A marine gamma spectrum analysis method and system based on gan and pinn

Abstract

The application discloses a marine gamma spectrum analysis method and system based on GAN and PINN, and relates to the field of marine radioactivity monitoring. In view of the problem that the prior art cannot simultaneously improve the signal-to-noise ratio and physical consistency, marine gamma spectrum data is acquired, and a multi-level benchmark data set is constructed; a physical prior constraint term and a nuclide prior knowledge constraint term are determined according to the marine gamma spectrum data; a physical constraint generative adversarial network is constructed, and the physical prior constraint term is embedded into the physical constraint generative adversarial network; a physical information neural network is constructed, and the nuclide prior knowledge constraint term is embedded into the physical information neural network; the two networks are optimized by using a cooperative training strategy; marine gamma spectrum data to be analyzed is input into the physical constraint generative adversarial network which has been cooperatively optimized, and a physically fidelity-enhanced spectrum is obtained; the physically fidelity-enhanced spectrum is input into the physical information neural network which has been cooperatively optimized, and the nuclide type, nuclide activity and corresponding uncertainty are output, so that the signal-to-noise ratio and physical consistency can be simultaneously improved.

Description

Technical Field

[0001] This invention relates to the field of marine radioactivity monitoring technology, and in particular to a marine gamma-ray spectrum analysis method and system based on GAN and PINN. Background Technology

[0002] Gamma spectroscopy is a core technology for in-situ online monitoring of marine radioactivity, but the marine environment presents significant technical bottlenecks: strong attenuation of gamma rays by seawater and background interference from complex environments lead to low in-situ detection count rates and poor statistical performance of the energy spectrum; at the same time, the requirements for miniaturization and low power consumption of monitoring equipment limit the size and efficiency of detectors, further exacerbating problems such as spectral line overlap, weakening of characteristic peaks and low signal-to-noise ratio, posing a huge challenge to nuclide identification and quantitative activity inversion.

[0003] Existing spectral interpretation techniques have significant limitations: traditional methods (such as stripping and inverse matrix methods) rely on accurate detector response functions and background subtraction, resulting in slow convergence and spurious peak generation in low count rate scenarios, making it difficult to meet the requirements of online real-time processing; pure data-driven deep learning models (such as conventional GANs and CNNs) lack physical constraints and are prone to physical distortion problems such as peak shift and spurious peak generation; although single physical information neural networks (PINNs) can embed physical laws, their accuracy is limited when directly processing low-quality energy spectra, making it difficult to balance signal-to-noise ratio improvement and physical consistency.

[0004] Therefore, developing a marine gamma-ray spectrum analysis method and system based on GAN and PINN is of great significance for balancing physical fidelity and high-precision analysis capabilities. Summary of the Invention

[0005] To address the difficulty in balancing signal-to-noise ratio improvement and physical consistency in existing technologies, this invention provides a method for analyzing ocean gamma spectra based on GANs and PINNs, specifically including the following steps: S1. Acquire ocean gamma spectrum data and construct a multi-level benchmark dataset based on the ocean gamma spectrum data; S2. Determine the physical prior constraints and nuclide prior knowledge constraints based on ocean gamma-ray energy spectrum data; S3. Construct a physical constraint generative adversarial network based on the benchmark dataset, and embed the physical prior constraint terms into the physical constraint generative adversarial network; S4. Construct a physical information neural network based on the benchmark dataset, and embed the nuclide prior knowledge constraint terms into the physical information neural network; S5. A pre-defined collaborative training strategy is adopted to jointly optimize the physical constraint generative adversarial network and the physical information neural network, so that the physical fidelity enhancement spectrum output by the physical constraint generative adversarial network can be adapted to the analytical requirements of the physical information neural network. At the same time, the physical consistency constraint of the physical information neural network guides the parameter optimization of the physical constraint generative adversarial network in reverse. S6. Input the ocean gamma spectrum data to be analyzed into the physical constraint generative adversarial network after collaborative optimization to obtain the physical fidelity enhanced spectrum; input the physical fidelity enhanced spectrum into the physical information neural network after collaborative optimization to output the nuclide type, nuclide activity and corresponding uncertainty.

[0006] Furthermore, in S1, a multi-level benchmark dataset is constructed based on ocean gamma-ray spectrum data, including: establishing an ocean detection scenario model with seawater medium and shielding envelope surrounding the detector using Monte Carlo simulation technology, simulating key ocean nuclides, and designing multiple combinations of different nuclide activities; performing long-term and short-term simulations on each nuclide activity combination, with long-term simulation generating high statistical ground truth spectra and short-term simulation generating low count rate observation spectra with different count rates, forming multiple pairs of paired data of high statistical ground truth spectra and low count rate observation spectra; supplementing with laboratory standard source measured gamma-ray spectrum data and in-situ low count rate ocean gamma-ray spectrum data at sea, constituting a multi-level benchmark dataset, which includes a training set, a validation set, and a test set.

[0007] Furthermore, in S2, determining physical prior constraints and nuclide prior knowledge constraints based on ocean gamma spectrum data includes: extracting key physical laws of the energy spectrum based on the physical formation mechanism of the ocean gamma spectrum and the characteristics presented by the ocean gamma spectrum data, and transforming the key physical laws of the energy spectrum into differentiable mathematical constraint rules to determine physical prior constraints; physical prior constraints include peak position locking constraints, peak shape regularization constraints, and background smoothness constraints; determining nuclide prior knowledge constraints based on the inherent physical properties of key ocean nuclides and the detection characteristics of the detector; nuclide prior knowledge constraints include nuclide characteristic peak shape constraints, nuclide branching ratio constraints, and detector detection efficiency constraints for nuclides.

[0008] Furthermore, in S3, constructing a physical constraint generative adversarial network based on a benchmark dataset includes: designing the basic architecture of the physical constraint generative adversarial network, which includes a generator and a discriminator, and determining the core training parameters; using an adaptive momentum estimation optimizer to alternately train the generator and the discriminator according to a set alternating training strategy; and training the physical constraint generative adversarial network using paired data of high statistical ground truth spectrum and low count rate observation spectrum in the training set.

[0009] Furthermore, in S3, embedding the physical prior constraint terms into the physical constraint generative adversarial network includes: embedding the physical prior constraint terms into the generator total loss function of the physical constraint generative adversarial network.

[0010] Furthermore, in step S4, constructing a physical information neural network based on a benchmark dataset includes: designing the basic architecture of the physical information neural network, which is a nuclide activity inversion network, including a backbone network and an output layer, and determining the core training parameters; using an adaptive momentum estimation optimizer to train the physical information neural network according to a set iterative training strategy; and using the high statistical ground truth spectrum and corresponding nuclide activity and background parameter data in the training set to train the physical information neural network.

[0011] Furthermore, in step S4, embedding the nuclide prior knowledge constraint term into the physical information neural network includes: embedding the nuclide prior knowledge constraint term as a network initialization condition or a soft constraint term into the network structure and total loss function of the physical information neural network.

[0012] Furthermore, in S5, a preset collaborative training strategy is used to jointly optimize the physical constraint generative adversarial network and the physical information neural network, including: using a fine-tuning strategy for the generative adversarial network guided by the physical information neural network, independently training the physical constraint generative adversarial network and the physical information neural network based on a multi-level benchmark dataset, fixing the parameters of the physical information neural network, embedding physical consistency constraints into the generator loss function of the physical constraint generative adversarial network, and fine-tuning the generator; or, using an end-to-end joint training strategy, constructing a joint network containing a differentiable physical model layer, the physical model layer being used to realize the calculation of the linear superposition model of the γ-energy spectrum, alternately fixing the parameters of any network in the physical constraint generative adversarial network and the physical information neural network, and training the other network using the total loss of the corresponding network respectively, to achieve mutual adaptation and collaborative optimization of the two networks.

[0013] Furthermore, the physical consistency constraint of the physical information neural network is achieved by backpropagating the loss corresponding to the physical consistency constraint to the physical constraint generative adversarial network, adjusting its generator parameters, so that the output physical fidelity enhancement spectrum is adapted to the spectral resolution requirements of the physical information neural network.

[0014] This invention also provides a marine gamma-ray spectrometry analysis system based on GAN and PINN, the system being used to execute the marine gamma-ray spectrometry analysis method based on GAN and PINN described in any of the preceding claims, the system comprising: The dataset construction module is used to acquire ocean gamma spectrum data and construct a multi-level benchmark dataset based on the ocean gamma spectrum data; The constraint determination module is used to determine physical prior constraints and nuclide prior knowledge constraints based on ocean gamma-ray spectrum data. The adversarial network building module is used to build a physical constraint generative adversarial network based on a benchmark dataset and embed physical prior constraint terms into the physical constraint generative adversarial network; The neural network adversarial module is used to construct a physical information neural network based on a benchmark dataset and embed nuclide prior knowledge constraints into the physical information neural network; The joint optimization module is used to jointly optimize the physical constraint generative adversarial network and the physical information neural network using a preset collaborative training strategy, so that the physical fidelity enhancement spectrum output by the physical constraint generative adversarial network can be adapted to the parsing requirements of the physical information neural network. At the same time, the physical consistency constraints of the physical information neural network guide the parameter optimization of the physical constraint generative adversarial network in reverse. The analysis module is used to input the ocean gamma spectrum data to be analyzed into a collaboratively optimized physical constraint generative adversarial network to obtain a physical fidelity enhanced spectrum; the physical fidelity enhanced spectrum is then input into a collaboratively optimized physical information neural network to output the nuclide type, nuclide activity and corresponding uncertainty.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires ocean gamma-ray spectrum data and constructs a multi-level benchmark dataset based on this data. It then determines physical prior constraints and nuclide prior knowledge constraints based on the ocean gamma-ray spectrum data. A physical constraint generative adversarial network (GAN) is constructed based on the benchmark dataset, and the physical prior constraints are embedded into the GAN. A physical information neural network (PIN) is constructed based on the benchmark dataset, and the nuclide prior knowledge constraints are embedded into the PIN. A pre-defined collaborative training strategy is used to jointly optimize the GAN and the PIN, ensuring that the physical fidelity enhanced spectrum output by the GAN adapts to the analytical requirements of the PIN. Simultaneously, the physical consistency constraints of the PIN guide the parameter optimization of the GAN. The ocean gamma-ray spectrum data to be analyzed is input into the collaboratively optimized GAN to obtain the physical fidelity enhanced spectrum. This enhanced spectrum is then input into the collaboratively optimized PIN, which outputs the nuclide type, nuclide activity, and corresponding uncertainty. Through the deep integration of the GAN and the PIN, complementary advantages and synergistic effects are achieved. On the one hand, the physical constraint generative adversarial network embeds physical prior constraints to ensure that the output of the ocean gamma spectrum enhancement results has both fidelity and physical rationality, providing a high-quality data foundation for subsequent analysis. On the other hand, the physical information neural network embeds nuclide prior knowledge constraints to improve the targeting and accuracy of nuclide identification and activity inversion. The pre-set collaborative training strategy allows the two networks to adapt to each other and dynamically optimize during training through bidirectional interaction. This avoids the physical distortion or analytical bias that may occur when a single network works independently, and also builds an efficient link from the original energy spectrum to the final analytical result. Ultimately, it achieves a comprehensive and accurate output of nuclide types, nuclide activities and corresponding uncertainties, taking into account both analytical efficiency and result reliability. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a marine gamma-ray spectrum analysis method based on GAN and PINN provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a marine gamma-ray spectrum analysis system based on GAN and PINN provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] The specific embodiments of the present invention will be described below.

[0020] To address the challenge of balancing signal-to-noise ratio (SNR) improvement with physical consistency in existing technologies, this invention acquires ocean gamma-ray spectrum data and constructs a multi-level benchmark dataset. Based on the ocean gamma-ray spectrum data, it determines physical prior constraints and nuclide prior knowledge constraints. A physical constraint generative adversarial network (GAN) is constructed, embedding the physical prior constraints into the GAN. A physical information neural network (PIN) is constructed, embedding the nuclide prior knowledge constraints into the PIN. A collaborative training strategy is used to jointly optimize the two networks. The ocean gamma-ray spectrum data to be analyzed is input into the collaboratively optimized GAN to obtain a physically enhanced spectrum. This physically enhanced spectrum is then input into the collaboratively optimized PIN to output the nuclide type, nuclide activity, and corresponding uncertainty, thus achieving a balance between SNR improvement and physical consistency.

[0021] Example 1 This invention provides a method for analyzing marine gamma spectra based on GANs and PINNs. Figure 1 This is a flowchart of a marine gamma-ray spectrum analysis method based on GAN and PINN provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the specific steps include the following: S1. Obtain ocean gamma spectrum data and construct a multi-level benchmark dataset based on the ocean gamma spectrum data.

[0022] Ocean gamma-ray spectral data refers to gamma-ray spectral data acquired by detectors in the in-situ marine environment, and also includes simulated data generated by Monte Carlo simulations and laboratory measured data. Gamma-ray spectral data includes spectral signals from both natural and artificial nuclides; for example, natural nuclides include... 40 K, 238 U series 232 Th series, artificial nuclides include 137 Cs、 60 Co. γ-ray energy spectrum data are characterized by low count rate, large statistical fluctuations, low signal-to-noise ratio, spectral line overlap, and significant background interference.

[0023] Specifically, a multi-level benchmark dataset is constructed based on ocean gamma-ray spectrum data, including: establishing an ocean detection scenario model with seawater medium and shielding surrounding the detector using Monte Carlo simulation technology, simulating key ocean nuclides, and designing multiple combinations of different nuclide activities; performing long-term and short-term simulations for each nuclide activity combination, with long-term simulations generating high statistical ground truth spectra and short-term simulations generating low count rate observation spectra with different count rates, forming multiple pairs of paired high statistical ground truth spectra and low count rate observation spectra; supplementing with laboratory standard source measured gamma-ray spectrum data and in-situ low count rate ocean gamma-ray spectrum data, forming a multi-level benchmark dataset, which includes a training set, a validation set, and a test set.

[0024] Monte Carlo simulation is a numerical computation method that simulates physical processes through random sampling. It's used to construct ocean exploration scenarios and simulate gamma-ray transport and detection processes. Commonly used tools include Geant4 Monte Carlo particle transport simulation tool or MCNP Monte Carlo N-particle transport program. Nuclide activity combinations are nuclide activity pairing schemes designed to cover typical marine environments; for example, designing 1000 different activity combinations, with individual nuclide activities ranging from 0.1 Bq / L to 100 Bq / L. Long-duration simulation refers to emitting ≥10... 9 The simulation process for each particle aims to generate a highly statistical, near-noise-free true spectrum. Short-time simulation refers to simulating the process at different measurement times to generate low-count-rate observation spectra. Laboratory standard source measured gamma-ray spectrum data, obtained through measurements using a laboratory standard radioactive source, supplements the dataset's authenticity. In-situ low-count-rate marine gamma-ray spectrum data refers to low-count-rate spectrum data directly acquired by detectors in the ocean, reproducing the spectrum characteristics of real ocean exploration scenarios. The multi-level benchmark dataset is a structured dataset built for model training, validation, and testing, containing three levels: training set, validation set, and test set, supporting supervised training and performance evaluation of the dual networks.

[0025] Using Geant4 or MCNP tools, a marine detection scenario model was established, incorporating seawater and a shielding enclosure around the detector, to recreate the physical environment for gamma-ray detection. Key marine nuclides were selected, and 1000 different nuclide activity combinations were designed, covering a typical marine environment range from 0.1 Bq / L to 100 Bq / L. Two simulations were performed for each activity combination: a long-term simulation generated high-statistical true spectra and recorded precise activities; a short-term simulation generated low-count-rate observation spectra at 5 cps, 10 cps, and 20 cps, resulting in at least 3000 paired data sets. Laboratory-measured gamma-ray energy spectrum data from standard sources and in-situ low-count-rate energy spectrum data from the sea were incorporated to compensate for the lack of realism in purely simulated data. The integrated data were divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively, for dual-network training, training process validation, and final performance testing.

[0026] The multi-level benchmark dataset covers various nuclide combinations, activity ranges, and count rate levels, meeting the supervised training requirements for physical constraints and nuclide activity inversion. This ensures that the model learns the core physical laws and data characteristics of the energy spectrum, providing high-quality data support for dual-network training. Paired data provides a reference for dual-network collaborative training, and experimental data ensures the effectiveness of the model in real marine environments, guaranteeing the reliability of subsequent collaborative optimization and performance evaluation.

[0027] S2. Determine physical prior constraints and nuclide prior knowledge constraints based on ocean gamma-ray spectrum data. Specifically, this includes: extracting key physical laws of the energy spectrum based on the physical formation mechanism of the ocean gamma-ray spectrum and the characteristics presented by the ocean gamma-ray spectrum data, and transforming these key physical laws into differentiable mathematical constraint rules to determine physical prior constraints; these physical prior constraints include peak position locking constraints, peak shape regularization constraints, and background smoothness constraints; and determining nuclide prior knowledge constraints based on the inherent physical properties of key ocean nuclides and the detection characteristics of the detector, including nuclide characteristic peak shape constraints, nuclide branching ratio constraints, and detector detection efficiency constraints for nuclides.

[0028] The physical prior constraints are a set of differentiable mathematical constraints based on the key physical laws governing the physical formation mechanism and feature extraction of gamma spectrum data. These constraints include peak position locking constraints, peak shape regularization constraints, and background smoothness constraints, ensuring the physical rationality of the spectrum enhancement process. The nuclide prior knowledge constraints are a set of constraints determined by the inherent physical properties of key marine nuclides and the detector's detection characteristics. These include nuclide characteristic peak shape constraints, nuclide branching ratio constraints, and detector efficiency constraints for nuclides, used to optimize the nuclide activity inversion process. The physical formation mechanism of the marine gamma spectrum refers to the process by which gamma rays propagate in the marine environment, interact with the seawater medium, and form an energy spectrum after detector response. This includes core laws such as detector response functions and linear superposition of energy spectra. Key physical laws governing the energy spectrum refer to the inherent characteristics followed during gamma spectrum formation, including fixed characteristic peak addresses, peak shapes conforming to the detector's theoretical response, and continuous background smoothness. Differentiable mathematical constraints transform physical laws into mathematical expressions recognizable by deep learning models, support gradient backpropagation, and can be embedded with loss functions to guide model training. Peak position locking constraints ensure that the characteristic peak addresses of known nuclides remain unchanged during energy spectrum enhancement. Peak shape regularization constraints make the enhanced characteristic peak shape approximate the detector's theoretical response function, suppressing non-physical sawtooth oscillations. Background smoothness constraints ensure the physical rationality and smoothness of the continuous portion of the energy spectrum background, reducing background interference. The inherent physical properties of key marine nuclides refer to the inherent characteristics of typical nuclides in the marine environment, such as characteristic peak shapes and branching ratios. Detector detection characteristics refer to the detector's response efficiency to gamma rays from different nuclides, i.e., detection efficiency, which affects the intensity characterization of nuclide signals.

[0029] This study elucidates the physical formation mechanism of marine gamma-ray spectra, and, considering the characteristics of low count rate, high noise, and overlapping spectral lines in marine gamma-ray spectra data, extracts key physical laws such as fixed peak positions, peak shapes conforming to theoretical responses, and continuous smooth background. These laws are then transformed into differentiable mathematical constraints, clarifying the specific expressions for peak position locking constraints, peak shape regularization constraints, and background smoothness constraints, forming physical prior constraint terms. Based on the inherent physical properties of key marine nuclides and the detection efficiency characteristics of detectors for different nuclides, the specific content of nuclide characteristic peak shape constraints, nuclide branching ratio constraints, and detection efficiency constraints is directly determined, forming nuclide prior knowledge constraint terms.

[0030] S3. Construct a physical constraint generative adversarial network based on the benchmark dataset, and embed the physical prior constraint terms into the physical constraint generative adversarial network.

[0031] Physically Constrained Generative Adversarial Networks (PC-GANs) are network models based on generative adversarial networks, embedding physical prior constraints on the gamma spectrum. They include generators and discriminators, and are primarily used for physical fidelity enhancement of the ocean gamma spectrum at low count rates.

[0032] Specifically, a physical constraint generative adversarial network (GAN) is constructed based on a benchmark dataset, including: designing the basic architecture of the GAN, which includes a generator and a discriminator, and determining the core training parameters; using an adaptive momentum estimation optimizer to alternately train the generator and discriminator according to a set alternating training strategy; and training the GAN using paired data of high statistical ground truth spectrum and low count rate observation spectrum in the training set. Embedding physical prior constraints into the GAN includes: embedding the physical prior constraints into the generator's total loss function of the GAN.

[0033] The generator is the core enhancement module of PC-GAN, employing a U-Net architecture adapted to one-dimensional energy spectrum data. It includes an encoder, bottleneck layer, decoder, and final layer. The input is a low-count-rate observed spectrum, and the output is a physically accurate enhanced spectrum. The discriminator is the PC-GAN's spectrum authenticity detection module, using a PatchGAN architecture. It challenges the generator by judging the authenticity of local energy spectrum fragments, guiding the generator to output an energy spectrum that more closely conforms to physical laws. Core training parameters refer to the key parameters that need to be determined during training, including the optimizer parameters (learning rate, loss function weights, training iteration strategy, etc.). The adaptive momentum estimation optimizer, also known as the Adam optimizer, optimizes network parameters by adaptively adjusting the learning rate and is used for training the PC-GAN generator and discriminator. The alternating training strategy is the core training method of PC-GAN, where the generator is trained once after each discriminator training iteration, iterating until the network converges, achieving dynamic adversarial and collaborative optimization between the generator and discriminator. The generator total loss function is a comprehensive loss formula that guides generator training. It integrates adversarial loss, content reconstruction loss, and physical constraint loss to ensure that the generator output has both high signal-to-noise ratio and physical realism.

[0034] First, the basic architecture of the physically constrained generative adversarial network (GAN) is designed. The generator adopts a spectral enhancement U-Net structure adapted to one-dimensional energy spectrum data, including an encoder, a bottleneck layer with self-attention mechanism, a decoder, and a final layer responsible for outputting physically faithful enhanced spectra. The discriminator adopts the PatchGAN architecture, using five convolutional layers to distinguish between genuine and fake local segments of the energy spectrum. Simultaneously, core training parameters are determined, including the selection of the Adam optimizer and setting the generator learning rate to 2×10⁻⁶. -4 The discriminator learning rate is set to 1×10. -4The Adam optimizer parameters were set to β1=0.5 and β2=0.999, along with initial weights for adversarial loss, L1 loss, and physical constraint loss. Subsequently, a progressive training strategy was employed, alternating between training the discriminator once and the generator once. This was done using paired data from the training set, consisting of high-statistic ground truth spectra and low-count-rate observation spectra. Training began with higher-count-rate data, gradually introducing lower-count-rate data at 5cps, 10cps, and 20cps. Finally, physical prior constraints such as peak locking, peak regularization, and background smoothness were transformed into corresponding physical constraint loss functions and embedded into the generator's total loss function, thus completing the construction of the PC-GAN.

[0035] By adapting the U-Net and PatchGAN architectures to one-dimensional energy spectra, employing an alternating training strategy, and combining supervised learning with paired data, PC-GAN can accurately grasp the enhancement mapping law from low count rate spectra to high statistical ground truth spectra, effectively improving the energy spectrum signal-to-noise ratio and feature integrity. Embedding physical prior constraints into the generator's total loss function rigorously suppresses non-physical distortions such as false peak generation, peak shift, and sawtooth oscillations in the peak region, ensuring that the enhanced spectrum conforms to the physical laws of γ-energy spectrum formation and possesses high physical fidelity. The dual guarantee of progressive training and physical constraints allows the network to stably adapt to challenging scenarios such as low count rates, strong background interference, and complex nuclide combinations in in-situ ocean environments, reducing the risk of overfitting. The generated physically fidelity enhanced spectrum provides high-quality data input for subsequent nuclide activity inversion in the Physical Information Neural Network (PINN).

[0036] S4. Construct a physical information neural network based on the benchmark dataset, and embed the nuclide prior knowledge constraint terms into the physical information neural network.

[0037] Physical Information Neural Network (PINN) is a type of neural network that integrates physical models and data-driven approaches. It takes the energy spectrum enhanced by generative adversarial networks as input and the nuclide types, nuclide activities, and background parameters as outputs. Its core is to embed the physical model of linear superposition of the γ-energy spectrum into the network structure or loss function, thus combining data learning capabilities with physical consistency.

[0038] Specifically, a physical information neural network (PIN) is constructed based on a benchmark dataset, including: designing the basic architecture of the PIN, which is a nuclide activity inversion network, including a backbone network and an output layer, and determining the core training parameters; training the PIN using an adaptive momentum estimation optimizer according to a set iterative training strategy; and training the PIN using the high statistical ground truth spectrum and corresponding nuclide activity and background parameter data from the training set. Nuclide prior knowledge constraints are embedded into the PIN, including: using nuclide prior knowledge constraints as network initialization conditions or soft constraints, and embedding them into the network structure and total loss function of the PIN.

[0039] The backbone network is the core feature extraction module of PINN, using ReLU activation function to extract high-level features from the energy spectrum. The output layer has M+4 nodes, where M is the number of nuclides to be solved, and outputs four background polynomial coefficients. The network initialization conditions are the initial parameter states set before network training, incorporating prior knowledge of the nuclides into the initial parameters to guide the network to converge quickly to a physically reasonable solution space. Soft constraints refer to flexible constraints embedded in the loss function, adjusting their impact on training through weights, ensuring the network follows the physical laws of the nuclides while learning data features.

[0040] The basic architecture of the Physical Information Neural Network (PINN) is designed. This network is a nuclide activity inversion network, with a backbone network consisting of a 5-layer fully connected network (FCN). The number of nodes in each layer is [1024, 512, 256, 128, 64], and the activation function is ReLU, responsible for extracting high-level features of the energy spectrum. The output layer uses the Softplus activation function to ensure that the nuclide activity value is non-negative. Core training parameters are determined, including the parameters of the Adam adaptive momentum estimation optimizer, as well as the weights of each term in the loss function and the number of iterations. An iterative training strategy is then adopted, using the high statistical ground truth spectrum and corresponding nuclide activity and background parameters from the benchmark dataset training set as input. The Adam optimizer minimizes the total loss function, iteratively updating the network parameters to complete the training of PINN. Simultaneously, prior knowledge constraints such as nuclide feature peak shape, branch ratio, and detector detection efficiency are incorporated as both initial parameters matching physical characteristics and soft constraints into the PINN network structure and total loss function, achieving deep guidance of training by physical priors.

[0041] S5. A pre-defined collaborative training strategy is adopted to jointly optimize the physical constraint generative adversarial network and the physical information neural network, so that the physical fidelity enhancement spectrum output by the physical constraint generative adversarial network can be adapted to the analytical requirements of the physical information neural network. At the same time, the physical consistency constraint of the physical information neural network guides the parameter optimization of the physical constraint generative adversarial network in reverse.

[0042] Specifically, a pre-defined collaborative training strategy is employed to jointly optimize the Physically Constrained Generative Adversarial Network (GAN) and the Physically Information Neural Network (PIN). This includes: using a fine-tuning strategy for the GAN guided by the PIN, independently training the GAN and PIN based on a multi-level benchmark dataset, fixing the parameters of the PIN, embedding physical consistency constraints into the generator loss function of the GAN, and fine-tuning the generator; or, using an end-to-end joint training strategy, constructing a joint network containing a differentiable physical model layer, where the physical model layer is used to implement the linear superposition model calculation of the γ-ray spectrum, alternately fixing the parameters of any network in the GAN and PIN, and training the other network using the total loss of the corresponding network, achieving mutual adaptation and collaborative optimization between the two networks. The physical consistency constraints of the PIN are implemented by backpropagating the loss corresponding to the physical consistency constraints to the GAN, adjusting its generator parameters to adapt the output physically faithful enhanced spectrum to the spectral interpretation requirements of the PIN.

[0043] Both collaborative training strategies construct an optimization mechanism for bidirectional flow of physical information, overcoming the limitations of simple model concatenation. The PINN-guided fine-tuning strategy allows PC-GAN to accurately adapt to the backend spectral interpretation requirements, ensuring the physical fidelity and spectral adaptability of the enhanced spectrum. End-to-end joint training aligns the physical objectives of the two networks from the initial training stage, improving overall optimization efficiency. Through collaborative optimization with shared physical objectives, the two networks not only guarantee the physical authenticity of the PC-GAN enhanced spectrum but also enable PINN to effectively utilize the enhanced data to stably invert nuclide activity. This significantly improves the accuracy, robustness, and physical interpretability of low-count-rate ocean gamma spectrum analysis, adapting to complex ocean scenarios such as extremely low signal-to-noise ratios and nuclide overlap, providing reliable support for end-to-end automated spectral interpretation.

[0044] S6. Input the ocean gamma spectrum data to be analyzed into the physical constraint generative adversarial network after collaborative optimization to obtain the physical fidelity enhanced spectrum; input the physical fidelity enhanced spectrum into the physical information neural network after collaborative optimization to output the nuclide type, nuclide activity and corresponding uncertainty.

[0045] First, the acquired marine gamma-ray spectrum data to be analyzed is preprocessed, such as by normalization and automatic energy scale correction. The preprocessed marine gamma-ray spectrum data is then input into the co-optimized PC-GAN. PC-GAN, leveraging embedded physical prior constraints such as peak position locking, peak shape regularization, and background smoothness, as well as parameters co-optimized with PINN, performs denoising, characteristic peak recovery, and background smoothing on the original spectrum with low count rate and high noise, outputting a physically enhanced spectrum with both high signal-to-noise ratio and physical fidelity. Subsequently, this physically enhanced spectrum is input into the co-optimized PINN. PINN, constrained by embedded prior knowledge such as nuclide characteristic peak shape, branching ratio, and detection efficiency, combined with a linear superposition physical model of gamma-ray spectra, uses automatic differentiation technology to invert nuclide types and activities, and quantifies the uncertainty of the dispersive spectral results using Monte Carlo Dropout technology, finally outputting nuclide types, activities, and corresponding confidence intervals.

[0046] The co-optimized PC-GAN ensures the physical fidelity of the enhanced spectrum, effectively avoiding non-physical distortions such as false peak generation and peak shift, providing a high-quality data foundation for PINN spectral analysis, and significantly improving the accuracy of nuclide identification and activity inversion, especially overcoming the challenge of resolving overlapping peak regions of nuclides under low signal-to-noise ratio conditions. The end-to-end analysis process formed by the dual-network collaboration eliminates the need for tedious manual intervention in background subtraction, peak finding, and other cumbersome steps, achieving a high degree of automation. It is well-suited to the real-time requirements of in-situ marine monitoring, overcoming the limitations of poor spatiotemporal resolution and delayed emergency response in traditional offline analysis. Simultaneously, the uncertainty of the PINN output ensures the verifiability and reliability of the spectral results. The collaborative adaptability of PC-GAN and PINN enables the system to stably cope with complex scenarios such as strong background interference and count rate fluctuations in the ocean, providing high-precision and robust technical support for in-situ marine radioactive monitoring and nuclear emergency decision-making.

[0047] Example 2 This invention also provides a marine gamma-ray energy spectrum analysis system based on GAN and PINN. Figure 2 This is a schematic diagram of the structure of a marine gamma-ray spectroscopy analysis system based on GAN and PINN provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the system is used to perform the method described in any of the above-mentioned methods, including: The dataset construction module is used to acquire ocean gamma spectrum data and construct a multi-level benchmark dataset based on the ocean gamma spectrum data; The constraint determination module is used to determine physical prior constraints and nuclide prior knowledge constraints based on ocean gamma-ray spectrum data. The adversarial network building module is used to build a physical constraint generative adversarial network based on a benchmark dataset and embed physical prior constraint terms into the physical constraint generative adversarial network; The neural network adversarial module is used to construct a physical information neural network based on a benchmark dataset and embed nuclide prior knowledge constraints into the physical information neural network; The joint optimization module is used to jointly optimize the physical constraint generative adversarial network and the physical information neural network using a preset collaborative training strategy, so that the physical fidelity enhancement spectrum output by the physical constraint generative adversarial network can be adapted to the parsing requirements of the physical information neural network. At the same time, the physical consistency constraints of the physical information neural network guide the parameter optimization of the physical constraint generative adversarial network in reverse. The analysis module is used to input the ocean gamma spectrum data to be analyzed into a collaboratively optimized physical constraint generative adversarial network to obtain a physical fidelity enhanced spectrum; the physical fidelity enhanced spectrum is then input into a collaboratively optimized physical information neural network to output the nuclide type, nuclide activity and corresponding uncertainty.

[0048] The ocean gamma spectral analysis system based on GAN and PINN provided in this embodiment performs the ocean gamma spectral analysis method based on GAN and PINN described in any of the above embodiments, and has the beneficial effects of any of the above embodiments, which will not be repeated here.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for analyzing marine gamma spectra based on GAN and PINN, characterized in that, include: S1. Acquire ocean gamma spectrum data and construct a multi-level benchmark dataset based on the ocean gamma spectrum data; S2. Determine the physical prior constraints and nuclide prior knowledge constraints based on ocean gamma-ray energy spectrum data; S3. Construct a physical constraint generative adversarial network based on the benchmark dataset, and embed the physical prior constraint terms into the physical constraint generative adversarial network; S4. Construct a physical information neural network based on the benchmark dataset, and embed the nuclide prior knowledge constraint terms into the physical information neural network; S5. A pre-defined collaborative training strategy is adopted to jointly optimize the physical constraint generative adversarial network and the physical information neural network, so that the physical fidelity enhancement spectrum output by the physical constraint generative adversarial network can be adapted to the analytical requirements of the physical information neural network. At the same time, the physical consistency constraint of the physical information neural network guides the parameter optimization of the physical constraint generative adversarial network in reverse. S6. Input the ocean gamma spectrum data to be analyzed into the physical constraint generative adversarial network after collaborative optimization to obtain the physical fidelity enhanced spectrum; input the physical fidelity enhanced spectrum into the physical information neural network after collaborative optimization to output the nuclide type, nuclide activity and corresponding uncertainty. In step S5, a preset collaborative training strategy is used to jointly optimize the physical constraint generative adversarial network and the physical information neural network, including: A fine-tuning strategy for generative adversarial networks (GANs) guided by a physical information neural network (PIN) is adopted. The physical constraint GAN and the physical information neural network are independently trained based on a multi-level benchmark dataset. The parameters of the physical information neural network are fixed, and the physical consistency constraint is embedded into the generator loss function of the physical constraint GAN to fine-tune the generator. or, An end-to-end joint training strategy is adopted to construct a joint network containing a differentiable physical model layer. The physical model layer is used to realize the calculation of the linear superposition model of the γ-energy spectrum. The parameters of any network in the physical constraint generative adversarial network and the physical information neural network are alternately fixed. The total loss of the corresponding network is used to train the other network, thereby realizing the mutual adaptation and collaborative optimization of the two networks.

2. The marine gamma-ray spectrum analysis method based on GAN and PINN according to claim 1, characterized in that, In S1, a multi-level benchmark dataset is constructed based on ocean gamma-ray spectrum data, including: Monte Carlo simulation technology was used to establish a marine detection scenario model that included a seawater medium and a shield enclosing the detector. This model simulated key marine nuclides and designed multiple combinations of different nuclide activities. Key marine nuclides included natural nuclides. 40 K, 238 U series 232 Th series and artificial nuclides 137 Cs、 60 Co; For each nuclide activity combination, long-term simulation and short-term simulation were performed respectively. Long-term simulation generated high statistical true value spectrum, and short-term simulation generated low count rate observation spectrum with different count rates, forming multiple pairs of paired data of high statistical true value spectrum and low count rate observation spectrum; Supplementing the laboratory standard source measured gamma spectrum data and the in-situ low count rate ocean gamma spectrum data at sea, a multi-level benchmark dataset is constructed, which includes a training set, a validation set, and a test set.

3. The marine gamma-ray energy spectrum analysis method based on GAN and PINN according to claim 1, characterized in that, In step S2, physical prior constraints and nuclide prior knowledge constraints are determined based on ocean gamma-ray energy spectrum data, including: Based on the physical formation mechanism of ocean gamma spectrum and the characteristics presented by ocean gamma spectrum data, key physical laws of the spectrum are extracted and transformed into differentiable mathematical constraint rules to determine physical prior constraint terms. The physical prior constraint terms include peak position locking constraint, peak shape regularity constraint, and background smoothness constraint. The prior knowledge constraints of nuclides are determined based on the inherent physical properties of key marine nuclides and the detection characteristics of detectors. These constraints include constraints on the characteristic peak shape of nuclides, the branching ratio of nuclides, and the detection efficiency of detectors for nuclides.

4. The marine gamma-ray energy spectrum analysis method based on GAN and PINN according to claim 2, characterized in that, In step S3, a physically constrained generative adversarial network is constructed based on a benchmark dataset, including: Design the physical constraint generative adversarial network infrastructure, which includes a generator and a discriminator, and determine the core training parameters; An adaptive momentum estimation optimizer is used to train the generator and discriminator alternately according to a set alternating training strategy; A physically constrained generative adversarial network is trained using paired data of high statistical ground truth spectrum and low count rate observation spectrum in the training set.

5. The marine gamma-ray spectrum analysis method based on GAN and PINN according to claim 4, characterized in that, In step S3, embedding the physical prior constraint terms into the physical constraint generative adversarial network includes: embedding the physical prior constraint terms into the generator total loss function of the physical constraint generative adversarial network.

6. The ocean gamma-ray energy spectrum analysis method based on GAN and PINN according to claim 1, characterized in that, In step S4, constructing a physical information neural network based on a benchmark dataset includes: Design the physical information neural network infrastructure, which is a nuclide activity inversion network, including a backbone network and an output layer, and determine the core training parameters; An adaptive momentum estimation optimizer is used to train the physical information neural network according to a set iterative training strategy; The physical information neural network is trained using the high statistical ground truth spectrum and the corresponding nuclide activity and background parameter data in the training set.

7. The marine gamma-ray spectrum analysis method based on GAN and PINN according to claim 6, characterized in that, In step S4, embedding the nuclide prior knowledge constraint term into the physical information neural network includes: using the nuclide prior knowledge constraint term as a network initialization condition or a soft constraint term, and embedding it into the network structure and total loss function of the physical information neural network.

8. The marine gamma-ray spectrum analysis method based on GAN and PINN according to claim 1, characterized in that, The physical consistency constraint of the physical information neural network is achieved in the following way: The loss corresponding to the physical consistency constraint is backpropagated to the physical constraint generative adversarial network, and its generator parameters are adjusted so that the output physical fidelity enhanced spectrum is adapted to the spectral resolution requirements of the physical information neural network.

9. A marine gamma-ray energy spectrum analysis system based on GAN and PINN, characterized in that, The system is used to execute the ocean gamma-ray spectrometry analysis method based on GAN and PINN as described in any one of claims 1-8, and the system comprises: The dataset construction module is used to acquire ocean gamma spectrum data and construct a multi-level benchmark dataset based on the ocean gamma spectrum data; The constraint determination module is used to determine physical prior constraints and nuclide prior knowledge constraints based on ocean gamma-ray spectrum data. The adversarial network building module is used to build a physical constraint generative adversarial network based on a benchmark dataset and embed physical prior constraint terms into the physical constraint generative adversarial network; The neural network adversarial module is used to construct a physical information neural network based on a benchmark dataset and embed nuclide prior knowledge constraints into the physical information neural network; The joint optimization module is used to jointly optimize the physical constraint generative adversarial network and the physical information neural network using a preset collaborative training strategy, so that the physical fidelity enhancement spectrum output by the physical constraint generative adversarial network can be adapted to the parsing requirements of the physical information neural network. At the same time, the physical consistency constraints of the physical information neural network guide the parameter optimization of the physical constraint generative adversarial network in reverse. The analysis module is used to input the ocean gamma spectrum data to be analyzed into a collaboratively optimized physical constraint generative adversarial network to obtain a physical fidelity enhanced spectrum; the physical fidelity enhanced spectrum is then input into a collaboratively optimized physical information neural network to output the nuclide type, nuclide activity and corresponding uncertainty.

Images