Liquid flash TDCR multi-nuclide beta spectrum analysis method and system based on artificial intelligence

Abstract

The invention belongs to the technical field of nuclear radiation measurement, and relates to a liquid flash TDCR multi-nuclide beta spectrum analysis method and system based on artificial intelligence. The method comprises the following steps: constructing a numerical model of a liquid flash detector spectrometer by using a Monte Carlo technology to simulate the energy spectrum response of a single nuclide under different quenching conditions; constructing a training database by adopting a parameterized data synthesis algorithm; constructing a multi-task mixed spectrum analysis neural network model; carrying out model training by utilizing the constructed database; and inputting an actual measurement spectrogram, and outputting the multi-nuclide absolute activity, the detector efficiency and the decomposition energy spectrum contribution curve in real time by using the trained neural network model. According to the method, differential distribution characteristics and quenching response curve characteristics of the beta continuous energy spectrum are creatively fused, a neural network architecture with physical mechanism constraints is constructed, and liquid flash TDCR multi-nuclide energy spectrum characteristic decoupling and accurate activity solving are achieved under the unknown quenching condition.

Description

Technical Field

[0001] The present invention belongs to the technical field of nuclear radiation measurement, and relates to a liquid scintillation TDCR multi-nuclide β spectrum analysis method and system based on artificial intelligence. Background Art

[0002] Radioactive decay mainly includes three types: α, β, and γ. Among them, due to the momentum conservation characteristics accompanied by neutrino emission in β decay, the energy of the released electrons shows a continuous energy spectrum distribution from zero to the maximum value, which is significantly different from the single energy peak of α particles and the discrete characteristic peaks of γ rays. This continuous energy spectrum characteristic makes β energy spectrum measurement irreplaceable in nuclear facility status monitoring (such as 90 Sr activity analysis), environmental radiation assessment ( 3 recognition of low-energy nuclides such as 14 H), and nuclear medicine diagnosis ( 3 C metabolism tracing), etc. It is worth noting that the analysis demand for mixed β nuclide systems is becoming increasingly prominent: dual-nuclide systems (such as 14 H- 3 C) have become standard analysis objects in scenarios such as biomedical tracing and synchronous detection of multiple pollutants, and the demand for mixed systems of three or more nuclides in the research and development of new nuclear drugs is also continuously increasing. However, the continuous distribution characteristics of the β energy spectra of different nuclides lead to a high degree of overlap in the energy regions. Especially under unknown quenching conditions, how to achieve rapid, convenient, and efficient calculation of the absolute activities of multi-nuclides has become a current technical bottleneck.

[0003] Currently, the liquid scintillation analyzer (LSA) is the core device for β energy spectrum measurement. Its international mainstream models include the PerkinElmer Tri-Carb series in the United States, the Hidex 300SL in Finland, and the domestic Shanghai Xinman LSA3000, Hubei Fangyuan FY2700-00, etc. These devices generally adopt a triple photomultiplier tube (PMT) coincidence measurement architecture, and achieve self-adaptive calibration of the single-nuclide detector efficiency through the counting ratio (TDCR) of triple coincidence (TAC) and double coincidence (DAC). The TDCR method effectively suppresses noise interference through triple-tube coincidence logic. Without the need for standard source calibration, it can eliminate the efficiency drift caused by the quenching effect, and significantly improve the measurement accuracy of single-nuclide low-activity samples (such as 3 H detection) in a high-background environment. This breakthrough advantage has enabled it to be included in the reference method for radioactive activity measurement by the International Bureau of Weights and Measures (BIPM) and is widely used in national standard laboratories.

[0004] However, the traditional TDCR method faces fundamental constraints in the analysis of multi-nuclide mixed systems: when the sample contains multiple β-decaying nuclides and the quenching degree is unknown, the existing methods must rely on external calibration to establish a quenching efficiency curve. Take 14Taking the C dual-isotope system as an example, the current solutions mainly rely on two mutually exclusive strategies: the Exclusion Method and the Inclusion Method. The Exclusion Method decouples nuclides through energy spectrum partitioning: selecting the high-energy region (e.g., >50 keV) as the 14 C characteristic window, and solving for the 14 C activity using the count rate and efficiency curve in this energy region, and then inverting the ³H activity through the composite signal in the low-energy region; the Inclusion Method allows energy region overlap and synchronously solves the dual-isotope activities by constructing a system of simultaneous equations (CPM A = H·ε hA + C·ε cA ; CPM B = H·ε hB + C·ε cB ). Although these two methods have been engineered and implemented in patented technologies (such as the efficiency calibration method in CN117949998A and the energy window interpolation method in CN114488264B). However, the existing TDCR methods have common technical defects: First, their experimental systems need to rely on external sources and standard sources of the nuclides to be measured to establish the multi-energy region quenching efficiency response curve. Due to the strict supervision of radioactive substances and the scenario limitations of conventional laboratories and portable devices, it is difficult to routinely equip standard sources (including the supporting γ reference source), which significantly increases the experimental complexity and hinders the standardized promotion; Second, the method has a subjective experience dependence on the energy region division. Manually setting a fixed energy window (such as 14 50-156 keV for 14 C) leads to fragmented utilization of energy spectrum information. Especially in the case of 35 C (Emax = 156 keV) and

[0005] 35 S (Emax = 167 keV) and other energy spectrum adjacent or strong quenching scenarios, the fuzzy energy spectrum boundary and empirical screening are likely to cause significant cross-interference and systematic errors; In addition, the traditional solution model constructs a system of simultaneous equations using fixed-dimensional parameters, and its rigid mathematical framework cannot adapt to the changes in the energy spectrum morphology, facing problems such as no solution, local optimal solution or dynamic calibration failure under complex quenching conditions. These defects jointly restrict the accuracy and applicable boundary of the multi-nuclide β energy spectrum analysis technology.In recent years, the excellent performance of deep neural networks in dealing with multivariable non-linear problems has provided a new path to break through the limitations of traditional methods. Patent CN119763688 demonstrates a neutron detector efficiency correction technology for convolutional neural networks, and the multi-nuclide β energy spectrum analysis is essentially a multi-dimensional inverse problem involving quenching parameters, energy spectrum distribution characteristics, and nuclide activities. By constructing a numerical model of the detector response to generate a training dataset, the neural network can establish a direct mapping between energy spectrum features and nuclide activities, theoretically avoiding the dependence on calibration sources and manual energy region division in traditional methods. This data-driven strategy provides a new technical paradigm for developing adaptive multi-nuclide analysis algorithms, indicating that the β energy spectrum analysis technology is evolving from experience-dependent to intelligent computing-based. Summary of the Invention

[0006] The present invention proposes a method and system for multi-nuclide β energy spectrum analysis of TDCR liquid scintillation based on a neural network (NN). The system consists of two core modules: a detector response modeling module and a multi-task mixed spectrum analysis neural network module. At the technical implementation level, first, the Monte Carlo numerical simulation method is used to construct a TDCR liquid scintillation detector response model, and the system simulates the β energy spectrum response curves of different nuclides (such as , , , etc.) under gradient quenching conditions; secondly, by dynamically adjusting the multi-nuclide concentration ratio and quenching intensity distribution, multi-component mixed energy spectra are generated to construct a database required for model training; then, a multi-layer perceptron neural network architecture is designed, and a non-linear mapping relationship between multi-nuclide energy spectrum response characteristics and absolute activity is established through an adaptive feature extraction mechanism. The core breakthrough of this method is to achieve high-precision analysis of multi-nuclide mixed β energy spectra through a fusion strategy of numerical simulation and deep learning, and simultaneously output the absolute activity of each nuclide and the corresponding β energy spectrum response.

[0007] The liquid scintillation TDCR multi-nuclide β spectrum analysis method based on artificial intelligence provided by the present invention includes the following steps: Step 1: Construct a numerical model of the liquid scintillation detector spectrometer to simulate the energy spectrum responses of single nuclides in liquid scintillation TDCR two-fold coincidence logic and three-fold coincidence under different quenching conditions; Step 2: Use a parametric data synthesis algorithm to construct a training database, including different quenching conditions and multi-component mixtures; Step 3: Construct a multi-task mixed spectrum analysis neural network model; use the constructed database for cascaded multi-task training of the model; Step 4: Input the measured spectrum diagram, and use the trained neural network model to output the absolute activities of multi-nuclides, detection efficiencies, and decomposed energy spectrum contribution curves in real time.

[0008] Preferably, in the first step, a numerical model of the liquid scintillation detector spectrometer is constructed using Monte Carlo techniques, including: detector geometry modeling, nuclide decay process simulation, liquid scintillation performance simulation, optical physics process simulation, photomultiplier tube response, and signal digitization process simulation.

[0009] Preferably, in the second step, the method for constructing the training database includes: using the constructed numerical model, randomly generating quenching values, and simulating the TDCR response energy spectra of any two-tube and three-tube coincidences of a single nuclide under different quenching conditions; secondly, dynamically adjusting the multi-nuclide concentration ratio and quenching intensity distribution to generate multi-component mixed energy spectra.

[0010] Preferably, in the third step, the multi-task mixed-spectrum analysis neural network model adopts a two-stage cascade design: in the first stage, the mixed-spectrum features are extracted through a shared fully connected layer, and the activities of each nuclide and the detector efficiency are predicted synchronously. The gradient magnitude adaptive dynamic weight allocation mechanism is used to dynamically balance the optimization objectives of the activity and the detector efficiency; in the second stage, the predicted results of the activity and the detector efficiency are fused with the shared features to reconstruct the original spectra of each nuclide and the mixed spectra of two-tube and three-tube.

[0011] Preferably, the input of the multi-task mixed-spectrum analysis neural network model is the TDCR response energy spectra of two-tube coincidence and three-tube coincidence of multi-nuclide mixtures, and the outputs are the activities of each nuclide, the detector efficiency, the original spectra of each nuclide, and the decomposed spectra of two-tube and three-tube.

[0012] The present invention also provides a liquid scintillation TDCR multi-nuclide β spectrum analysis system based on artificial intelligence. The system includes a Monte Carlo numerical simulation module for liquid scintillation spectrometers and a training module for the multi-task mixed-spectrum analysis neural network model; the Monte Carlo numerical simulation module for liquid scintillation spectrometers is used to simulate the liquid scintillation TDCR two-channel coincidence logic and three-channel coincidence energy spectra responses of a single nuclide under different quenching conditions; the training module constructs a training database based on the simulated energy spectra responses and uses the training database to train the multi-task mixed-spectrum analysis neural network model; the multi-task mixed-spectrum analysis neural network model is used to predict the measured spectrum graph input, and real-time output the absolute activities of multi-nuclides, detection efficiency, and the decomposed energy spectrum contribution curves.

[0013] Preferably, the system further includes a display module for visually displaying the output results.

[0014] The liquid scintillation TDCR multi-nuclide β spectrum analysis method and system provided by the present invention break through the technical bottlenecks of traditional methods in three dimensions: calibration source dependence, energy spectrum analysis fragmentation, and model rigid constraints. The present invention innovatively integrates the differential distribution characteristics of the β continuous energy spectrum and the characteristics of the quenching response curve, constructs a neural network architecture with physical mechanism constraints, and realizes the decoupling of multi-nuclide energy spectrum characteristics and the accurate solution of activities under unknown quenching conditions. Description of the Drawings

[0015] Figure 1 This is the flow chart for the modeling and optimization of the numerical model of the LSA detector in the embodiments of the present invention; Figure 2 This is a schematic diagram of the cascaded multi-task mixed-spectrum analysis neural network model in the embodiments of the present invention; Figure 3 This is a schematic diagram of an application example of the liquid scintillation TDCR multi-nuclide β-spectrum analysis system based on artificial intelligence provided by the present invention. Detailed Embodiments

[0016] To facilitate the understanding of the present invention, the present invention will be described in more detail below with reference to the accompanying drawings and specific embodiments. The preferred embodiments of the present invention are shown in the drawings. All equivalent technical deformations and improvement measures derived from the core technical solution of the present invention, as long as the implementation schemes achieved by conventional experimental means, fall within the scope of legal protection, specifically covering but not limited to the following technical dimensions: optimization of the detection cavity structure at the hardware adaptation level, upgrading of the detector material components, and alternative solutions for optoelectronic conversion devices; strategies for expanding training samples in the data processing link, optimization of the data batch processing strategy, enhancement of the filtering algorithm for input / output energy spectra, iteration of data cleaning and denoising techniques, and upgrading of the parallel computing architecture; replacement of the architectures of the fully connected layer and the convolutional / transposed convolutional layer in the model architecture dimension, dynamic configuration of the number of neurons, selection of activation functions, dynamic adjustment of the number of training times and early stopping mechanism, adaptive learning rate scheduling strategy, and optimization of the multi-task loss function; reconstruction of the model evaluation index at the algorithm verification level, exploration of the regularization parameter space, and adjustment of Monte Carlo simulation parameters; expansion of the nuclide database at the extended application level, improvement of signal processing technology, and optimization of the detector shielding material, etc. This patent strictly follows the principle of central limitation. The protection scope of the claims is not limited to the embodiments described in the specification. Any improvement scheme based on equivalent technical principles and achieved by conventional engineering optimization means, including but not limited to the integration of attention mechanisms, optimization of residual connection topologies, adjustment of batch normalization strategies, development of training data spectrum enhancement technologies, optimization of detector cavity materials, etc., regardless of their forms of expression and optimization degrees, are regarded as falling within the legal protection boundary of the claims of this patent. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosed content of the present invention more thorough and comprehensive.

[0017] Embodiment 1 The method for analyzing the multi-nuclide β-spectrum of liquid scintillation TDCR based on artificial intelligence provided by the present invention includes the following steps: Step 1: Numerical simulation of the liquid scintillation detector spectrometer. The present invention adopts a hierarchical modeling strategy to construct a full-link digital twin of the liquid scintillation spectrometer to achieve the energy spectrum response of a single nuclide under different quenching conditions. And through comparison and optimization of actual measurement and simulation, a numerical model of the liquid scintillation detector spectrometer is finally constructed.

[0018] Modeling Process of Liquid Scintillation Spectrometer Figure 1 As shown in the figure, it includes the following steps: I. First, a detector physical model is constructed based on the Monte Carlo method, specifically including: 1. Detector Modeling (1) Geometric Structure: Taking a cylindrical hollow cavity as an example, the inner diameter of the cavity is 100 mm, the height is 200 mm, and the inner layer of the cavity is coated with a 0.2 mm reflective layer. Three photomultiplier tubes (PMTs) with a diameter of 50 mm and a sensitive area diameter of 48 mm are installed on the side of the cavity and are symmetrically distributed at 120°. The PMTs are coupled to the test cavity through a quartz window (refractive index 1.46). The sample is placed at the center of the test cavity.

[0019] (2) Material Properties: The material of the test cavity is polytetrafluoroethylene with a density of 0.89 g / cm³. The inner reflective layer of the cavity is made of polyethylene material, and a bidirectional scattering distribution function (BSDF) is defined on the surface of the reflective layer to simulate the mixed effect of diffuse reflection and specular reflection. The sample bottle and the optical windows of the PMTs are both fused silica glass with a density of 2.2 g / cm³. Its refractive index is wavelength-dependent and is described by the Sellmeier equation, such as n = 1.470 @400 nm. The refractive index of the liquid scintillator is 1.59, and the light attenuation length is 35 cm.

[0020] 2. Parameterization of Liquid Scintillation Performance (1) Light Yield Model Based on the Birks quenching formula, the photon yield per unit path length is defined , where the light yield coefficient photons / Me, (Birks quenching constant), and the energy deposition rate is calculated by the continuous slowing down approximation (CSDA) of β particles in the liquid scintillator.

[0021] (2) Nonlinear Quenching Correction Lookup table interpolation of Scintillation Yield vs. dE / dx is introduced to cover the saturation effect in the low energy region (<100 keV) to ensure that the light yield is consistent with the experimental calibration data.

[0022] 3. Modeling of Radioactive Source Samples (1) β Emission Energy Spectrum Based on the nuclide decay database (such as NuDat 3.0), the Fermi function is called to calculate the theoretical β energy spectrum: ; where, is the mass of the electron, is the speed of light, is the maximum energy of β particles, is the nuclear Coulomb correction factor.

[0023] (2)Spatial distribution The radiation source is evenly distributed in the sample bottle to avoid boundary effects; the number of decays is set to times to ensure that the statistical fluctuation of the energy spectrum < 1%.

[0024] (3)Background noise Simulate the spatio-temporal distribution of the background in the simulated environment (such as cosmic ray muons, natural radionuclides), and superimpose it on the main signal to enhance the authenticity of the simulation.

[0025] 4. Refined modeling of physical processes (1)Particle transport Simulate the ionization loss, δ-ray generation, and Cherenkov radiation threshold determination (triggered when the β velocity exceeds the speed of light in the liquid scintillator) that occur during the transport of β particles (electrons / positrons).

[0026] (2)Photon generation and collection The number of photons generated by each energy deposition event is based on the Birks formula, and the photon emission direction is isotropically distributed; photons experience reflection / refraction during transmission: the interface behavior is tracked through a geometric optical model; the absorption / scattering of photons in the liquid scintillator: the probability is calculated based on the absorption length and scattering length of the liquid scintillator.

[0027] Secondly, it is the readout electronics system, including the simulation of the PMT gain process and the simulation of the electronics response. Specifically: Simulation of the PMT gain process: By establishing the cascade amplification of the photomultiplier tube (PMT), defining the quantum efficiency of PMT photoelectric conversion (20% @ 400 nm), and simulating the full-link conversion from a single-photon pulse to the PMT readout signal. Focus on quantifying the impact of PMT gain fluctuations and dark noise interference on the energy spectrum shape.

[0028] Simulation of signal digitization: Convert the analog pulse into discrete energy spectrum data through the ADC model. The detector response function generated at this stage is combined with the physical simulation data to form a simulated energy spectrum, including digital signals and TDCR and other information.

[0029] By systematically comparing the Monte Carlo simulation data with the experimentally measured energy spectrum and trigger coincidence information (TDCR), a parameter optimization mechanism for the detector physical model and the front-end electronics system is established. After iterative calibration, the relative deviation between the simulated values and the experimentally measured values of the quenching response curve and detection efficiency can be controlled within a 5% confidence interval. The present invention constructs a universal simulation correction framework, which only requires a set of reference nuclides (such as The experimental data of ( ) can complete the parameter calibration of the detector response characteristics, without the need for the measured data support of all nuclides to be measured. The calibrated Monte Carlo simulation system only needs an open nuclide database, and can achieve the accurate reconstruction of the energy spectrum characteristics of any nuclide through particle transport calculation and energy deposition model.

[0030] Step 2: Generation of multi-nuclide mixed energy spectrum data.

[0031] Construct an enhanced training database through a parameterized data synthesis algorithm. Using the verified numerical model of the liquid scintillation detector, first randomly generate quenching parameters (light yield and Birks constant , and generate the basic quenching spectra of each nuclide in the two-tube / three-tube configuration , covering the energy range of 0.1 - 10 MeV, and simulate the TDCR response energy spectra of any two-tube and three-tube coincidences of a single nuclide under different quenching conditions.

[0032] On this basis, dynamically adjust the multi-nuclide concentration ratio and quenching intensity distribution to generate multi-component mixed energy spectra: apply random activity , simulate different nuclide activity ratios; introduce PMT noise (Gaussian distribution ) and electronics baseline drift to simulate the real measurement environment. Finally, generate 10 7 groups of mixed spectrum sample databases (80% for the training set, 10% for the validation set, 10% for the test set), covering the two-nuclide and three-nuclide mixed scenarios, providing a high-fidelity and strong generalization training basis for the neural network.

[0033] Step 3: Cascade multi-task mixed spectrum analysis neural network.

[0034] The present invention constructs a deep learning model dedicated to the decomposition of mixed nuclear β energy spectra. The main process of this part is shown in Figure 2 , and its core architecture adopts a two-stage cascade design to achieve double optimization from physical quantity inversion to spectrum shape fine-tuning.

[0035] 1. Construction of the model The first stage (physical quantity inversion): The TDCR multi-nuclide mixed spectra of the two-tube and three-tube are input and high-dimensional features are extracted through a shared fully connected layer (two 2-layer FC networks with 128 and 645 neurons respectively, ReLU activation, Dropout regularization), and independent tasks are branched out: (1) Activity prediction, a 3-layer FC network outputs the activity coefficients of each nuclide, and a non-negative constraint is applied (Softplus output); (2) Efficiency prediction, a 2-layer FC network outputs the detection efficiency coefficients of each nuclide, and a numerical 0 - 1 constraint is applied (Sigmoid output); (3) Adopt gradient magnitude adaptive weight allocation: Dynamically balance the optimization weights of activity (MSE loss) and efficiency (MSE loss); The second stage (spectrum shape reconstruction and parameter fine-tuning): Concatenate the activity, efficiency, and shared features and input them into a spectrum shape reconstructor with a lower learning rate; Reconstruct the two-tube and three-tube spectral quenching spectra independent of each nuclide through an FC network (Huber loss), apply non-negativity constraints (Softplus output), and perform smoothing filtering for denoising; Generate two-tube / three-tube mixed spectra based on the physical mixing formula: ; ; Among them, is the th nuclide, N is the total number of nuclide types in the mixed sample.

[0036] 2. Training and evaluation of the model (1) Training strategy Two-stage joint training: Unlock the parameters of the second stage after pre-training for 200 rounds; Early stopping mechanism: Monitor the loss function value of the validation set and tolerate 10 rounds without improvement; Optimizer: AdamW (lr = 1e-3 in the first round, 1e-4 in the second round).

[0037] During the training process, the model monitors the validation set based on the early stopping strategy using a mixed spectrum dataset containing different quenching conditions and concentration ratios to ensure generalization, and saves the best model according to the validation results.

[0038] (2) Evaluation system Table 1 Evaluation parameters Index Physical meaning Target preset value MSE (activity) Component identification accuracy >0.95 MSE (efficiency) Detector response modeling accuracy <0.05 SSIM (spectrum shape) Mixed spectrum reconstruction fidelity >0.95 .

[0039] Step 4: Use the trained neural network model to output the absolute activities, detection efficiencies, and decomposition energy spectrum contribution curves of multiple nuclides in real time. The specific process is shown in Figure 3 .

[0040] The user inputs the measured spectrum, and the model makes predictions on it. After loading the trained model, efficiency prediction, activity prediction, and contribution spectrum generation are performed. The predicted activity values and decomposition spectra are displayed to the user for easy analysis and decision-making. This process not only improves the analysis efficiency but also reduces the dependence on professional knowledge.

[0041] Embodiment 2 The present invention also provides an artificial intelligence-based liquid scintillation TDCR multi-nuclide β spectrum analysis system, which includes a liquid scintillation spectrometer Monte Carlo numerical simulation module, a training module, and a cascaded multi-task mixed spectrum analysis neural network model; the liquid scintillation spectrometer Monte Carlo numerical simulation module is used to simulate the energy spectrum response of a single nuclide under different quenching conditions; the training module constructs a training database based on the simulated energy spectrum response and uses the training database to train the cascaded multi-task mixed spectrum analysis neural network model; the cascaded multi-task mixed spectrum analysis neural network model is used to predict the input measured spectrum diagram and output the absolute activity of multi-nuclides, the detector efficiency, and the decomposed energy spectrum contribution curve in real time. The output results are displayed through a display module.

[0042] The online analysis engine loads the pre-trained model (best_model.pth) and outputs the prediction results in real time. The human-computer interaction interface provides a spectrum comparison view, a dynamic rendering of the confidence heat map (nuclide probability mapped by color scale), and a data export module (compatible with CSV / ROOT format).

Claims

1. An artificial intelligence-based liquid scintillation TDCR multi-nuclide β spectrum analysis method, characterized in that, It includes the following steps: Step 1: Construct a numerical model of the liquid scintillation detector spectrometer to simulate the energy spectrum responses of the two-channel coincidence logic sum and three-channel coincidence of liquid scintillation TDCR for a single radionuclide under different quenching conditions; Step 2: Construct a training database using a parametric data synthesis algorithm, including different quenching conditions and multi-component mixtures; Step 3: Construct a multi-task mixed-spectrum analysis neural network model and perform cascade multi-task training of the model using the constructed database; Step 4: Input the measured spectrum and use the trained neural network model to output the absolute activities of multiple radionuclides, detector efficiency, and the decomposed energy spectrum contribution curves in real time.

2. The method for analyzing multi-nuclide β spectra of liquid scintillation TDCR based on artificial intelligence according to claim 1, wherein In the above Step 1, a numerical model of the liquid scintillation detector spectrometer is constructed using Monte Carlo techniques, including: detector geometry modeling, radionuclide decay process simulation, liquid scintillation performance simulation, optical physics process simulation, photomultiplier tube response, and signal digitization process simulation.

3. The method for multi - nuclide β - spectrum analysis of liquid scintillation TDCR based on artificial intelligence according to claim 1, wherein In the above Step 2, the method for constructing the training database includes: using the constructed numerical model to randomly generate quenching values and simulate the TDCR response energy spectra of any two-tube and three-tube coincidences of a single radionuclide under different quenching conditions; secondly, dynamically adjusting the concentration ratios of multiple radionuclides and the quenching intensity distribution to generate multi-component mixed energy spectra.

4. The method for multi - nuclide β - spectrum analysis of liquid scintillation TDCR based on artificial intelligence according to claim 1, wherein, In the above Step 3, the multi-task mixed-spectrum analysis neural network model adopts a two-stage cascade design: in the first stage, the mixed-spectrum features are extracted through a shared fully connected layer, and the activities and detector efficiencies of each radionuclide are predicted synchronously. A gradient magnitude adaptive dynamic weight allocation mechanism is used to dynamically balance the optimization objectives of activity and detector efficiency; in the second stage, the predicted results of activity and detector efficiency are fused with the shared features to reconstruct the original spectra of each radionuclide and the two-tube and three-tube mixed spectra.

5. The method for multi - nuclide β - spectrum analysis of liquid scintillation TDCR based on artificial intelligence according to claim 4, wherein, The input of the multi-task mixed-spectrum analysis neural network model is the TDCR response energy spectra of the two-channel coincidence logic sum and three-channel coincidence of multiple radionuclide mixtures, and the output is the activities of each radionuclide, detector efficiency, the original spectra of each radionuclide, and the decomposed spectra of the two-tube and three-tube.

6. An artificial intelligence-based liquid scintillation TDCR multi-nuclide β spectrum analysis system, characterized in that: The system includes a Monte Carlo numerical simulation module for liquid scintillation spectrometers and a training module for the multi-task mixed-spectrum analysis neural network model; the Monte Carlo numerical simulation module for liquid scintillation spectrometers is used to simulate the energy spectrum responses of the two-channel coincidence logic sum and three-channel coincidence of liquid scintillation TDCR for a single radionuclide under different quenching conditions; the training module constructs a training database based on the simulated energy spectrum responses and uses the training database to train the multi-task mixed-spectrum analysis neural network model; the multi-task mixed-spectrum analysis neural network model is used to predict the input measured spectrum and output the absolute activities of multiple radionuclides, detection efficiency, and the decomposed energy spectrum contribution curves in real time; when the system runs, it implements the steps of the method described in any one of claims 1-5.

7. The liquid scintillation TDCR multi-nuclide β spectrum analysis system based on artificial intelligence according to claim 6, wherein, The system further includes a display module for visually displaying the output results.

Images