the source term intensity distribution output by the source term inversion head.

[0071] This embodiment employs a one-dimensional U-Net as the backbone network of the physical information neural network. Utilizing a symmetrical encoding / decoding structure consisting of an encoder, bottleneck layer, and decoder, it efficiently mines multi-scale spatiotemporal features of radon data distributed along the survey line, accurately preserving spatial details and temporal variation patterns of the linear survey line. A concentration prediction head and a source term inversion head are set up in parallel at the decoder end. Based on the same backbone features, multiple tasks can be simultaneously output, including anomaly probability identification and quantitative inversion of radon source terms, effectively improving model computational efficiency and feature reuse rate. For different task output characteristics, Sigmoid activation functions and ReLU activation functions are used respectively to ensure the probabilistically reasonable output of the concentration probability mask and the stable regression of physical quantities of the intensity distribution of fault radon release source terms.

[0072] Meanwhile, the model's total loss function integrates data fitting loss, unsteady physical residual loss, and source term sparsity regularization loss and applies weighted constraints: the data fitting loss ensures consistency between the prediction results and the actual annotations; the unsteady physical residual loss introduces radon migration coupling physical constraints to prevent the model output from deviating from the actual geological and physical mechanisms; and the source term sparsity regularization loss based on the L1 norm fits the geological prior characteristics of local radon release from concealed faults, effectively suppressing global false source term interference, reducing invalid anomaly outputs, and thus improving the accuracy of radon multi-peak anomaly identification and the reliability of source term inversion.

[0073] In some embodiments, the post-processing of the concentration probability mask to extract peak features and obtain a global peak feature set includes: binarizing the concentration probability mask by setting a preset threshold, and extracting continuous high-response regions as potential peak regions; for each potential peak region: using the corrected radon concentration value of each sampling point in the potential peak region as a weight, calculating the weighted average of the spatial coordinates as the peak center position; taking the maximum corrected radon concentration value in the potential peak region as the peak height; determining the length of the interval where the corrected radon concentration is greater than or equal to the half-peak height by interpolation as the half-width; calculating the average value of the concentration probability mask in the potential peak region as the identification confidence of the corresponding peak; calculating the ratio of the cube of the third central moment to the standard deviation of the corrected radon concentration value in the potential peak region as the peak shape skewness, used to characterize the peak shape symmetry and assist in inferring the fault tendency; and summarizing the peak features extracted from all candidate anomaly segments to form the global peak feature set.

[0074] In some embodiments, the formula for calculating the data fitting loss is: In the formula, To fit the loss value to the data; The number of sampling points contained in the k-th candidate anomaly segment; Provide the expert-labeled true mask for the i-th sampling point in the k-th candidate anomaly segment; It is the concentration probability mask for the i-th sampling point in the k-th candidate anomaly segment.

[0075] In some embodiments, the formula for calculating the unsteady physical residual loss is: In the formula, This represents the unsteady physical residual loss value. The number of sampling points contained in the k-th candidate anomaly segment; The timing of the time sequence sampling for each sampling point; For spacetime points The physical residual at the location is ideally zero; among which, ; ; ; In the formula, Let be the first partial derivative of concentration with respect to time, representing the rate of change of radon concentration over time; Let be the first partial derivative of concentration with respect to space, representing the rate of change of radon concentration along the measurement line; To predict the radon concentration field, it is obtained by multiplying the concentration probability mask point by point with the input concentration curve; Let be the second partial derivative of concentration with respect to space, representing the degree of diffusion of concentration in space; The diffusion coefficient is denoted as . This refers to the velocity of underground airflow. The intensity distribution of radon gas release source terms in the fault zone.

[0076] In some embodiments, the formula for calculating the source term sparsity regularization loss is: In the formula, The source term sparsity regularization loss value; The number of sampling points contained in the k-th candidate anomaly segment; The intensity of the fault radon gas release source term at the i-th sampling point; This is the absolute value of the source term strength.

[0077] In some embodiments, the preset multi-peak-fault mapping rule base is defined at least as the following mapping relationships: Single-peak mode: containing only one peak with a confidence level not less than a preset threshold, mapped to a simple main fault, with the fault center determined by the position of the source term peak and the width of the non-zero region of the source term determining the width of the fracture zone; Bimodal symmetrical mode: containing two peaks with a spacing within the preset fracture zone width, both peaks having an absolute skewness value less than a preset threshold or the product of the two peak skewness values ​​being greater than zero and approximately equal, mapped to a fault fracture zone with a certain width, and the dip is determined to be symmetrical and without dip; Bimodal asymmetrical mode: containing two peaks with a spacing between them, and both peaks having an absolute skewness value less than a preset threshold or the product of the two peak skewness values ​​being greater than zero and approximately equal, mapped to a fault fracture zone with a certain width, and the dip is determined to be symmetrical and without dip; Peak skewness with opposite signs and absolute values ​​both greater than a preset threshold is mapped to a dipping fault, with the fault dip direction determined by the skewness sign; Three-peak M-type pattern: containing 3 peaks, with the amplitude of the middle peak significantly lower than a certain proportion of the average values ​​of the two side peaks, and the two side peaks being roughly symmetrical, is mapped to a main fault plus secondary ruptures on both sides or a graben structure; Multi-peak dense cluster pattern: containing no less than 4 peaks and the distance between adjacent peaks being less than a preset density threshold, is mapped to a complex strong fracture zone or a fault intersection area; If the conditions of the multi-peak dense cluster pattern and the bi-peak asymmetric pattern are met simultaneously, the bi-peak asymmetric pattern rule is applied preferentially.

[0078] The following will describe an exemplary application of the embodiments of the present invention in a practical application scenario.

[0079] This embodiment addresses the problems in existing soil radon detection technologies, such as significant environmental interference, difficulty in analyzing multi-peak anomalies, inability to quantitatively invert the distribution and activity flux of radon release sources along faults, and reliance on manual experience for fault parameter inversion. It proposes a method for intelligent localization and source term inversion of hidden faults using a non-steady-state physical information neural network. This embodiment eliminates systematic biases caused by meteorological factors by constructing a multivariate collaborative environmental interference correction model. Simultaneously, it introduces a deep neural network that integrates the non-steady-state radon diffusion-convection-source term coupling equations to construct a learnable fault source term inversion mechanism. This enables the solution of the inverse problem of inverting the distribution of radon release sources in fault fracture zones from time-series radon concentration data. Furthermore, by combining the source term intensity distribution and a multi-peak-fault mapping rule base with a graph neural network, multi-peak features are transformed into fault geometric parameters and activity fluxes, ultimately generating interpretable, verifiable, and standardized fault localization results.

[0080] This embodiment provides another method for locating hidden faults based on intelligent identification of radon multi-peak anomalies, which includes the following: Step 1: Spatially continuous high-density soil radon concentration collection and data preprocessing.

[0081] Sampling points were set up along a pre-defined survey line at 1-meter intervals. An alpha spectrometer was used to continuously sample at multiple time points every 30 minutes to obtain raw soil radon concentration time-series data, forming a concentration sequence over time. GPS coordinates and environmental parameters such as temperature, humidity, and atmospheric pressure were recorded simultaneously at each sampling point. Subsequently, the raw soil radon concentration time-series data for each sampling point underwent preprocessing including time-domain denoising, coordinate alignment, and normalization to eliminate random noise and measurement bias, constructing a standardized radon concentration time-series sequence.

[0082] First, soil gas sampling points are laid out along the pre-defined survey line in the target detection area at intervals no greater than 1 meter to ensure that the spatial sampling density meets the resolution requirements for multi-peak anomalies. At each sampling point, a dedicated probe is inserted 20±2 cm below the ground surface using high-precision... The particle spectrometer performs continuous sampling at a frequency of 30 minutes per sampling point, continuously collecting data at N time points (e.g., N=6, corresponding to 3 hours) to obtain raw soil radon concentration time-series data. Simultaneously, the location and environmental parameters of each sampling point were collected, and the planar coordinates of the sampling points were recorded using RTK-GPS. A digital temperature and humidity sensor was used to collect ambient temperature T (°C) and ambient humidity H (%), and a MEMS digital barometric pressure sensor was used to collect atmospheric pressure P (hPa).

[0083] Then, the time-series data of the original soil radon concentration at each sampling point were analyzed. Temporal denoising was performed using a sliding window mid-range filtering method. The window length was set to 30 seconds. The multi-time-time sampling sequence for each point was smoothed, and its temporal mean was extracted as the normalized radon concentration Rn for that point after time-domain smoothing. (Plane coordinates) Perform coordinate alignment and spatial calibration operations to obtain the planar coordinates of all sampling points recorded by RTK-GPS. Compared with the cumulative distance, the sampling point positions on the survey line were fine-tuned using piecewise linear interpolation to ensure that the spatial distribution maintained a strict spacing of 1m. Environmental parameters needed to be normalized, with the collected ambient temperature T (°C), ambient humidity H (%), and atmospheric pressure P (hPa) normalized to the [0, 1] interval.

[0084] The final output is a standardized time series of radon concentration. Plane coordinates And ambient temperature T, ambient humidity H and atmospheric pressure P.

[0085] Step 2: Correct the original soil radon concentration using a multivariate collaborative environmental disturbance correction model.

[0086] A multivariate collaborative environmental disturbance correction model was constructed, and the standardized radon concentration time series was dynamically corrected using environmental parameters such as ambient temperature, ambient humidity, and atmospheric pressure to obtain the corrected radon concentration time series.

[0087] The process of constructing a multivariate collaborative environmental disturbance correction model is as follows: 500 independent samples were collected and divided into training, validation and test sets in a ratio of 7:2:1.

[0088] The long-term time-series average of radon concentration in the area to be tested Consider it as a background value. ( (where is the measured original soil radon concentration for the j-th sample). For each sample j, calculate its deviation relative to the background value as a monitoring label. ;in, The deviation of radon concentration in sample j caused by environmental factors.

[0089] Lightweight multilayer perceptron (MLP) is used as In its implementation, the MLP network has four layers: an input layer with three neurons receiving T, H, and P signals respectively; hidden layer 1 with eight neurons using ReLU activation; hidden layer 2 with eight neurons using ReLU activation; and an output layer with one neuron representing the prediction bias. .

[0090] The goal is to optimize the model parameters by minimizing the mean squared error between the predicted and actual deviations. The formula is as follows: ; in, This represents the set of all weights and bias parameters to be optimized in the MLP neural network; This represents the environmental bias predicted by the MLP model for the j-th sample. This represents the actual environmental deviation of the j-th sample. in, During training, the Adam optimizer was used with a learning rate of 0.001 and 200 training epochs. An early stopping strategy was employed to prevent overfitting. These represent the ambient temperature, ambient humidity, and atmospheric pressure of sample j, respectively.

[0091] Evaluate model performance on the retained validation subset, requiring the coefficient of determination. If the target is met, freeze all optimized network parameters. and the training completed It serves as a universal calibration module for data processing of all subsequent target survey lines.

[0092] In actual concealed fault detection, the (T, H, P) values ​​of any sampling point on the target survey line are input into a pre-trained... Output environmental deviation The standardized radon concentration time series was obtained according to the following formula. : ; In the formula, The environmental deviation is a one-dimensional data point of length N, corresponding to the correction value of all N sampling points on the current survey line.

[0093] Step 3: Use dynamic thresholds to perform initial screening for anomalies and extract candidate segments.

[0094] A window-adaptive dynamic threshold mechanism is employed to identify potential anomaly regions, narrowing the scope of subsequent intelligent analysis and improving overall computational efficiency. The input is the corrected radon concentration time series. and one-dimensional spatial coordinate sequence ,in This represents the time-series soil radon concentration at the i-th sampling point after environmental interference correction. This represents the cumulative distance along the survey line from each sampling point, satisfying the following condition: The filtering and extraction steps are as follows: Set the physical window length to 10m, and dynamically determine the neighborhood range for each sampling point i. Find all samples that satisfy... Point j forms a local window. .

[0095] For each local window Calculate its local mean and local standard deviation using the following formula: ; ; ; in, Let be the average radon concentration within a local window centered on the i-th sampling point. For local windows Standard deviation of internal radon concentration; Indicates a local window Summing is performed on all sample point indices j; This represents the corrected time-series radon concentration at the j-th sampling point.

[0096] Introducing local coefficient of variation Dynamically adjust sensitivity The formula is as follows: ; In the formula, Let be the coefficient of variation for the i-th sampling point; The threshold for anomaly detection at the i-th sampling point; This is a sensitivity adjustment function, and its variable is the local coefficient of variation. The announcement is as follows: Among them, the sensitivity function The formula is as follows: ; like If i is an outlier, then the marked point i is an outlier.

[0097] Connectivity analysis is performed on all outlier masks to merge spatially consecutive outliers into independent segments. The physical length of each segment is then calculated using the x-coordinate. The calculation formula is as follows: ; in, Let be the coordinates of the end position of the k-th candidate segment. Let these be the coordinates of the starting position of the k-th candidate segment. Initially, retain... If the segment, or If it is, then it will be directly removed.

[0098] Furthermore, to eliminate isolated noise or along-the-flow interference, only length is retained. The output is a structured list, where each element includes the segment number. Spatial start and end coordinates and Segment length The index range in the original sequence (starting index) and end index ), and the corresponding corrected radon concentration time series subsequence .

[0099] Step 4: Multi-peak intelligent recognition and feature extraction that integrates physical constraints.

[0100] We employ the training concept of a physical information neural network that integrates the physical mechanisms of unsteady radon diffusion-convection-source term coupling. We construct a deep network containing a learnable fault source term mapping head to solve the inverse problem of retrieving fault source distribution from time-series concentration data, and complete the intelligent identification and structured feature extraction of multi-peak structures.

[0101] This step takes the candidate anomaly segments output from step 3 as input. For each corrected radon concentration time-series subsequence, a physical information neural network training approach that integrates the coupled physical mechanisms of unsteady radon diffusion-convection-source terms is adopted. A deep network containing a learnable fault source term mapping head is constructed to solve the inverse problem of retrieving the radon release source distribution in the fault fracture zone from the time-series concentration data. It also completes the intelligent identification and structured feature extraction of multi-peak structures. The specific implementation process is as follows: For the k-th candidate segment (k=1,2,...,K), its input is a spatial coordinate subsequence. and the corrected radon concentration time series subsequence ,in Let x be the number of sampling points contained in the k-th candidate anomaly segment, and r be the number of sampling points in m. Spatial coordinates Normalized to the interval [0, 1], denoted as Normalize the time coordinate t to the interval [0, 1], denoted as .Will , and spliced ​​into a three-dimensional tensor .

[0102] A one-dimensional U-Net is used as the backbone network, with two output heads (concentration prediction head and source term inversion head) connected in parallel at the decoder end. The overall network structure is an encoder-decoder architecture: the input is a three-dimensional tensor. The encoder consists of four downsampling blocks, each containing a 1D convolutional layer, batch normalization, and a ReLU activation function, and progressively reduces the sequence length through max pooling layers. The bottleneck layer is a single-layer 1D convolution used to fuse global context information. The decoder consists of four upsampling blocks, symmetrical to the encoder, and uses transposed convolutions to achieve upsampling, progressively restoring the sequence length, and finally outputting a single-channel density probability mask in parallel. Intensity distribution of fault radon emission source terms . This represents the probability that each point belongs to the "peak region," and is mapped to the [0,1] interval by the Sigmoid activation function, completing the initial localization of multi-peak intelligent recognition; The ReLU activation function ensures non-negativity, representing the intensity of fault radon gas release sources at each spatial location.

[0103] To ensure that the model output conforms to unsteady physical laws and geological realities, the loss function... It consists of three parts, namely the data fitting loss value. Unsteady physical residual loss value and source term sparsity regularization loss value The formula is as follows: ; ; in, Generated through expert annotation =1.0, , .

[0104] The unsteady-state physical residual loss is constructed based on the unsteady-state radon diffusion-convection-source term equations, and its form is as follows: ; Where C is the radon concentration; t is time; and D is the diffusion coefficient. This refers to the velocity of underground airflow. The intensity distribution of radon gas release source terms in the fault zone.

[0105] Concentration curves predicted by the model (Obtained by multiplying the probability mask by the input concentration point by point), its partial derivative is calculated using an automatic fractionation technique: ; ; ; Next, the intensity distribution of the above derivatives and fault radon gas release source terms will be analyzed. Substituting into the unsteady-state control equations, we obtain the residual R(x, t) at each spatiotemporal point: ; Unsteady physical residual loss value Defined as: ; Source term sparsity regularization loss value The L1 norm of the radon release source term intensity distribution in the fault is obtained by calculating the L1 norm, which is used to force the source term to be spatially concentrated in the fault fracture zone region, reflecting the geological a priori constraints on radon release in the fault: ; The probability mask output for each candidate segment Post-processing is then performed, including threshold segmentation, connected component analysis, and precise peak center location. Specifically, a threshold is set. A binary mask is obtained; then, continuous regions are extracted, each region being considered a potential peak; for the j-th peak region, the concentration-weighted average position is taken as the peak center, and the formula is as follows: ; in, The weighted center position of the j-th peak. Let j be the set of sampling points contained in the j-th peak. Let i be the spatial coordinates of the i-th sampling point. Let be the corrected time-series radon concentration at the i-th sampling point.

[0106] Simultaneously, for the j-th peak, its peak height is calculated. Half-height and width The concentration was determined to be greater than or equal to half the peak height by interpolation. The interval length is used to characterize the peak's width, reflecting the spatial scale of the fault fracture zone; the confidence level is calculated. The confidence level is defined as the predicted mask within the peak region. The average value, i.e. Its function is to evaluate the reliability of the model's peak identification result; and to calculate the skewness. The formula for is used to quantify the symmetry of the peak shape, and it is as follows: ; in, The set of indices of all sampling points belonging to the j-th peak of the k-th segment; This represents the number of sampling points contained in the peak. , . Indicates an approximately symmetrical distribution. This indicates a positive skewness (right skew), which geologically can be interpreted as the gas originating from the left (along the direction of the survey line, the left side is a 90° counterclockwise rotation perpendicular to the survey line), or the fault dipping to the left. This indicates a negative skewness (left skew), which geologically can be interpreted as the gas originating from the right side (the right side is a 90° clockwise rotation perpendicular to the survey line direction), or the fault dipping to the right.

[0107] Preliminary screening The peaks are then analyzed further, and combined with the source term intensity distribution at the corresponding locations. Calculate the fault radon release flux of this peak. : ; in, This is the radon release coefficient from underground to the surface (a constant value, related to soil porosity and moisture content); Let be the spatial region of the j-th peak.

[0108] For the k-th candidate segment, the final output is a structured list. All peak features are summarized into the global output AllPeaks={ , ,..., }

[0109] Step 5: Invert fault geometric parameters based on the mapping rule base.

[0110] A multi-peak-fault mapping mechanism based on the intensity distribution of radon emission source terms in faults is established: the fault center is determined by the peak position of the source term distribution, the width of the fracture zone is determined by the width of the non-zero region of the source term, and the radon emission flux in the fault is quantitatively calculated by the integrated flux of the source term, realizing a quantitative inversion from concentration observation to fault activity assessment. The fault strike and branch topology are reconstructed to achieve intelligent inference from point anomalies to the fault system.

[0111] The input for this step is the global peak feature set output from step 4. and intensity distribution of fault radon emission source terms First, allpeaks are sorted by spatial coordinate x and clustered according to local proximity to form several multi-peak combined units (MPUs), each MPU corresponding to a potential fault or fault complex. Simultaneously, the intensity distribution of fault radon emission source terms is utilized. The peak position is used to accurately locate the fault center, the width of the fracture zone is determined by the width of the non-zero region of the source term, and the source term integral flux is used to determine the fault center. Quantitative evaluation of fault activity. Here, the pre-defined multi-peak-fault mapping rule base is shown in Table 1.

[0112] Table 1. Multi-peak-fault mapping rule base In the table above, skewness reflects the asymmetry of the peak shape. Positive skewness indicates that radon gas is continuously supplied from the left side, pointing to the leftward dip of the fault (the hanging wall is on the left, and vice versa). If both the conditions of multi-peak clustering and bimodal asymmetry are met, the bimodal asymmetry rule is applied preferentially. The coding of dip_direction is shown in Table 2.

[0113] Table 2 Encoding Rules The inversion process for each MPU follows this logic: First, determine the number of peaks. If there is only one valid peak, apply the single-peak rule; if there are two valid peaks, calculate the product of peak spacing and skewness. Determine its symmetry; if there are ≥3 valid peaks, check whether it conforms to the M-type or multi-peak dense cluster pattern.

[0114] Then, the intensity distribution of radon gas release sources in the fault was used. The precise location of the fault (source term peak coordinates), the range of the fracture zone (source term non-zero region), and the active flux (source term integral value) are determined, and the specific parameters of the structured fault parameter object are shown in Table 3.

[0115] Table 3 Structured Fault Parameter Objects The source distribution is the core inversion parameter, which is directly output by the non-steady-state physical information neural network; All geometric parameters (center_x, gouge_width) are preferentially distributed from the source term. Multi-peak characteristics are only used for auxiliary verification; radon_flux is a quantitative indicator of fault activity, and activity_level is its hierarchical representation; The final output is a list of standardized fault parameters, InFerredFaults, where each element is a FaultsSegment object.

[0116] Step 6: Generation and standardized output of fault location results.

[0117] The final output includes the fault centerline, fracture zone extent, multi-peak annotation map, fault radon emission flux distribution map, and confidence interval thermogram. Results are generated in standard formats such as GeoJSON, CSV, and GeoTIFF. The entire process forms a closed-loop system of high-density acquisition, intelligent identification, physical inversion, and standard output, demonstrating good engineering practicality and scalability.

[0118] This step requires the input of the original survey line coordinates and fault parameter list. Each FaultsSegment contains parameters as shown in Table 4. The original survey line coordinate information includes the GPS coordinates of the survey line's starting point. and survey line direction angle , From the coordinates of the starting point of the survey line and endpoint coordinates The calculation yields the following formula: ; in, The difference between the x-coordinates of the endpoint and the starting point. The difference between the ordinates of the endpoint and the starting point. The unit is radians, which need to be converted to degrees. .

[0119] First, perform a geographic coordinate transformation. Convert the one-dimensional fault center location (center_x) from step 5 into actual geographic coordinates. , The formula is as follows: ; Then, based on the fault zone width gouge_width, the left and right boundary points of the fault zone are generated, with the coordinates of the left boundary being... The coordinates of the right boundary are .

[0120] Finally, the system automatically generates fault location maps, multi-peak annotation maps, fault radon emission flux distribution maps, and confidence interval heatmaps. The fault location map layer features include a centerline, fault zone extent, and radon emission flux levels (high / medium / low). The centerline includes `fault_id`, `confidence`, and `peak_pattern`, while the fault zone extent includes `gouge_width` and `dip_direction`. The multi-peak annotation map overlays the identified peak positions, widths, and confidence levels onto the original radon concentration profile. The output format is PDF, with peak positions marked by vertical lines and color depth indicating confidence levels. The fault radon emission flux distribution map displays the source term intensity distribution along the survey line. The curves are marked with peak positions as fault centers and non-zero intervals as fracture zone ranges. The confidence interval heatmap generates buffer zone heat rendering on the map based on the confidence values ​​of each fault. High confidence intervals (≥0.85) are marked in red, medium confidence intervals (0.7~0.85) are marked in orange, and low confidence intervals (≤0.7) are marked in gray.

[0121] This embodiment effectively solves the key technical challenges of traditional radon detection methods in locating concealed faults, including reliance on manual experience, susceptibility to environmental interference, difficulty in analyzing multi-peak anomaly structures, inability to quantitatively invert fault source distribution and active flux, and low positioning accuracy. Specifically, the scheme introduces high-density continuous sampling and a multi-variable collaborative environmental interference correction mechanism, significantly reducing false anomalies caused by factors such as temperature, humidity, and air pressure, while maintaining the temporal continuity of the data. The unsteady radon diffusion-convection-source term coupling physical equations are embedded into a deep learning model to construct a learnable source term inversion mechanism, breaking through the assumptions of traditional steady-state models and enabling the solution of the inverse problem of inverting fault source distribution from observed concentrations. Therefore, it can not only automatically and interpretably identify and extract multi-peak anomaly features but also quantitatively output the key activity parameter of fault radon release flux. Furthermore, by combining the source term intensity distribution with a multi-peak-fault mapping rule base, the scheme can accurately invert the fault center location, fracture zone width, and dip, and reconstruct the topology of complex fault systems. The standardized output significantly improves the efficiency and reliability of the process from field detection to engineering decision-making. Overall, this scheme upgrades the detection of concealed faults from qualitative interpretation to quantitative inverse source inversion guided by unsteady physics, providing high-precision, high-efficiency, and high-availability technical support for urban seismic safety assessment, site selection of major infrastructure, and the construction of active fault databases.

[0122] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.

[0123] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0124] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not performed.

[0125] 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 variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for locating hidden faults based on intelligent identification of radon multi-peak anomalies, characterized in that, The method includes: The original soil radon concentration time series data and synchronously collected environmental parameters are obtained from multiple sampling points along a preset survey line in the area to be tested. The original soil radon concentration time series data are preprocessed to obtain a standardized radon concentration time series. Based on the environmental parameters, a multivariate collaborative environmental disturbance correction model is constructed. The standardized radon concentration time series is dynamically corrected using the multivariate collaborative environmental disturbance correction model to obtain the corrected radon concentration time series. The sliding window dynamic threshold method is used to perform anomaly screening on the corrected radon concentration time series. Spatially continuous anomaly points are merged into candidate anomaly segments, and the time series concentration subsequence corresponding to each candidate anomaly segment is extracted. A physical information neural network model is constructed that integrates the physical constraints of unsteady radon diffusion-convection-source term coupling. The time-series concentration subsequences corresponding to each candidate anomaly segment are input into the physical information neural network model, and the concentration probability mask and the intensity distribution of fault radon emission source term are output in parallel. The concentration probability mask is post-processed to extract peak features, and a global peak feature set is obtained. Based on a preset multi-peak-fault mapping rule library, the fault center coordinates are determined by the peak position of the intensity distribution of the fault radon emission source term, the width of the fracture zone is determined by the width of the non-zero region of the source term, the fault radon emission flux is determined by the integrated flux of the source term, and the fault dip is inferred by combining the peak characteristics to generate hidden fault parameters. By combining the hidden fault parameters with the geographic coordinates of the survey line, standardized fault location results, including fault location maps, source term distribution maps, and confidence interval heat maps, are generated and output.

2. The method according to claim 1, characterized in that, The process involves constructing a multivariate collaborative environmental disturbance correction model based on the environmental parameters, and then using this model to dynamically correct the standardized radon concentration time series to obtain a corrected radon concentration time series, including: Temperature, humidity, and atmospheric pressure are input as covariates into a pre-trained multivariate collaborative environmental disturbance correction model, and the output is the environmental deviation. The corrected radon concentration time series is obtained by subtracting the environmental bias from the standardized radon concentration time series. The multivariate collaborative environmental interference correction model is trained by a lightweight multilayer perceptron network, which includes: an input layer with 3 neurons, two hidden layers each with 8 neurons and using the ReLU activation function, and a single-neuron output layer for outputting the prediction bias.

3. The method according to claim 1, characterized in that, The method of using a sliding window dynamic threshold to perform anomaly screening on the corrected radon concentration time series merges spatially continuous anomaly points into candidate anomaly segments, including: Taking each sampling point in the corrected radon concentration time series as the center, sampling points within a neighborhood range that satisfy a preset physical length are selected along the measurement line direction to form a local window; Calculate the local mean and local standard deviation of radon concentration within the local window; The local coefficient of variation is calculated based on the local mean and the local standard deviation, and the sensitivity function is determined based on the local coefficient of variation. The corrected radon concentration value at the current sampling point is compared with a dynamic threshold determined based on the local mean, the local standard deviation, and the sensitivity function. If the corrected radon concentration value is greater than the dynamic threshold, the current sampling point is marked as an outlier. Spatial connectivity analysis is performed on all marked outliers to merge consecutive outliers into independent initial outlier segments, and candidate outlier segments are selected based on a set physical length threshold.

4. The method according to claim 1, characterized in that, The physical information neural network model uses a one-dimensional U-Net as the backbone network, including an encoder, a bottleneck layer, and a decoder, with a concentration prediction head and a source term inversion head set in parallel at the end of the decoder; wherein, The input to the physical information neural network model is a three-dimensional tensor formed by splicing the normalized spatial coordinates, time coordinates and the corrected radon concentration value. The concentration prediction head outputs the concentration probability mask through the Sigmoid activation function, and the source term inversion head outputs the intensity distribution of the fault radon gas release source term through the ReLU activation function; The total loss function of the physical information neural network model is composed of a weighted sum of data fitting loss, unsteady physical residual loss, and source term sparsity regularization loss. The data fitting loss measures the difference between the predicted concentration probability mask and the expert-annotated true mask. The unsteady physical residual loss is constructed based on the unsteady radon diffusion-convection-source term equation. The source term sparsity regularization loss is the L1 norm of the source term intensity distribution output by the source term inversion head.

5. The method according to claim 4, characterized in that, The post-processing of the concentration probability mask to extract peak features yields a global peak feature set, including: The concentration probability mask is binarized by setting a preset threshold, and continuous high response regions are extracted as potential peak regions. For each potential peak region: the weighted average of the spatial coordinates is calculated as the peak center position, using the corrected radon concentration value of each sampling point within the potential peak region as the weight. The maximum corrected radon concentration within the potential peak region is taken as the peak height; The length of the interval where the corrected radon concentration is greater than or equal to the half-peak height is determined by interpolation and is taken as the half-width at half-peak. The average value of the concentration probability mask within the potential peak region is used as the identification confidence level of the corresponding peak. The ratio of the cube of the third central moment to the standard deviation of the corrected radon concentration value within the potential peak region is calculated as the peak shape skewness, which is used to characterize the peak shape symmetry and assist in inferring the fault tendency. The peak features extracted from all candidate anomaly segments are summarized to form the global peak feature set.

6. The method according to claim 4, characterized in that, The formula for calculating the data fitting loss is: ； In the formula, The loss value is used to fit the data; The number of sampling points contained in the k-th candidate anomaly segment; Provide the expert-labeled true mask for the i-th sampling point in the k-th candidate anomaly segment; It is the concentration probability mask for the i-th sampling point in the k-th candidate anomaly segment.

7. The method according to claim 4, characterized in that, The formula for calculating the unsteady physical residual loss is as follows: ； In the formula, This represents the unsteady physical residual loss value. The number of sampling points contained in the k-th candidate anomaly segment; The timing of the time sequence sampling for each sampling point; For spacetime points The physical residual at the location is ideally zero; among which, ； ； ； ； In the formula, Let be the first partial derivative of concentration with respect to time, representing the rate of change of radon concentration over time; Let be the first partial derivative of concentration with respect to space, representing the rate of change of radon concentration along the measurement line; To predict the radon concentration field, it is obtained by multiplying the concentration probability mask point by point with the input concentration curve; Let be the second partial derivative of concentration with respect to space, representing the degree of diffusion of concentration in space; The diffusion coefficient is denoted as . This refers to the velocity of underground airflow. The intensity distribution of radon gas release source terms in the fault zone.

8. The method according to claim 4, characterized in that, The formula for calculating the sparse regularization loss of the source term is: ； In the formula, The source term sparsity regularization loss value; The number of sampling points contained in the k-th candidate anomaly segment; The intensity of the fault radon gas release source term at the i-th sampling point; This is the absolute value of the source term strength.

9. The method according to claim 1, characterized in that, The preset multi-peak-fault mapping rule base is defined at least as follows: Single-peak mode: contains only one peak with a confidence level not less than a preset threshold, which is mapped to a simple main fault. The fault center is determined by the position of the source term peak and the width of the non-zero region of the source term is determined by the width of the fracture zone. Bimodal symmetric mode: containing two peaks with a spacing within the preset fracture zone width, and the absolute values ​​of the skewness of the two peaks are both less than the preset threshold or the product of the skewness of the two peaks is greater than zero and approximately equal, which is mapped to a fault fracture zone with a certain width, and the dip is determined to be symmetrical and undiverted. Bimodal asymmetric mode: containing two peaks with opposite signs of skewness and absolute values ​​of both peaks greater than a preset threshold, it is mapped to a dip fault, and the dip direction of the fault is determined by the sign of skewness; The three-peak M-type pattern contains three peaks, with the amplitude of the middle peak being significantly lower than the average of the two side peaks by a certain proportion, and the two side peaks being roughly symmetrical, which is a mapping of a main fault plus secondary ruptures or graben structures on both sides. Multi-peak dense cluster mode: contains no less than 4 peaks and the distance between adjacent peaks is less than the preset dense threshold, which is mapped to a complex strong fracture zone or fault intersection area; If the conditions of both the multi-peak dense cluster mode and the bimodal asymmetric mode are met simultaneously, the bimodal asymmetric mode rule shall be applied preferentially.

Images