CN-121679665-B - Gamma energy spectrum analysis method and system based on deep learning physical information network

CN121679665BCN 121679665 BCN121679665 BCN 121679665BCN-121679665-B

Abstract

The application provides a gamma energy spectrum analysis method and a gamma energy spectrum analysis system based on a deep learning physical information network, and relates to the technical field of gamma energy spectrum analysis. The method comprises the steps of obtaining sample data containing energy spectrums and corresponding analysis data, preprocessing the sample data, constructing an energy spectrum analysis model based on a multi-scale convolution neural network, an attention mechanism module and a physical information neural model, training the energy spectrum analysis model by utilizing the preprocessed sample data to generate a target energy spectrum analysis model, inputting the energy spectrums to be analyzed into the target energy spectrum analysis model, and outputting analysis results. By constructing a multi-module collaborative energy spectrum analysis model, the accurate decomposition of energy spectrum multi-scale characteristics, the efficient interaction of cross-characteristic information and the collaborative prediction of multiple parameters are realized, the problems that the spectrum peaks are difficult to separate and identify due to overlapping, the anti-interference capability is weak and the like in the traditional method are solved, and the accuracy, the efficiency and the degree of automation of energy spectrum analysis are improved.

Inventors

WANG JIANYE
GUO HU
ZHANG ZIHENG
YANG MINGHAN
LI GANG

Assignees

中国科学院合肥物质科学研究院

Dates

Publication Date: 20260508
Application Date: 20260212

Claims (10)

1. The gamma energy spectrum analysis method based on the deep learning physical information network is characterized by comprising the following steps of: acquiring sample data containing energy spectrum and corresponding analysis data, and preprocessing the sample data; constructing an energy spectrum analysis model based on a multi-scale convolutional neural network, an attention mechanism module and a physical information neural model, wherein the attention mechanism module performs weight distribution of space dimension on a feature map extracted by the multi-scale convolutional neural network; Training the energy spectrum analysis model by utilizing the preprocessed sample data to generate a target energy spectrum analysis model; and inputting the energy spectrum to be analyzed into a target energy spectrum analysis model, and outputting an analysis result.
2. The method of claim 1, wherein obtaining sample data comprising spectra and corresponding analytical data comprises: The method comprises the steps of obtaining real sample data, carrying out multidimensional data enhancement on the real sample data, adjusting the content of target elements in the real sample data to generate simulation data containing element content and corresponding energy spectrum, carrying out self-adaptive trace energy address displacement on the energy spectrum according to temperature change, and generating gamma energy spectrum corresponding to different detection environments aiming at response functions of different types of detectors.
3. The method of claim 1, wherein preprocessing the sample data comprises: the energy spectrum is standardized, and the formula is: wherein x is the input energy spectrum; And The mean and standard deviation of the global training set are respectively.
4. A method according to any one of claims 1 to 3, wherein training the spectrum analysis model using the pre-processed sample data comprises: inputting the preprocessed sample data into an energy spectrum analysis model to generate a predicted value of the model about element mass fraction; Constructing an MAE Loss function, an MES Loss function and a SmoothL Loss function according to the actual element mass fraction and the predicted value of the model about the element mass fraction, wherein the formulas are as follows: Wherein, the A predicted value for the model with respect to the element mass fraction; L is the energy spectrum length, K is the element category number; The MAE, MES, and SmoothL Loss penalty functions are weighted to generate a data-driven training penalty function L data , with the following formula: Wherein, the 、、 Is a super parameter and is used for balancing the weight of each loss; And training the energy spectrum analysis model set times by using the loss function of the data driving training.
5. The method of claim 4, wherein a semi-supervised learning strategy is employed to fine tune the top-level parameters using the simulation data in the sample data and the backbone network of the training model, and then using the real data in the sample data.
6. The method of claim 4, wherein the physical information constraint loss term is added during training using a physical information neural model to generate a total loss function, the formula being as follows: L data is a loss function of data driving training, namely data error, L physics is a physical information loss function, namely penalty item violating a physical rule; And The weight coefficients of the penalty term for data errors and violations of the physical laws, respectively.
7. The method of claim 6 wherein the physical information constraint loss term comprises a nuclide characteristic peak energy constraint and an element content sum normalization constraint, wherein the nuclide characteristic peak energy constraint comprises a peak area response of a corresponding channel and a specific element, characteristic peak information of the corresponding element is obtained according to a response matrix, and the element content sum normalization constraint comprises that the element contents are non-negative and should be close to 1.
8. The method of claim 6, wherein the physical information constraint loss term is expressed as: Wherein, the The element content sum is normalized, and the non-negativity is realized by adding ReLu activation functions to the output layer; The energy constraint of the characteristic peak of the nuclide is realized; the content of the ith element predicted for the model; A unit mass fraction response spectrum for the i-th element; The mass fraction of the element; Is the sum of the statistical error and the background spectrum, S pred is the predicted energy spectrum, and S true is the real energy spectrum.
9. A gamma energy spectrum analysis system based on a deep learning physical information network, comprising: the data generation module is used for acquiring sample data containing energy spectrum and corresponding analysis data and preprocessing the sample data; the model construction model is used for constructing an energy spectrum analysis model based on the multi-scale convolutional neural network, the attention mechanism module and the physical information neural model, wherein the attention mechanism module performs weight distribution of space dimension on the feature map extracted by the multi-scale convolutional neural network; The model training module is used for training the energy spectrum analysis model by utilizing the preprocessed sample data to generate a target energy spectrum analysis model; the energy spectrum analysis module is used for inputting the energy spectrum to be analyzed into the target energy spectrum analysis model and outputting analysis results.
10. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 8 when the computer program is executed.

Description

Gamma energy spectrum analysis method and system based on deep learning physical information network Technical Field The application relates to the technical field of gamma energy spectrum analysis, in particular to a gamma energy spectrum analysis method and system based on a deep learning physical information network. Background The traditional gamma energy spectrum analysis mainly relies on manual experience to perform peak identification and nuclide judgment, and has the problems of low efficiency, strong subjectivity and the like. In the existing neutron activation gamma energy spectrum analysis method, a method based on peak type fitting needs to be preset with mathematical functions to describe the shape of a full-energy peak, and the area and the position of the peak are solved by utilizing a nonlinear least square method iteration. The method is highly dependent on accurate estimation of initial parameters, has high calculation complexity, and greatly reduces analysis precision in the stratum element detection environment in the face of complex energy spectrum overlapped by spectrum peaks and high Compton background interference. The method based on template matching mainly comprises a Monte Carlo least square method The Monte Carlo Library Least Squares (MCLLS), and the method firstly uses Monte Carlo simulation to generate a standard response spectrum library of each element, wherein the spectrum library is defined in an iterative way until the iteration is the most suitable for a library set of the sample to be studied. Assuming that the measured spectrum is a linear superposition of these standard spectra, the element content can be inversely solved by the least squares method. The main difficulties with this approach are the numerical instability of the least squares stage (matrix inversion may insert numerical instability when the matrix is reversible), overdetermination of the system of equations, linear dependence in the library, gamma radiation scattering. For example, for least squares, when the system of equations is overdetermined, small data disturbances will result in large oscillations of the solution, which is numerical instability, and for gamma radiation scattering problems, compton scattering will result in continuous background and peak broadening of the energy spectrum, which makes complex nonlinear differences between the standard spectrum library and the actual measured spectrum. The traditional machine learning-based methods such as support vector machines and artificial neural networks require manual feature extraction as input, so that not only is feature engineering complex, but also expert experience is relied on, and global context information in energy spectrum is difficult to capture. In summary, in the existing solution method, the method based on the peak type fitting has high calculation complexity, the method based on the template matching has poor adaptability, and the method based on the traditional machine learning has complex characteristic engineering, so that the existing solution method has the problems of difficult separation of spectrum peak overlapping, difficult identification of weak peaks, weak anti-interference capability and the like, and has low accuracy and low efficiency of energy spectrum analysis. Disclosure of Invention Based on the above, it is necessary to provide a gamma energy spectrum analysis method and system based on a deep learning physical information network, which are capable of realizing accurate decomposition of energy spectrum multi-scale characteristics, efficient interaction of cross-characteristic information and multi-parameter collaborative prediction by constructing a multi-module collaborative energy spectrum analysis model, solving the problems of difficult separation of spectrum peak overlapping, difficult identification of weak peaks, weak anti-interference capability and the like in the traditional method, and improving the accuracy, efficiency and automation degree of energy spectrum analysis. In a first aspect, the present application provides a gamma spectrum analysis method based on a deep learning physical information network, including: acquiring sample data containing energy spectrum and corresponding analysis data, and preprocessing the sample data; constructing an energy spectrum analysis model based on a multi-scale convolutional neural network, an attention mechanism module and a physical information neural model, wherein the attention mechanism module performs weight distribution of space dimension on a feature map extracted by the multi-scale convolutional neural network; Training the energy spectrum analysis model by utilizing the preprocessed sample data to generate a target energy spectrum analysis model; and inputting the energy spectrum to be analyzed into a target energy spectrum analysis model, and outputting an analysis result. In one embodiment, obtaining sample data including energy spectra and