A multi-gamma-source azimuthal inversion method, device, equipment and storage medium

Abstract

The present application relates to a kind of multi-gamma source azimuth inversion method, device, equipment and storage medium, wherein multi-gamma source azimuth inversion method includes: using Monte Carlo numerical simulation method to construct data set;Establish the convolutional neural network for multi-gamma source inversion;Wherein, cosine annealing scheduler is introduced to adjust the learning rate of convolutional neural network, for dynamically adjusting the training process of convolutional neural network, and, center loss function is introduced to reconstruct the loss function of convolutional neural network to adjust the classification performance and convergence of convolutional neural network training process;Based on data set, the convolutional neural network is trained to obtain multi-gamma source azimuth inversion model;Collect the sample to be inverted and input multi-gamma source azimuth inversion model, and based on multi-gamma source azimuth inversion model, the number of gamma source in the sample to be inverted and gamma source position are output.The present application can cover the dual target of source number classification and coordinate positioning, and provides an innovative solution for radiation field real-time dynamic reconstruction.

Description

Technical Field

[0001] This invention relates to the field of nuclear radiation field inversion technology, and in particular to a method, apparatus, equipment and storage medium for multi-γ source orientation inversion. Background Technology

[0002] Radiation field inversion and reconstruction technology plays a crucial role in nuclear safety, environmental monitoring, and other fields, serving as an indispensable key technology for ensuring personnel safety and optimizing nuclear facility operations. This technology measures radiation field data by deploying a limited number of detectors within a specific area, thereby visualizing the radiation environment and providing strong support for developing scientifically effective protective measures. This reduces personnel radiation dose and protects personnel health. Furthermore, radiation field inversion and reconstruction technology can optimize the operational procedures of nuclear facilities, improving operational efficiency and reducing radiation risks by preventing personnel from entering high-radiation areas.

[0003] In the event of a nuclear accident or other emergency, radiation field inversion and reconstruction technology can rapidly and accurately assess the radiation environment, quickly define the scope of the accident's impact, provide a basis for the rapid development of emergency plans, guide effective rescue and cleanup efforts, and mitigate the consequences of the accident. Specifically, radiation field inversion and reconstruction technology is a technique that uses limited monitoring data, mathematical models, and computational methods to infer the spatial distribution of the radiation field. Through this technology, the number, location, and activity of radioactive sources can be reconstructed, and the spatial distribution of the radiation field can be obtained.

[0004] In terms of nuclear facility safety, using a limited number of detectors to measure radiation field data enables visualization of the radiation environment, allowing for the development of more scientific and effective protective measures, reducing personnel radiation doses, protecting personnel health, and optimizing nuclear facility operation procedures. By avoiding personnel entering high-radiation areas, operational efficiency is improved and radiation risks are reduced. In the event of a nuclear accident or other emergency, rapid and accurate assessment of the radiation environment allows for the swift determination of the accident's impact area, providing a basis for the rapid development of emergency plans, guiding effective rescue and cleanup efforts, and mitigating the consequences of the accident.

[0005] With the development of artificial intelligence and deep learning technologies, neural network-based methods have shown significant advantages in complex data modeling, pattern recognition, and solving nonlinear problems. Training neural networks with data collected from a small number of dose detectors can accurately invert radiation dose fields in key areas, improving the efficiency of radiation field assessment. This technological advancement makes real-time radiation field inversion and reconstruction possible, providing new scientific tools and methodological support for research and applications in related fields. As technology continues to advance, radiation field inversion and reconstruction technology is also constantly developing and improving. In the future, this technology will become more intelligent and automated, capable of real-time and accurate inversion and reconstruction of radiation fields, providing more efficient and reliable technical support for fields such as nuclear safety. Simultaneously, this technology will be combined with other technologies to form a more comprehensive safety monitoring and early warning system, providing stronger support for ensuring personnel safety, optimizing nuclear facility operations, and other aspects.

[0006] In summary, radiation field inversion and reconstruction technology has broad application prospects and significant practical implications in fields such as nuclear safety. With continuous technological development and improvement, it is believed that this technology will play an even more important role in the future, providing stronger support for ensuring personnel safety and optimizing nuclear facility operations. However, radiation field inversion and reconstruction technology is still in its early stages and cannot effectively handle the problem of the number of unknown sources or the problem of radiation field inversion for sources with unknown locations. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a multi-γ source orientation inversion method, apparatus, equipment and storage medium.

[0008] To achieve the above-mentioned objective, this invention provides a multi-gamma source orientation inversion method, comprising the following steps:

[0009] S1. Construct the dataset using the Monte Carlo numerical simulation method;

[0010] S2. Establish a convolutional neural network for multi-γ source inversion; wherein, a cosine annealing scheduler is introduced to adjust the learning rate of the convolutional neural network to dynamically adjust the training process of the convolutional neural network, and a center loss function is introduced to reconstruct the loss function of the convolutional neural network to adjust the classification performance and convergence of the training process of the convolutional neural network.

[0011] S3. Train the convolutional neural network based on the dataset to obtain a multi-γ source orientation inversion model;

[0012] S4. Collect the samples to be inverted and input them into the multi-γ source orientation inversion model, and output the number and location of γ sources in the samples to be inverted based on the multi-γ source orientation inversion model.

[0013] According to one aspect of the present invention, step S1, the step of constructing the dataset using the Monte Carlo numerical simulation method, includes:

[0014] S11. Establish a ground environment model of preset dimensions based on MCNP software;

[0015] S12. Arrange the detector array in the ground environment model;

[0016] S13. Randomly arrange 1 to 4 γ-ray sources in the standard atmospheric medium of the ground environment model; wherein the activity range of the γ-ray sources is set between 1 Ci and 10 Ci;

[0017] S14. Use geometric grid cards to define the spatial structure of the ground environment model, and use material grid cards to configure the properties of different media;

[0018] S15. Based on the MCNP software, simulate and record the motion trajectory and radiation field data of photoelectric effect, Compton scattering and electron pair effect in the ground environment model to generate a dataset for training.

[0019] According to one aspect of the present invention, in step S11, the step of establishing a ground environment model of a preset size based on MCNP software, the ground environment model adopts a 10m×10m cement ground environment model.

[0020] In step S12, the detector array is arranged in a uniform coverage manner in the ground environment model.

[0021] According to one aspect of the present invention, in step S2, the step of establishing a convolutional neural network for multi-γ source inversion, the convolutional neural network includes: an input layer, a feature extraction module connected to the input layer, a classification decision module connected to the feature extraction module, and an output layer connected to the classification decision module;

[0022] The feature extraction module includes: several cascaded convolutional layers and max pooling layers;

[0023] Each convolutional layer is followed by a ReLU activation function, which is used to extract the spatial features of the radiation field intensity distribution of the samples in the dataset layer by layer, and to compress the spatial dimension step by step through the max pooling layer;

[0024] The classification decision module includes: a Flatten layer and multiple fully connected layers;

[0025] The Flatten layer is connected to the last max pooling layer of the feature extraction module and is used to expand the features output by the feature extraction module.

[0026] The fully connected layer has three layers, with the first fully connected layer connected to the Flatten layer and the third fully connected layer connected to the output layer.

[0027] The fully connected layer is configured with a tanh activation function to output the probability of the number of γ sources, and with a Softmax function to output the probability of the γ source coordinates classification.

[0028] According to one aspect of the present invention, in step S2, a cosine annealing scheduler is introduced to adjust the learning rate of the convolutional neural network. In this step of dynamically adjusting the training process of the convolutional neural network, the learning rate gradually decreases from an initial value according to a cosine curve to a preset minimum value; wherein the learning rate is expressed as:

[0029] ;

[0030] in, Indicates the learning rate. This represents the lower bound of the learning rate. Indicates the upper limit of the learning rate. Indicates the current training step count. This represents the total number of steps in the training cycle.

[0031] According to one aspect of the present invention, in step S2, in the step of establishing a convolutional neural network for multi-γ source inversion, the reconstructed loss function is a composite loss function that fuses the cross-entropy loss function and the center loss function;

[0032] The composite loss function is expressed as follows:

[0033] ;

[0034] ;

[0035] ;

[0036] in, Represents the composite loss function. Represents the cross-entropy loss function. Represents the central loss function. This represents the weighting coefficients that balance the two loss functions. Category labels representing coordinates, This indicates the number of classification categories, which includes 1 to 4 gamma radiation sources and their coordinates. This represents the probability values ​​corresponding to different classification labels predicted by the trained convolutional neural network. These represent the different classification labels predicted by the trained convolutional neural network. This represents the sample data in the input dataset. Indicates the center value of the label. The subscript indicates the total number of labels for each category. Indicates the order of label categories.

[0037] To achieve the above-mentioned objective, the present invention provides a multi-gamma source orientation inversion device, comprising:

[0038] The dataset generation module is used to construct datasets using the Monte Carlo numerical simulation method;

[0039] A convolutional neural network building module is used to build a convolutional neural network for multi-γ source inversion; wherein, a cosine annealing scheduler is introduced to adjust the learning rate of the convolutional neural network to dynamically adjust the training process of the convolutional neural network, and a center loss function is introduced to reconstruct the loss function of the convolutional neural network to adjust the classification performance and convergence of the training process of the convolutional neural network.

[0040] The training module trains the convolutional neural network based on the dataset to obtain a multi-γ source orientation inversion model.

[0041] The identification module is used to load the obtained multi-γ source orientation inversion model, collect samples to be inverted and input them into the multi-γ source orientation inversion model, and output the number and location of γ sources in the samples to be inverted based on the multi-γ source orientation inversion model.

[0042] To achieve the above-mentioned objectives, the present invention provides a device comprising at least one processor, at least one memory, and a data bus;

[0043] The processor and the memory communicate with each other via the data bus;

[0044] The memory stores program instructions that can be executed by the processor, which calls the program instructions to execute the aforementioned multi-gamma source orientation inversion method.

[0045] To achieve the above-mentioned objectives, the present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned multi-gamma source orientation inversion method.

[0046] According to one aspect of the present invention, this method addresses the challenge of identifying the number and coordinates of sources in multi-source radiation field inversion. It proposes a systematic approach based on Monte Carlo simulation and convolutional neural networks, successfully solving the problem of simultaneously determining the number of unknown radiation sources and their azimuth coordinates in radiation field inversion. By optimizing the detector array layout, introducing a cosine annealing strategy, and employing a composite loss function, the robustness of the obtained multi-γ source azimuth inversion model under complex scenarios is improved.

[0047] According to one aspect of the present invention, this approach can encompass the dual objectives of source quantity classification and coordinate positioning, providing an innovative solution for real-time dynamic reconstruction of the radiation field.

[0048] According to one aspect of the present invention, in the identification of the number of radioactive sources, the identification accuracy for 2 to 4 γ-radiation sources reaches 89.25%; in coordinate inversion, an innovative strategy of joint calculation using composite loss functions is adopted, which enables single-source positioning accuracy to exceed 98% (e.g., x-coordinate and y-coordinate positioning accuracies of 99.2% and 98.9%, respectively), dual-source positioning accuracy to exceed 81% (e.g., x-coordinate and y-coordinate positioning accuracies of the first source of 83.55% and 85.25%, respectively, and x-coordinate and y-coordinate positioning accuracies of the second source of 81.8% and 81.85%, respectively), and triple-source positioning accuracy to exceed 50% (e.g., x-coordinate and y-coordinate positioning accuracy of the first source of 89.25% and 98.9%, respectively). The accuracy rates were 76.13% and 75.97% respectively. The accuracy rates for the second source x-coordinate and y-coordinate positioning were 50.4% and 50.13% respectively, and the accuracy rates for the third source x-coordinate and y-coordinate positioning were 72.57% and 71.93% respectively. The four-source positioning accuracy exceeded 41% (the accuracy rates for the first source x-coordinate and y-coordinate positioning were 68.78% and 69.75%, the second source x-coordinate and y-coordinate positioning were 42.2% and 41.98%, the third source x-coordinate and y-coordinate positioning were 41.93% and 42.15%, and the fourth source x-coordinate and y-coordinate positioning were 67.68% and 66.85%).

[0049] According to one aspect of the present invention, a strategy of dynamically adjusting the learning rate using a cosine annealing scheduler is employed. This strategy fully utilizes the periodic restart mechanism of the cosine annealing scheduler, enabling the convolutional neural network in this scheme to escape local minima when approaching the optimal solution, thereby enhancing the parameter fine-tuning capability. In the task of radiation source identification, by introducing a cosine annealing scheduler (100 epochs period), the rate of decrease in the loss value during the training process of the convolutional neural network can be significantly improved, exhibiting fluctuation characteristics synchronized with the learning rate period. In particular, by adjusting the learning rate through the introduced cosine annealing scheduler, the local optimum problem caused by multi-source feature coupling can be effectively alleviated, significantly improving the performance of multi-source identification.

[0050] According to one aspect of the present invention, this approach achieves high-precision inversion of the number and spatial location of unknown radioactive sources, providing a new development approach for radiation field inversion and having significant application value for improving nuclear emergency response capabilities. Attached Figure Description

[0051] Figure 1 This is a step diagram of a multi-γ source orientation inversion method according to an embodiment of the present invention;

[0052] Figure 2This is a schematic diagram of a grid card according to one embodiment of the present invention, wherein, Figure 2 (a) is a schematic diagram of a geometric grid card. Figure 2 (b) is a schematic diagram of the material grid card;

[0053] Figure 3 This is a structural diagram of a convolutional neural network according to one embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of single-source radiation field measurement in one embodiment of the present invention;

[0055] Figure 5 This is a comparison diagram of the experimental measurement results of a single-source radiation field in one embodiment of the present invention and the results of this solution;

[0056] Figure 6 This is a schematic diagram of multi-source radiation field measurement in one embodiment of the present invention;

[0057] Figure 7 This is a comparison diagram of the experimental measurement results of multi-source radiation field in one embodiment of the present invention and the results of the present invention. Detailed Implementation

[0058] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The embodiments cannot be described in detail here, but the embodiments of the present invention are not limited to the following embodiments.

[0059] like Figure 1 As shown, according to one embodiment of the present invention, a multi-γ source orientation inversion method of the present invention includes the following steps:

[0060] S1. Construct the dataset using the Monte Carlo numerical simulation method;

[0061] S2. Establish a convolutional neural network for multi-γ source inversion; wherein, a cosine annealing scheduler is introduced to adjust the learning rate of the convolutional neural network to dynamically adjust the training process of the convolutional neural network, and a center loss function is introduced to reconstruct the loss function of the convolutional neural network to adjust the classification performance and convergence of the training process of the convolutional neural network.

[0062] S3. Train the convolutional neural network based on the dataset to obtain a multi-γ source orientation inversion model;

[0063] S4. Collect the samples to be inverted and input them into the multi-γ source orientation inversion model, and output the number and location of γ sources in the samples to be inverted based on the multi-γ source orientation inversion model.

[0064] like Figure 1As shown, according to one embodiment of the present invention, step S1, the step of constructing the dataset using the Monte Carlo numerical simulation method, includes:

[0065] S11. Establish a ground environment model of a preset size based on MCNP software. MCNP software is a general-purpose particle transport simulation platform capable of handling neutron, photon, and electron transport problems in complex three-dimensional geometric structures. Therefore, the use of MCNP software to construct the dataset can generate a large amount of high-precision simulation data, thus solving the problems of high data acquisition costs and long cycles. Furthermore, in the step of establishing a ground environment model of a preset size based on MCNP software, a 10m × 10m cement ground environment model is used.

[0066] S12. Arrange the detector array in the ground environment model; In this embodiment, the detector array is arranged in a uniform coverage manner in the ground environment model. For example, if 10,000 detectors are uniformly placed in the ground environment model (XY plane), the detectors are arranged with an interval of 0.1m × 0.1m. Of course, if the number of detectors is limited by cost and placement environment conditions, the number of detectors can be reduced to 2,500 detectors with an interval of 0.2m × 0.2m, 1,600 detectors with an interval of 0.25m × 0.25m, or 100 detectors with an interval of 1m × 1m. Thus, different detection accuracies can be achieved through different detector arrays, and the data accuracy of the dataset can be flexibly adjusted. In this embodiment, since the uniform distribution of the detectors makes their intervals fixed, the arrangement of the detectors can be used to perform matching grid division of the arrangement area, thereby matching the detection results of the detectors with the corresponding grid coordinates.

[0067] S13. Randomly arrange 1 to 4 gamma-ray sources in the standard atmospheric medium of the ground environment model; wherein the activity range of the gamma-ray sources is set between 1 Ci and 10 Ci; and the emission direction / location of the gamma-ray sources is isotropic.

[0068] S14. Geometric raster cards are used to define the spatial structure of the ground environment model, and material raster cards are used to configure the properties of different media; such as Figure 2 As shown in (a), the geometric raster card records the dimensions and material allocation of the three-dimensional region division in the ground environment model. For example, the dimensions are set to 10m × 10m, the ground material is configured as cement, and the environment is configured as air. Figure 2 As shown in (b), the material cell card records the element composition, density, and nuclide cross-section library from the MCNP DATA dataset using MCNP software.

[0069] S15. Based on MCNP software, simulate and record the motion trajectory and radiation field data of photoelectric effect, Compton scattering, and electron-electron pair effect in the ground environment model to generate a training dataset; wherein, the corresponding absorbed dose rate should also be recorded in the obtained dataset. In this embodiment, a dataset consisting of 210,000 sets of three-dimensional radiation field data can be constructed for subsequent training of convolutional neural networks; further, 200,000 sets of three-dimensional radiation field data are selected from the constructed dataset as training samples, and 10,000 sets of three-dimensional radiation field data are selected as test samples.

[0070] In this embodiment, the obtained three-dimensional radiation field data undergoes structured processing to facilitate the training of a convolutional neural network. Specifically, the absorbed dose rate in the dataset is normalized to construct a two-dimensional dose matrix, and the γ-ray source parameters are encoded into a label vector [x1, y1, a1, ..., x]. n ,y n ,a n By employing a gridded coordinate mapping strategy, the continuous coordinate regression problem is transformed into a discrete classification problem, achieving a unified algorithm for source quantity prediction and coordinate localization. Consequently, all data is serialized into a "dataset.pth" binary file for use as input to train a convolutional neural network. A FloatTensor from the PyTorch framework stores the dose matrix, and an IntTensor stores the label vector.

[0071] like Figure 3As shown, according to one embodiment of the present invention, in step S2, the step of establishing a convolutional neural network for multi-γ source inversion includes: an input layer, a feature extraction module connected to the input layer, a classification decision module connected to the feature extraction module, and an output layer connected to the classification decision module. In this embodiment, the input layer is used to receive samples, and its size can be set to 1×10×10 to represent that the input is a single-channel two-dimensional image or matrix of size 10×10. Further, the feature extraction module includes: several cascaded convolutional layers and max-pooling layers. In this embodiment, two convolutional layers are provided, and the max-pooling layers are configured in a one-to-one correspondence with the convolutional layers. In this implementation, the first convolutional layer and the first max-pooling layer can output features of 32×11×11. Specifically, with an input dimension of 1×10×10, using a 2×2 convolutional kernel with a stride of 1, and adding a white margin, the output dimension becomes 32×11×11. Similarly, the second convolutional layer and the first max-pooling layer can output features of 64×12×12. Again, with an input dimension of 32×11×11, using a 2×2 convolutional kernel with a stride of 1, and adding a white margin, the output dimension becomes 64×12×12. In this embodiment, each convolutional layer is followed by a ReLU activation function to extract the spatial features of the radiation field intensity distribution of samples in the dataset layer by layer, and the spatial dimension is progressively compressed through the max-pooling layer.

[0072] Furthermore, the classification decision module includes a Flatten layer and multiple fully connected layers. The Flatten layer is connected to the last max-pooling layer of the feature extraction module and is used to expand the features output by the feature extraction module. The Flatten layer uses the feature expansion output by the view function. There are three fully connected layers, with the first fully connected layer connected to the Flatten layer and the third fully connected layer connected to the output layer. In this embodiment, the first fully connected layer is used to flatten the features output by the feature extraction module into a 1×256 vector. Specifically, its input dimension is 1×9216, and it uses a fully connected function and tanh. The first fully connected layer uses an activation function and dropout regularization to achieve an output dimension of 1×256. The second fully connected layer reduces the output dimension of the first layer to 1×64. Specifically, its input dimension is 1×256, and it uses a fully connected function, wrapped with a softplus activation function and dropout regularization to achieve an output dimension of 1×64. The third fully connected layer reduces the output dimension of the second layer to 1×4 or 1×10 (depending on the scenario). Specifically, its input dimension is 1×64, and it uses a fully connected function to output a dimension of 1×4 or 1×10, yielding different results and corresponding probability values. In this implementation, the fully connected layer is configured with a tanh activation function to output the probability of the number of γ sources, and a softmax function to output the classification probability of the γ source coordinates.

[0073] According to one embodiment of the present invention, the learning rate is a core hyperparameter for training convolutional neural networks. A learning rate that is too small will lead to long training times and a tendency to get stuck in local optima; a learning rate that is too large will cause the learned parameters to change drastically, making it difficult to reach the optimal solution. Therefore, this scheme introduces a cosine annealing scheduler to optimize the training process through dynamic adjustment strategies, and verifies the generalization ability of the convolutional neural network using Monte Carlo simulation data. Thus, in step S2, the cosine annealing scheduler is introduced to adjust the learning rate of the convolutional neural network. In this step of dynamically adjusting the training process of the convolutional neural network, the learning rate gradually decays from its initial value to a preset minimum value according to a cosine curve; wherein the learning rate is expressed as:

[0074] ;

[0075] in, Indicates the learning rate. This represents the lower bound of the learning rate. Indicates the upper limit of the learning rate. Indicates the current training step count. This represents the total number of steps in a training cycle. In this implementation, the total number of steps in a training cycle... It can be set to 100 epochs. Additionally, the lower bound of the learning rate... It can be set to 0, which is the upper limit of the learning rate. It can be set to 0.001.

[0076] Through the above settings, a strategy of dynamically adjusting the learning rate using a cosine annealing scheduler is adopted. This strategy fully utilizes the periodic restart mechanism of the cosine annealing scheduler, enabling the convolutional neural network in this scheme to escape local minima when approaching the optimal solution, thus enhancing the parameter fine-tuning capability. In the task of radiation source identification, by introducing a cosine annealing scheduler (period of 100 epochs), the rate of decrease in the loss value during the training process of the convolutional neural network can be significantly improved, exhibiting fluctuation characteristics synchronized with the learning rate period. In particular, by adjusting the learning rate through the introduced cosine annealing scheduler, the local optimum problem caused by multi-source feature coupling can be effectively alleviated, significantly improving the performance of multi-source identification.

[0077] According to one embodiment of the present invention, in step S2, the reconstructed loss function in the step of establishing a convolutional neural network for multi-γ source inversion is a composite loss function that fuses the cross-entropy loss function and the center loss function; in this embodiment, the composite loss function is expressed as:

[0078] ;

[0079] ;

[0080] ;

[0081] in, Represents the composite loss function. Represents the cross-entropy loss function. Represents the central loss function. This represents the weighting coefficient that balances the two loss functions; it can be set to 0.1. Category labels representing coordinates, with values ​​ranging from 1 to 10. This indicates the number of classification categories, which includes 1 to 4 gamma radiation sources and their coordinates. This represents the probability values ​​corresponding to different classification labels predicted by the trained convolutional neural network. These represent the different classification labels predicted by the trained convolutional neural network. This represents the sample data in the input dataset. This represents the center value of the label, with values ​​ranging from 0.5, 1.5, 2.5…9.5. This indicates the total number of labels for the category, and can be set to 10 categories, with subscripts... This indicates the order of the tag categories, with values ​​ranging from 1 to 10.

[0082] According to one embodiment of the present invention, a multi-γ source orientation inversion device is provided, comprising: a dataset generation module, a convolutional neural network (CNN) construction module, a training module, and a recognition module; wherein, the dataset generation module is used to construct a dataset using a Monte Carlo numerical simulation method; the CNN construction module is used to establish a CNN for multi-γ source inversion; wherein, a cosine annealing scheduler is introduced to adjust the learning rate of the CNN to dynamically adjust the training process of the CNN, and a center loss function is introduced to reconstruct the loss function of the CNN to adjust the classification performance and convergence of the CNN training process; the training module trains the CNN based on the dataset to obtain a multi-γ source orientation inversion model; the recognition module is used to load the obtained multi-γ source orientation inversion model to collect samples to be inverted and input them into the multi-γ source orientation inversion model, and output the number and location of γ sources in the samples to be inverted based on the multi-γ source orientation inversion model.

[0083] Specific limitations regarding the multi-gamma source azimuth inversion device can be found in the limitations of the vector data extraction method based on high-fidelity scenes mentioned above, and will not be repeated here. Each module in the aforementioned multi-gamma source azimuth inversion device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0084] According to one embodiment of the present invention, an apparatus is provided, including at least one processor, at least one memory, and a data bus. In this embodiment, the processor and the memory communicate with each other via the data bus; wherein, the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the aforementioned multi-gamma source orientation inversion method.

[0085] In this embodiment, the memory may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc.

[0086] In this embodiment, the processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0087] In this embodiment, the device may also be equipped with a display screen and an input device. The display screen may be an LCD screen or an e-ink screen. The input device may be a touch layer covering the display screen, or a button, trackball, or touchpad set on the housing, or an external keyboard, touchpad, or mouse, etc.

[0088] According to one embodiment of the present invention, a storage medium is provided thereon storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned multi-gamma source orientation inversion method.

[0089] To further illustrate this plan, further examples will be provided.

[0090] Example

[0091] In this embodiment, a radiation field detection experiment is conducted in a laboratory to verify whether the error range between the radiation field data obtained by using this scheme and the real environment is within an acceptable range. The experimental steps involve measuring the dose rate under different scenarios in the laboratory, then using the multi-gamma source orientation inversion method of this scheme to perform simulations, and drawing conclusions by comparing the radiation field dose rates.

[0092] a) Single-source radiation field verification

[0093] Laboratory experiments were conducted in a shielded laboratory (background dose rate 0.10 μSv / h), using a RadEye G-10 detector to measure the activity of the Cs-137 radioactive source. The dose rate distribution of the Cs-137 radiation source and detector was measured at a linear interval of 20 cm to 500 cm (see [reference]). Figure 4 The average and peak dose rates (in μSv / h) at each distance point were recorded. In this embodiment, a RadEye G-10 detector was used to measure the activity of the Cs-137 radioactive source. During the dose rate distribution process, each measurement lasted 5 minutes and the interval between each measurement was 20 cm.

[0094] This scheme: A single-source model is constructed using MCNP software, and the photon energy is set to 0.661 MeV (corresponding to...). decay spectrum) The number of particles was calculated and converted into a single-particle dose rate, with the statistical error controlled within 1%.

[0095] like Figure 5 As shown, a comparison between the inversion results of this scheme and laboratory experimental data indicates that: at close to medium distances (20-300 cm): the error is <10%, and the inversion results of this scheme are in high agreement with the laboratory experimental data; at long distances (450 cm): the dose rate is as low as... At the scale of [scale value], background environmental interference caused a maximum error of 50%, which was optimized to within 20% after background noise filtering.

[0096] b) Verification of multi-source radiation fields

[0097] Laboratory experiment: High / low activity Cs-137 sources were arranged in a 250 cm × 250 cm area. and See also Figure 6 The activity of the Cs-137 radioactive source was measured using a RadEye G-10 detector. During the dose rate distribution process, each measurement lasted 3 minutes. Therefore, Latin hypercube sampling (LHS) was used to solve the sampling problem of 6 variables (source and detector coordinates). Each coordinate variable was divided into 10 layers (25 cm per layer). Ten samples were extracted, covering a uniform distribution in the parameter space. The sampling results are shown in Table 1. In Table 1, S1_x and S1_y represent the x and y coordinates of the first source (in cm); S2_x and S2_y represent the x and y coordinates of the second source (in cm); D_x and D_y represent the x and y coordinates of the detector (in cm); and stratification indicates the layer number in the Latin hypercube sampling, ranging from 1 to 10, without units.

[0098] Table 1. Results of Latin hypercube sampling

[0099]

[0100] This scheme: A dual-source model is constructed using MCNP software, with the photon energy set to 0.661 MeV (corresponding to...). decay spectrum) and maintenance The number of particles was calculated and converted into a single-particle dose rate, with the statistical error controlled within 1%.

[0101] like Figure 7As shown, the results of the inversion based on this scheme, when combined with 10 sets of experimental results from laboratory data, demonstrate consistency: the relative error for 90% of the data points is ≤10%; in high-fluctuation scenarios (such as experiments 3 / 4 / 9), the error remains stable at peak dose rates of 1.0-1.2 μSv / h. ;

[0102] Therefore, the comparative experimental results show that the error of this solution is <10% in the range of 20-300cm, and is optimized to 20% at long distances through noise filtering; LHS sampling achieves full coverage of the parameter space with 10 sets of experiments, and 90% of the data error is ≤10%, which meets the requirements of complex scenarios and proves that this solution has engineering applicability in single / multi-source scenarios.

[0103] The above description is merely an example of a specific solution of the present invention. For any devices and structures not described in detail herein, it should be understood that they are implemented using common devices and methods already available in the art.

[0104] The above description is merely one embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements 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 multi-γ source orientation inversion method, characterized in that, Includes the following steps: S1. Construct the dataset using the Monte Carlo numerical simulation method; S2. Establish a convolutional neural network for multi-γ source inversion; wherein, a cosine annealing scheduler is introduced to adjust the learning rate of the convolutional neural network to dynamically adjust the training process of the convolutional neural network, and a center loss function is introduced to reconstruct the loss function of the convolutional neural network to adjust the classification performance and convergence of the training process of the convolutional neural network. The convolutional neural network includes: an input layer, a feature extraction module connected to the input layer, a classification decision module connected to the feature extraction module, and an output layer connected to the classification decision module; The input layer has a size of 1×10×10; The feature extraction module includes: two cascaded convolutional layers and a max pooling layer; The max pooling layer and the convolutional layer are set to correspond one-to-one. The feature size output by the first convolutional layer and the corresponding max pooling layer is 32×11×11, and the feature size output by the second convolutional layer and the corresponding max pooling layer is 64×12×12. Each convolutional layer is followed by a ReLU activation function, which is used to extract the spatial features of the radiation field intensity distribution of the samples in the dataset layer by layer, and to compress the spatial dimension step by step through the max pooling layer; The classification decision module includes: a Flatten layer and multiple fully connected layers; The Flatten layer is connected to the last max pooling layer of the feature extraction module and is used to expand the features output by the feature extraction module. The fully connected layer has three layers, with the first fully connected layer connected to the Flatten layer and the third fully connected layer connected to the output layer. The first fully connected layer is used to flatten the features output by the feature extraction module into a 1×256 vector. The second fully connected layer is used to reduce the dimension of the output of the first fully connected layer to 1×64. The third fully connected layer is used to reduce the dimension of the output of the second fully connected layer to 1×4 or 1×10. The fully connected layer is configured with a tanh activation function to output the probability of the number of γ sources, and with a Softmax function to output the probability of the γ source coordinates classification. The learning rate gradually decreases from an initial value to a preset minimum value according to a cosine curve; wherein, the learning rate is expressed as: ； in, Indicates the learning rate. This represents the lower bound of the learning rate. Indicates the upper limit of the learning rate. Indicates the current training step count. This represents the total number of steps in the training cycle; The reconstructed loss function is a composite loss function that combines the cross-entropy loss function and the center loss function; The composite loss function is expressed as follows: ； ； ； in, Represents the composite loss function. Represents the cross-entropy loss function. Represents the central loss function. This represents the weighting coefficients that balance the two loss functions. Category labels representing coordinates, This indicates the number of classification categories, which includes 1 to 4 gamma radiation sources and their coordinates. This represents the probability values ​​corresponding to different classification labels predicted by the trained convolutional neural network. These represent the different classification labels predicted by the trained convolutional neural network. This represents the sample data in the input dataset. Indicates the center value of the label. The subscript indicates the total number of labels for each category. Indicates the order of label categories; S3. Train the convolutional neural network based on the dataset to obtain a multi-γ source orientation inversion model; S4. Collect the samples to be inverted and input them into the multi-γ source orientation inversion model, and output the number and location of γ sources in the samples to be inverted based on the multi-γ source orientation inversion model.

2. The multi-γ source orientation inversion method according to claim 1, characterized in that, Step S1, which involves constructing the dataset using the Monte Carlo numerical simulation method, includes: S11. Establish a ground environment model of preset dimensions based on MCNP software; S12. Arrange the detector array in the ground environment model; S13. Randomly arrange 1 to 4 γ-ray sources in the standard atmospheric medium of the ground environment model; wherein the activity range of the γ-ray sources is set between 1 Ci and 10 Ci; S14. Use geometric grid cards to define the spatial structure of the ground environment model, and use material grid cards to configure the properties of different media; S15. Based on the MCNP software, simulate and record the motion trajectory and radiation field data of photoelectric effect, Compton scattering and electron pair effect in the ground environment model to generate a dataset for training.

3. The multi-γ source orientation inversion method according to claim 2, characterized in that, In step S11, the step of establishing a ground environment model of a preset size based on MCNP software, the ground environment model adopts a 10m×10m cement ground environment model. In step S12, the detector array is arranged in a uniform coverage manner in the ground environment model.

4. A multi-gamma source orientation inversion device, characterized in that, include: The dataset generation module is used to construct datasets using the Monte Carlo numerical simulation method; A convolutional neural network building module is used to build a convolutional neural network for multi-γ source inversion; wherein, a cosine annealing scheduler is introduced to adjust the learning rate of the convolutional neural network to dynamically adjust the training process of the convolutional neural network, and a center loss function is introduced to reconstruct the loss function of the convolutional neural network to adjust the classification performance and convergence of the training process of the convolutional neural network. The convolutional neural network includes: an input layer, a feature extraction module connected to the input layer, a classification decision module connected to the feature extraction module, and an output layer connected to the classification decision module; The input layer has a size of 1×10×10; The feature extraction module includes: two cascaded convolutional layers and a max pooling layer; The max pooling layer and the convolutional layer are set to correspond one-to-one. The feature size output by the first convolutional layer and the corresponding max pooling layer is 32×11×11, and the feature size output by the second convolutional layer and the corresponding max pooling layer is 64×12×12. Each convolutional layer is followed by a ReLU activation function, which is used to extract the spatial features of the radiation field intensity distribution of the samples in the dataset layer by layer, and to compress the spatial dimension step by step through the max pooling layer; The classification decision module includes: a Flatten layer and multiple fully connected layers; The Flatten layer is connected to the last max pooling layer of the feature extraction module and is used to expand the features output by the feature extraction module. The fully connected layer has three layers, with the first fully connected layer connected to the Flatten layer and the third fully connected layer connected to the output layer. The first fully connected layer is used to flatten the features output by the feature extraction module into a 1×256 vector. The second fully connected layer is used to reduce the dimension of the output of the first fully connected layer to 1×64. The third fully connected layer is used to reduce the dimension of the output of the second fully connected layer to 1×4 or 1×10. The fully connected layer is configured with a tanh activation function to output the probability of the number of γ sources, and with a Softmax function to output the probability of the γ source coordinates classification. The learning rate gradually decreases from an initial value to a preset minimum value according to a cosine curve; wherein, the learning rate is expressed as: ； in, Indicates the learning rate. This represents the lower bound of the learning rate. Indicates the upper limit of the learning rate. Indicates the current training step count. This represents the total number of steps in the training cycle; The reconstructed loss function is a composite loss function that combines the cross-entropy loss function and the center loss function; The composite loss function is expressed as follows: ； ； ； in, Represents the composite loss function. Represents the cross-entropy loss function. Represents the central loss function. This represents the weighting coefficients that balance the two loss functions. Category labels representing coordinates, This indicates the number of classification categories, which includes 1 to 4 gamma radiation sources and their coordinates. This represents the probability values ​​corresponding to different classification labels predicted by the trained convolutional neural network. These represent the different classification labels predicted by the trained convolutional neural network. This represents the sample data in the input dataset. Indicates the center value of the label. The subscript indicates the total number of labels for each category. Indicates the order of label categories; The training module trains the convolutional neural network based on the dataset to obtain a multi-γ source orientation inversion model. The identification module is used to load the obtained multi-γ source orientation inversion model, collect samples to be inverted and input them into the multi-γ source orientation inversion model, and output the number and location of γ sources in the samples to be inverted based on the multi-γ source orientation inversion model.

5. A device, characterized in that, Includes at least one processor, at least one memory, and a data bus; The processor and the memory communicate with each other via the data bus; The memory stores program instructions that can be executed by the processor, which invokes the program instructions to execute the multi-γ source orientation inversion method according to any one of claims 1 to 3.

6. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-γ source orientation inversion method according to any one of claims 1 to 3.

Images