A muon incidence position prediction method and system

Abstract

The application discloses a muon incident position prediction method and system, and belongs to the technical field of muon detection. The application sets up a multi-layer muon detection array in a to-be-detected area, and establishes a reverse propagation correction track based on muon events obtained by each detection unit. After voxelization of the to-be-detected area, track response, energy response, scattering response and propagation uncertainty response are generated according to the distance between voxels and the reverse propagation correction track, and voxel fingerprint features are spliced. Variance is taken for the track, energy and scattering response at each time, and reverse mapping is performed to obtain stable evolution features. The fingerprint features and the stable evolution features are fused to construct a stability-gated correction-type enhanced adjacency matrix, and a graph convolution classification network is used for position prediction. The application significantly improves the prediction accuracy of the muon incident position through multi-dimensional response fusion and stability enhancement mechanism, and can be widely applied to muon transmission imaging and cosmic ray monitoring scenes.

Description

Technical Field

[0001] This invention relates to the field of muon detection technology, and more specifically to a method and system for predicting the incident position of muons. Background Technology

[0002] Muon imaging is a non-destructive testing technique that utilizes the energy loss and scattering effects of cosmic ray muons as they penetrate matter to detect the internal structure of the area being inspected. Due to the strong penetrating power of muons, they can be used for internal state detection in large structures, underground spaces, and complex environments. Current muon detection technologies typically employ multi-layer muon arrays to detect muons passing through the area to be inspected. The upper and lower detection units record the incident position, exit position, and propagation direction of the muons. The trajectory of the muons is fitted based on multiple detection points, and further combined with information such as the deflection angle and energy changes of the muons in the medium, an inversion analysis is performed on the spatial regions that the muons may have traversed.

[0003] Existing muon incident position prediction methods involve deploying multiple detector arrays around the detection area to collect position signals generated by muons as they pass through different detector units. Muon propagation trajectories are established based on the spatial geometric relationships between the detector layers, and the possible incident regions of muons are determined using a trajectory reverse extrapolation method. Simultaneously, the detection area is divided into multiple spatial units. By statistically analyzing the number of muons passing through each spatial unit, the degree of trajectory intersection, and the change in scattering angle, the response intensity corresponding to each spatial position is calculated. The position with the highest response intensity is then selected as the predicted muon incident position.

[0004] However, there are still some shortcomings in practical applications: existing methods usually rely mainly on the geometric intersection of trajectories or simple statistical responses to determine the position, without fully considering the stable changes of muon response characteristics at different time scales. This makes it easy for incident position shifts to occur when there is noise, local occlusion or complex medium environment in the detection data, which reduces the accuracy of muon incident position prediction. Summary of the Invention

[0005] To address the aforementioned shortcomings in the prior art, the present invention provides a method and system for predicting the incident position of muons, which solves the problem of low accuracy in predicting the incident position of muons in the prior art.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for predicting the incident position of a muon, comprising the following steps:

[0007] S1. Set up a multi-layer muon detection array in the area to be detected, record each muon event for each detection unit, and establish a backpropagation correction trajectory based on the muon events;

[0008] S2. Divide the area to be detected into multiple voxels. Based on the distance between the voxels and the backpropagation correction trajectory, generate trajectory response, energy response, scattering response and propagation uncertainty response, and stitch them together to obtain the fingerprint features of the voxels.

[0009] S3. Take the variance of the trajectory response, energy response and scattering response at each time point, and back-map them to obtain the trajectory stability value, energy stability value and scattering stability value, and then splice them to obtain the stable evolution characteristics of the voxel.

[0010] S4. Based on fingerprint features and stable evolution features, construct a stability-gated modified enhanced adjacency matrix and use a graph convolutional integral network to obtain the predicted muon incident position.

[0011] Furthermore, S2 includes the following sub-steps:

[0012] S21. Divide the area to be detected into multiple voxels and calculate the shortest distance between the voxels and the backpropagation correction trajectory.

[0013] S22. Construct a Gaussian kernel function based on the shortest distance, and superimpose the contribution values ​​of all backpropagation correction trajectories to obtain the trajectory response of each voxel.

[0014] S23. Multiply the energy loss of the muon by the contribution value of the backpropagation correction trajectory, and sum them up to obtain the energy response;

[0015] S24. Calculate the scattering response based on the deflection angle and the contribution value of the backpropagation correction trajectory;

[0016] S25. Obtain the propagation uncertainty response based on the shortest distance, energy response, and scattering response;

[0017] S26. The trajectory response, energy response, scattering response and propagation uncertainty response belonging to the same voxel are spliced ​​together to obtain the fingerprint features of the corresponding voxel.

[0018] Furthermore, the process of calculating the scattering response in S24 includes: squaring the deflection angle and theoretical deflection angle for each muon, taking the ratio of the two to obtain the relative scattering intensity, multiplying the relative scattering intensity by the contribution value of the backpropagation correction trajectory, accumulating all the multiplication results, and taking the average value to obtain the scattering response of each voxel.

[0019] Furthermore, S25 includes the following sub-steps:

[0020] S251. The ratio of the square of the shortest distance to the square of the standard deviation of the position measurement is taken as the detection uncertainty component.

[0021] S252. The ratio of the detection uncertainty component to the maximum detection uncertainty component is taken as the detection uncertainty.

[0022] S253. The ratio of the scattering response to the maximum scattering response is taken as the scattering uncertainty;

[0023] S254. Take the average energy response of each voxel to obtain the average energy response.

[0024] S255. Subtract the energy response of each voxel from the mean energy response and take the absolute value to obtain the energy deviation value.

[0025] S256. The ratio of the energy deviation value to the maximum energy deviation value is taken as the energy loss uncertainty.

[0026] S257. The average of the detection uncertainty, scattering uncertainty and energy loss uncertainty is taken to obtain the propagation uncertainty response.

[0027] Furthermore, S3 includes the following sub-steps:

[0028] S31. Obtain the variance of the trajectory response at each time point within the time window of the same voxel, and perform reverse mapping to obtain the trajectory stability value.

[0029] S32. Obtain the variance of the energy response at each time point within the time window of the same voxel, and perform inverse mapping to obtain the stable energy value;

[0030] S33. Obtain the variance of the scattering response at each time point within the time window of the same voxel, and perform inverse mapping to obtain the scattering stability value.

[0031] S34. By concatenating the trajectory stability value, energy stability value, and scattering stability value, the stable evolution characteristics of the voxel are obtained.

[0032] Furthermore, S4 includes the following sub-steps:

[0033] S41. Use fingerprint features as node features;

[0034] S42. Obtain the adjacency coefficient based on the spatial distance between two voxels;

[0035] S43. Calculate the cosine similarity of the node features of two voxels and enhance the adjacency coefficient to obtain the enhanced adjacency coefficient.

[0036] S44. Based on the enhanced adjacency coefficient, construct the enhanced adjacency matrix, and then construct the node feature matrix using the features of each node.

[0037] S45. Based on the stable evolution characteristics and the enhanced adjacency matrix, the stability-gated modified enhanced adjacency matrix is ​​obtained;

[0038] S46. Input the stability-gated modified enhanced adjacency matrix and node feature matrix into the graph convolutional integral network to obtain the predicted muon incident position.

[0039] Furthermore, the enhancement formula in S43 is:

[0040] ,

[0041] in, For the first Individual elements and the first The enhanced adjacency coefficient of individual units, For the first Individual elements and the first Cosine similarity of individual units and For voxel numbers, For the first Individual elements and the first Adjacency coefficient of individual elements.

[0042] Furthermore, S45 includes the following sub-steps:

[0043] S451. Calculate the gating weights for the stable evolution characteristics of each voxel;

[0044] S452. Use the gating weights of each voxel as diagonal elements to generate diagonal node gating weights.

[0045] S453. Apply the diagonal node gating weights to the enhanced adjacency matrix to obtain the stability-gated modified enhanced adjacency matrix.

[0046] Furthermore, the graph convolutional network in S46 includes: normalization units, graph convolution, classification layers, and prediction layers;

[0047] The normalization unit is used to normalize the stability-gated modified enhanced adjacency matrix;

[0048] Graph convolution is used to process the normalized stability-gated modified augmented adjacency matrix and node feature matrix to obtain the depth feature matrix;

[0049] The classification layer is used to classify the depth features of each voxel in the depth feature matrix to obtain the voxel incident probability distribution.

[0050] The prediction layer is used to weight each voxel position with the voxel incident probability distribution to obtain the predicted muon incident position.

[0051] A muon incident position prediction system includes: a trajectory generation module, a fingerprint feature acquisition module, a stable evolution feature acquisition module, and a muon position prediction module;

[0052] The trajectory generation module is used to set up a multi-layer muon detection array in the area to be detected, record each muon event for each detection unit, and establish a backpropagation correction trajectory based on the muon events;

[0053] The fingerprint feature acquisition module is used to divide the area to be detected into multiple voxels, generate trajectory response, energy response, scattering response and propagation uncertainty response based on the distance between the voxels and the backpropagation correction trajectory, and then stitch them together to obtain the fingerprint features of the voxels.

[0054] The stable evolution feature acquisition module is used to take the variance of the trajectory response, energy response and scattering response at each time step, and then back-map them to obtain the trajectory stable value, energy stable value and scattering stable value, and then splice them to obtain the stable evolution feature of the voxel.

[0055] The muon location prediction module is used to construct a stability-gated modified enhanced adjacency matrix based on fingerprint features and stable evolution features, and to obtain the predicted muon incident position using a graph convolutional integral network.

[0056] The beneficial effects of this invention are as follows:

[0057] This invention establishes a backpropagation correction trajectory based on the direction of muon event motion by setting up a multi-layer muon detection array in the detection area, and combines it with a voxel space partitioning method to jointly characterize the trajectory information, energy change information, scattering response information, and propagation uncertainty information during the muon propagation process, forming a voxel fingerprint feature containing multi-dimensional physical response characteristics. Compared with the existing method that only uses the trajectory intersection relationship for position prediction, it can more comprehensively describe the response of different spatial locations to muon propagation behavior and improve the accuracy of incident position characterization.

[0058] Furthermore, this invention addresses the problem of response fluctuations caused by environmental changes, detection noise, and medium inhomogeneity during muon detection. By performing variance analysis on trajectory response, energy response, and scattering response, the stable variation patterns of each response feature are extracted, and the trajectory stable value, energy stable value, and scattering stable value are obtained by inverse mapping. This constructs stable evolution features, reduces the interference of instantaneous abnormal data on prediction results, and improves the robustness of prediction results.

[0059] Furthermore, this invention constructs a stability-gated modified enhanced adjacency matrix based on voxel fingerprint features and stable evolution features. Considering the spatial correlation between different voxels, stability information is introduced to dynamically adjust the correlation strength between voxels, so that spatial regions with similar muon propagation features and stable responses obtain higher correlation weights, while suppressing abnormal correlations that are greatly affected by noise, thereby improving the effectiveness of spatial feature propagation.

[0060] Furthermore, this invention employs a graph convolutional integral network for feature aggregation and position discrimination, which fully utilizes the spatial topological relationships between voxels and the muon propagation response law to achieve accurate prediction of the muon incident position. Compared with traditional methods based on geometric fitting and statistical response, this invention can adapt to the incident position prediction requirements under complex media environments, local occlusion, and fluctuations in probe data, thus improving the accuracy and stability of incident position prediction during muon imaging detection. Attached Figure Description

[0061] Figure 1 This is a flowchart of a method for predicting the incident position of muons. Detailed Implementation

[0062] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0063] Example 1, as Figure 1 As shown, a method for predicting the incident position of a muon includes the following steps:

[0064] S1. Set up a multi-layer muon detection array in the area to be detected, record each muon event for each detection unit, and establish a backpropagation correction trajectory based on the muon events;

[0065] S2. Divide the area to be detected into multiple voxels. Based on the distance between the voxels and the backpropagation correction trajectory, generate trajectory response, energy response, scattering response and propagation uncertainty response, and stitch them together to obtain the fingerprint features of the voxels.

[0066] S3. Take the variance of the trajectory response, energy response and scattering response at each time point, and back-map them to obtain the trajectory stability value, energy stability value and scattering stability value, and then splice them to obtain the stable evolution characteristics of the voxel.

[0067] S4. Based on fingerprint features and stable evolution features, construct a stability-gated modified enhanced adjacency matrix and use a graph convolutional integral network to obtain the predicted muon incident position.

[0068] In this embodiment, the muon event in S1 includes: the three-dimensional coordinates of the geometric center of the detection unit, the zenith angle of the propagation direction, and the azimuth angle of the propagation direction.

[0069] Each muon event is: ,in, For the first The Muzi incident. For the first The three-dimensional coordinates of the geometric center of the detection unit hit by the muon impact. For the first The direction of the propagation of a muse is at the zenith angle. For the first The direction and azimuth of the propagation of a muon.

[0070] In this embodiment, S1, establishing the backpropagation correction trajectory based on the direction of motion, includes:

[0071] Calculate the direction vector:

[0072] ,

[0073] in, For the first The direction vector of each muon. For the first The direction of the propagation of a muse is at the zenith angle. For the first The direction and azimuth of the propagation of a muon;

[0074] Based on the direction vector, establish the initial backpropagation trajectory:

[0075] ,

[0076] in, For the first The initial backpropagation trajectory of the muon, For the first The three-dimensional coordinates of the geometric center of the detection unit hit by the muon impact. To trace the distance of transmission;

[0077] The initial backpropagation trajectory is corrected to obtain the corrected backpropagation trajectory:

[0078] ,

[0079] ,

[0080] ,

[0081] in, For the first The backpropagation correction trajectory of a muon. For the first The angle of the mummie For the fixed empirical constants in Molière multiple Coulomb scattering, For the first The momentum of the muon itself. For the first The ratio of the speed of a muon to the speed of light in a vacuum. For the first The thickness of the medium through which a muon passes within the medium being measured. The intrinsic radiation length of the medium under test. This is a mapping function.

[0082] Mapping function The conversion from scalar angle to vector is achieved by using spherical angle decomposition: using the original incident direction vector of the muon. Using the central axis, the scattering deflection angle The decomposition is divided into two spatially orthogonal deflection components. The corresponding directional offset vector Δ in three-dimensional space is then calculated using the spherical coordinate-Cartesian coordinate conversion relationship. .

[0083] In this embodiment, S2 includes the following sub-steps:

[0084] S21. Divide the area to be detected into multiple voxels and calculate the shortest distance between the voxels and the backpropagation correction trajectory.

[0085] S22. Construct a Gaussian kernel function based on the shortest distance, and superimpose the contribution values ​​of all backpropagation correction trajectories to obtain the trajectory response of each voxel.

[0086] S23. Multiply the energy loss of the muon by the contribution value of the backpropagation correction trajectory, and sum them up to obtain the energy response;

[0087] S24. Calculate the scattering response based on the deflection angle and the contribution value of the backpropagation correction trajectory;

[0088] S25. Obtain the propagation uncertainty response based on the shortest distance, energy response, and scattering response;

[0089] S26. The trajectory response, energy response, scattering response and propagation uncertainty response belonging to the same voxel are spliced ​​together to obtain the fingerprint features of the corresponding voxel.

[0090] In this embodiment, the formula for calculating the shortest distance in S21 is: ,in, For the first The backpropagation correction trajectory of the first muon and the first The shortest distance of voxels To find the shortest distance by traversing the backtracking distance, For the first Individual factors, For the first The backpropagation correction trajectory of a muon. To obtain the L2 norm, the center coordinates of the voxels are used when calculating the distance;

[0091] The formula for calculating the trajectory response in S22 is: ,in, For the first Trajectory response of individual elements To preset fixed hyperparameters, It is a natural constant. The contribution value for correcting the trajectory during backpropagation;

[0092] The formula for calculating the energy response in S23 is: ,in, For the first Energy response of individual elements For the first Energy loss during the penetration process of a muon;

[0093] The process of calculating the scattering response in S24 includes: squaring the deflection angle and theoretical deflection angle for each muon, and taking the ratio of the two to obtain the relative scattering intensity; multiplying the relative scattering intensity by the contribution value of the backpropagation correction trajectory; summing all the multiplication results and taking the average to obtain the scattering response for each voxel. ,in, For the first Scattering response of individual units For the first The angle of the mummie The theoretical scattering deflection angle is calculated based on muon energy and medium thickness. To participate in the The number of muons calculated from the individual element response.

[0094] This invention constructs a Gaussian kernel function based on the shortest distance, accumulates the contributions of backpropagation correction trajectories corresponding to multiple muon events, and obtains a voxel-level trajectory response. This gives regions closer to the trajectory a higher response weight, improving the continuity and accuracy of spatial positioning information. Simultaneously, it fuses the energy loss information during muon penetration of the detection area with the trajectory contribution value to obtain an energy response. Furthermore, it constructs a scattering response by combining the relationship between the muon deflection angle and the theoretical deflection angle, enabling the voxel features to simultaneously include spatial propagation characteristics, energy attenuation characteristics, and medium scattering characteristics, thus enhancing the characterization capability of complex media internal states.

[0095] In this embodiment, S25 includes the following sub-steps:

[0096] S251. The ratio of the square of the shortest distance to the square of the standard deviation of the position measurement is taken as the detection uncertainty component.

[0097] S252. The ratio of the detection uncertainty component to the maximum detection uncertainty component is taken as the detection uncertainty.

[0098] S253. The ratio of the scattering response to the maximum scattering response is taken as the scattering uncertainty;

[0099] S254. Take the average energy response of each voxel to obtain the average energy response.

[0100] S255. Subtract the energy response of each voxel from the mean energy response and take the absolute value to obtain the energy deviation value.

[0101] S256. The ratio of the energy deviation value to the maximum energy deviation value is taken as the energy loss uncertainty.

[0102] S257. The average of the detection uncertainty, scattering uncertainty and energy loss uncertainty is taken to obtain the propagation uncertainty response.

[0103] The standard deviation of position measurement is the standard deviation of the position measurement deviation caused by the size limitation of the detection unit, electronic readout error and positioning error when the detection unit in the muon detection array measures the incident position of the muon. It is used to characterize the uncertainty of the detection unit in measuring the spatial coordinates of the muon.

[0104] This invention obtains the detection uncertainty component by combining the spatial distance between the voxel and the backpropagation correction trajectory with the standard deviation of position measurement. The maximum detection uncertainty is then normalized to quantify the trajectory deviation impact at different spatial locations. By normalizing the scattering response, scattering uncertainty is obtained, effectively characterizing the propagation deviation caused by multiple scattering in different medium regions. Simultaneously, by analyzing the deviation of each voxel's energy response from the overall mean energy response, energy loss uncertainty is constructed to identify regions of abnormal energy attenuation. Finally, the detection uncertainty, scattering uncertainty, and energy loss uncertainty are fused to obtain the propagation uncertainty response. This ensures that voxel characteristics not only include the muon propagation path, energy change, and scattering characteristics but also reflect the reliability during propagation, reducing position prediction errors caused by detection errors, medium inhomogeneity, and abnormal energy loss.

[0105] In this embodiment, S3 includes the following sub-steps:

[0106] S31. Obtain the variance of the trajectory response at each time point within the time window of the same voxel, and perform reverse mapping to obtain the trajectory stability value.

[0107] S32. Obtain the variance of the energy response at each time point within the time window of the same voxel, and perform inverse mapping to obtain the stable energy value;

[0108] S33. Obtain the variance of the scattering response at each time point within the time window of the same voxel, and perform inverse mapping to obtain the scattering stability value.

[0109] S34. By concatenating the trajectory stability value, energy stability value, and scattering stability value, the stable evolution characteristics of the voxel are obtained.

[0110] The formulas for the reverse mapping in S31, S32, and S33 are:

[0111] ,

[0112] in, For the first Stable values ​​of individual elements For the first Variance of individual elements To avoid parameters with a denominator of 0.

[0113] This invention performs time-series statistical analysis on the trajectory response, energy response, and scattering response of the same voxel within a time window, calculates the variance of the changes in various response characteristics, and converts the variance into a stable value through a reverse mapping method. This allows voxels with smaller variances and more stable response changes to obtain higher stability evaluations, thereby effectively extracting the stable law of muon propagation response changes over time.

[0114] The propagation uncertainty response of S25 is used to measure the reliability of the match between a single muon propagation evidence and the corresponding spatial location; while S3 is used to describe the temporal consistency of the spatial voxel response during continuous detection, reflecting whether the response characteristics of the region remain stable.

[0115] In this embodiment, S4 includes the following sub-steps:

[0116] S41. Use fingerprint features as node features;

[0117] S42. Obtain the adjacency coefficient based on the spatial distance between two voxels;

[0118] S43. Calculate the cosine similarity of the node features of two voxels and enhance the adjacency coefficient to obtain the enhanced adjacency coefficient.

[0119] S44. Based on the enhanced adjacency coefficient, construct the enhanced adjacency matrix, and then construct the node feature matrix using the features of each node.

[0120] S45. Based on the stable evolution characteristics and the enhanced adjacency matrix, the stability-gated modified enhanced adjacency matrix is ​​obtained;

[0121] S46. Input the stability-gated modified enhanced adjacency matrix and node feature matrix into the graph convolutional integral network to obtain the predicted muon incident position.

[0122] In this embodiment, the expression for the adjacency coefficient in S42 is: ,in, For the first Individual factors, For the first Individual factors, This is the spatial attenuation parameter.

[0123] In this embodiment, the enhancement formula in S43 is:

[0124] ,

[0125] in, For the first Individual elements and the first The enhanced adjacency coefficient of individual units, For the first Individual elements and the first Cosine similarity of individual units and The voxel number. For the first Individual elements and the first Adjacency coefficient of individual elements.

[0126] This invention calculates the cosine similarity between different voxel node features and enhances the adjacency coefficient obtained based on spatial distance, so that voxels that are spatially close and have similar muon propagation response features have higher association weights, while reducing the association influence between voxels that are spatially close but have large differences in response features, thereby achieving joint modeling of spatial relationships and muon physical response features.

[0127] In this embodiment, S45 includes the following sub-steps:

[0128] S451. For the stable evolution characteristics of each voxel, calculate the gating weights:

[0129] ,

[0130] in, For the first Gating weights of individual elements, For the first Stable evolutionary characteristics of individual elements For learnable weight matrix, For learnable weight bias, The voxel number. For the Sigmoid function;

[0131] S452. Generate diagonal node gating weights by using the gating weights of each voxel as diagonal elements:

[0132] ,

[0133] in, For diagonal node gating weights, Operators for generating diagonal matrices The gating weights for the first voxel. For the first Gating weights of individual elements, The number of voxels;

[0134] S453. Apply the diagonal node gating weights to the enhanced adjacency matrix to obtain the stability-gated modified enhanced adjacency matrix:

[0135] ,

[0136] in, For stability-gated modified enhanced adjacency matrix, For diagonal node gating weights, To enhance the adjacency matrix.

[0137] This invention generates gated weights based on voxel stable evolution features and uses these weights to dynamically modify the enhanced adjacency matrix, enabling the spatial relationships in the graph structure to adaptively adjust based on the stability of voxel responses over time. Specifically, by performing a nonlinear mapping on the stable evolution features, gated weights reflecting different voxel stability are obtained, allowing voxels with higher stability to have stronger information transmission capabilities, while suppressing the influence of voxels with larger response fluctuations on adjacency relationships. Furthermore, by constructing a diagonal node gated weight matrix and applying it to both sides of the enhanced adjacency matrix, bidirectional adjustment of the relationships between nodes is achieved, allowing stability information to simultaneously influence the information propagation strength between source and target nodes. Compared to methods that construct fixed adjacency relationships based solely on spatial distance and feature similarity, the stability-gated modified enhanced adjacency matrix constructed in this invention can dynamically optimize the graph structure according to the stability of voxel responses. This allows graph convolutional integral networks to focus more on spatial regions with stable physical responses, reducing interference from unstable response regions on the feature aggregation process, and improving the graph structure's expressive power as well as the accuracy and robustness of muon incident position prediction.

[0138] In this embodiment, the graph convolutional network in S46 includes: a normalization unit, a graph convolution, a classification layer, and a prediction layer;

[0139] The normalization unit is used to normalize the stability-gated modified enhanced adjacency matrix:

[0140] ,

[0141] in, For the normalized stability-gated modified enhanced adjacency matrix, For degree matrix, , The diagonal elements of the degree matrix, for Middle element;

[0142] Graph convolution is used to process the normalized stability-gated modified augmented adjacency matrix and node feature matrix to obtain the depth feature matrix:

[0143] ,

[0144] in, For the first The output of layered graph convolutional GCN, For the first The input of layered graph convolutional network (GCN) The node feature matrix, For the first The weight matrix of a layered graph convolutional network (GCN). The layer number of the graph convolutional GCN;

[0145] The classification layer is used to classify the depth features of each voxel in the depth feature matrix, resulting in the voxel incident probability distribution:

[0146] ,

[0147] ,

[0148] in, For the first Classification output of individual elements For the first In-depth characteristics of individual elements This is the classification weight matrix. For classification bias, It is a natural exponential function. For the first Classification output of individual elements For the first Voxel incident probability distribution of individual voxels;

[0149] The prediction layer is used to weight each voxel position with the voxel incident probability distribution to obtain the predicted muon incident position:

[0150] ,

[0151] in, To predict the incident position of the muon.

[0152] In this embodiment, a 3-layer graph convolution is used, and the output of the last layer is used as the depth feature matrix.

[0153] This invention processes the stability-gated modified enhanced adjacency matrix using a normalization unit, ensuring that the association weights between different voxels are on a uniform scale. This reduces the impact of numerical differences in the adjacency matrix on the graph feature propagation process and improves the stability of graph structure computation. Furthermore, graph convolution is used to jointly process the normalized adjacency matrix and the node feature matrix, enabling the network to fully integrate spatial association information between voxels, as well as trajectory response, energy response, scattering response, propagation uncertainty response, and stable evolution characteristics, achieving deep feature extraction of muon propagation patterns. Subsequently, a classification layer classifies the deep features of each voxel to obtain the muon incident probability distribution corresponding to different voxels, transforming incident position prediction from a single position judgment to a comprehensive analysis based on spatial probability distribution. Finally, a prediction layer weighted and fuses the spatial positions of each voxel with their corresponding incident probabilities to obtain continuous predicted muon incident positions, reducing the position quantization error caused by voxel discrete partitioning. Compared to traditional methods based on trajectory fitting or response statistics, this invention can comprehensively utilize spatial topological relationships, multi-physics response characteristics, and stability information to improve the accuracy and stability of muon incident position prediction in complex media and noisy environments.

[0154] Example 2, a muon incident position prediction system, includes: a trajectory generation module, a fingerprint feature acquisition module, a stable evolution feature acquisition module, and a muon position prediction module;

[0155] The trajectory generation module is used to set up a multi-layer muon detection array in the area to be detected, record each muon event for each detection unit, and establish a backpropagation correction trajectory based on the muon events;

[0156] The fingerprint feature acquisition module is used to divide the area to be detected into multiple voxels, generate trajectory response, energy response, scattering response and propagation uncertainty response based on the distance between the voxels and the backpropagation correction trajectory, and then stitch them together to obtain the fingerprint features of the voxels.

[0157] The stable evolution feature acquisition module is used to take the variance of the trajectory response, energy response and scattering response at each time step, and then back-map them to obtain the trajectory stable value, energy stable value and scattering stable value, and then splice them to obtain the stable evolution feature of the voxel.

[0158] The muon location prediction module is used to construct a stability-gated modified enhanced adjacency matrix based on fingerprint features and stable evolution features, and to obtain the predicted muon incident position using a graph convolutional integral network.

[0159] The specific implementation process of Example 2 is the same as that of Example 1.

[0160] This invention establishes a backpropagation correction trajectory based on the direction of muon event motion by setting up a multi-layer muon detection array in the detection area, and combines it with a voxel space partitioning method to jointly characterize the trajectory information, energy change information, scattering response information, and propagation uncertainty information during the muon propagation process, forming a voxel fingerprint feature containing multi-dimensional physical response characteristics. Compared with the existing method that only uses the trajectory intersection relationship for position prediction, it can more comprehensively describe the response of different spatial locations to muon propagation behavior and improve the accuracy of incident position characterization.

[0161] Furthermore, this invention addresses the problem of response fluctuations caused by environmental changes, detection noise, and medium inhomogeneity during muon detection. By performing variance analysis on trajectory response, energy response, and scattering response, the stable variation patterns of each response feature are extracted, and the trajectory stable value, energy stable value, and scattering stable value are obtained by inverse mapping. This constructs stable evolution features, reduces the interference of instantaneous abnormal data on prediction results, and improves the robustness of prediction results.

[0162] Furthermore, this invention constructs a stability-gated modified enhanced adjacency matrix based on voxel fingerprint features and stable evolution features. Considering the spatial correlation between different voxels, stability information is introduced to dynamically adjust the correlation strength between voxels, so that spatial regions with similar muon propagation features and stable responses obtain higher correlation weights, while suppressing abnormal correlations that are greatly affected by noise, thereby improving the effectiveness of spatial feature propagation.

[0163] Furthermore, this invention employs a graph convolutional integral network for feature aggregation and position discrimination, which fully utilizes the spatial topological relationships between voxels and the muon propagation response law to achieve accurate prediction of the muon incident position. Compared with traditional methods based on geometric fitting and statistical response, this invention can adapt to the incident position prediction requirements under complex media environments, local occlusion, and fluctuations in probe data, thus improving the accuracy and stability of incident position prediction during muon imaging detection.

[0164] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the incident position of a muon, characterized in that, Includes the following steps: S1. Set up a multi-layer muon detection array in the area to be detected, record each muon event for each detection unit, and establish a backpropagation correction trajectory based on the muon events; S2. Divide the area to be detected into multiple voxels. Based on the distance between the voxels and the backpropagation correction trajectory, generate trajectory response, energy response, scattering response and propagation uncertainty response, and stitch them together to obtain the fingerprint features of the voxels. S3. Take the variance of the trajectory response, energy response and scattering response at each time point, and back-map them to obtain the trajectory stability value, energy stability value and scattering stability value, and then splice them to obtain the stable evolution characteristics of the voxel. S4. Based on fingerprint features and stable evolution features, construct a stability-gated modified enhanced adjacency matrix and use a graph convolutional integral network to obtain the predicted muon incident position.

2. The muon incident position prediction method according to claim 1, characterized in that, S2 includes the following steps: S21. Divide the area to be detected into multiple voxels and calculate the shortest distance between the voxels and the backpropagation correction trajectory. S22. Construct a Gaussian kernel function based on the shortest distance, and superimpose the contribution values ​​of all backpropagation correction trajectories to obtain the trajectory response of each voxel. S23. Multiply the energy loss of the muon by the contribution value of the backpropagation correction trajectory, and sum them up to obtain the energy response; S24. Calculate the scattering response based on the deflection angle and the contribution value of the backpropagation correction trajectory; S25. Obtain the propagation uncertainty response based on the shortest distance, energy response, and scattering response; S26. The trajectory response, energy response, scattering response and propagation uncertainty response belonging to the same voxel are spliced ​​together to obtain the fingerprint features of the corresponding voxel.

3. The muon incident position prediction method according to claim 2, characterized in that, The process of calculating the scattering response in S24 includes: squaring the deflection angle and theoretical deflection angle of each muon respectively, taking the ratio of the two to obtain the relative scattering intensity, multiplying the relative scattering intensity by the contribution value of the backpropagation correction trajectory, accumulating all the multiplication results, and taking the average value to obtain the scattering response of each voxel.

4. The muon incident position prediction method according to claim 2, characterized in that, S25 includes the following sub-steps: S251. The ratio of the square of the shortest distance to the square of the standard deviation of the position measurement is taken as the detection uncertainty component. S252. The ratio of the detection uncertainty component to the maximum detection uncertainty component is taken as the detection uncertainty. S253. The ratio of the scattering response to the maximum scattering response is taken as the scattering uncertainty; S254. Take the average energy response of each voxel to obtain the average energy response. S255. Subtract the energy response of each voxel from the mean energy response and take the absolute value to obtain the energy deviation value. S256. The ratio of the energy deviation value to the maximum energy deviation value is taken as the energy loss uncertainty. S257. The average of the detection uncertainty, scattering uncertainty and energy loss uncertainty is taken to obtain the propagation uncertainty response.

5. The muon incident position prediction method according to claim 1, characterized in that, S3 includes the following steps: S31. Obtain the variance of the trajectory response at each time point within the time window of the same voxel, and perform reverse mapping to obtain the trajectory stability value. S32. Obtain the variance of the energy response at each time point within the time window of the same voxel, and perform inverse mapping to obtain the stable energy value; S33. Obtain the variance of the scattering response at each time point within the time window of the same voxel, and perform inverse mapping to obtain the scattering stability value. S34. By concatenating the trajectory stability value, energy stability value, and scattering stability value, the stable evolution characteristics of the voxel are obtained.

6. The muon incident position prediction method according to claim 1, characterized in that, S4 includes the following steps: S41. Use fingerprint features as node features; S42. Obtain the adjacency coefficient based on the spatial distance between two voxels; S43. Calculate the cosine similarity of the node features of two voxels and enhance the adjacency coefficient to obtain the enhanced adjacency coefficient. S44. Based on the enhanced adjacency coefficient, construct the enhanced adjacency matrix, and then construct the node feature matrix using the features of each node. S45. Based on the stable evolution characteristics and the enhanced adjacency matrix, the stability-gated modified enhanced adjacency matrix is ​​obtained; S46. Input the stability-gated modified enhanced adjacency matrix and node feature matrix into the graph convolutional integral network to obtain the predicted muon incident position.

7. The muon incident position prediction method according to claim 6, characterized in that, The enhancement formula in S43 is: ， in, For the first Individual elements and the first The enhanced adjacency coefficient of individual units, For the first Individual elements and the first Cosine similarity of individual units and For voxel numbers, For the first Individual elements and the first Adjacency coefficient of individual elements.

8. The muon incident position prediction method according to claim 6, characterized in that, S45 includes the following sub-steps: S451. Calculate the gating weights for the stable evolution characteristics of each voxel; S452. Use the gating weights of each voxel as diagonal elements to generate diagonal node gating weights. S453. Apply the diagonal node gating weights to the enhanced adjacency matrix to obtain the stability-gated modified enhanced adjacency matrix.

9. The muon incident position prediction method according to claim 6, characterized in that, The graph convolutional network in S46 includes: a normalization unit, a graph convolution, a classification layer, and a prediction layer; The normalization unit is used to normalize the stability-gated modified enhanced adjacency matrix; Graph convolution is used to process the normalized stability-gated modified augmented adjacency matrix and node feature matrix to obtain the depth feature matrix; The classification layer is used to classify the depth features of each voxel in the depth feature matrix to obtain the voxel incident probability distribution. The prediction layer is used to weight each voxel position with the voxel incident probability distribution to obtain the predicted muon incident position.

10. A muon incident position prediction system, implemented based on the muon incident position prediction method according to any one of claims 1 to 9, characterized in that, It includes: a trajectory generation module, a fingerprint feature acquisition module, a stable evolution feature acquisition module, and a muon position prediction module; The trajectory generation module is used to set up a multi-layer muon detection array in the area to be detected, record each muon event for each detection unit, and establish a backpropagation correction trajectory based on the muon events; The fingerprint feature acquisition module is used to divide the area to be detected into multiple voxels, generate trajectory response, energy response, scattering response and propagation uncertainty response based on the distance between the voxels and the backpropagation correction trajectory, and then stitch them together to obtain the fingerprint features of the voxels. The stable evolution feature acquisition module is used to take the variance of the trajectory response, energy response and scattering response at each time step, and then back-map them to obtain the trajectory stable value, energy stable value and scattering stable value, and then splice them to obtain the stable evolution feature of the voxel. The muon location prediction module is used to construct a stability-gated modified enhanced adjacency matrix based on fingerprint features and stable evolution features, and to obtain the predicted muon incident position using a graph convolutional integral network.

Images