Marine in-situ radioactivity measurement method based on MLP network and deconvolution

Abstract

The invention discloses an ocean in-situ radioactivity monitoring method, which is characterized in that feature extraction is carried out based on a maximum likelihood expected value maximization (MLEM) algorithm, and spectrum unfolding operation is carried out by adopting a multi-layer perceptron. The method comprises the following steps: firstly, measuring FWHM related parameters of a CeBr3 detector by adopting a standard source, simulating the response of the detector to gamma rays in a marine environment based on a Monte Carlo method, and establishing a full-spectrum response matrix H in a range of 0-2048keV; based on the response matrix H, reconstructing a gamma energy spectrum y obtained by the detector by using an MLEM iterative algorithm; and the reconstructed spectrum is used as the input of a multi-layer perceptron (MLP), and qualitative and quantitative analysis of radionuclides in the energy spectrum is realized. According to the method, the discrete characteristic peak signal is convolved into the single path address corresponding to the characteristic energy, so that the accuracy of the MLP model in the spectrum unfolding process is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of marine in-situ radioactivity measurement and analysis, and specifically relates to a method for marine in-situ radioactivity monitoring based on deconvolution energy spectrum reconstruction and MLP network. Background Art

[0002] As an efficient, low-carbon clean energy, the peaceful use of nuclear energy has become a strategic choice for global energy transformation. As of 2022, there are more than 450 coastal nuclear power units in operation worldwide. The radioactive nuclides (such as 137 Cs, 131 I) Marine proliferation events pose a significant threat to the fishery economy and offshore ecological security. The International Atomic Energy Agency (IAEA) has explicitly called for the establishment of a real-time, intelligent radioactivity monitoring system.

[0003] There is a significant technical gap in the current monitoring technology: the first generation of manual sampling method relies on the collection of discrete water samples by ship cruise, and the analysis of the radionuclide activity data by laboratory gamma spectrometer. Although the measurement accuracy of this method can reach Bq / m 3 However, they have inherent limitations, such as low sampling frequency (typically quarterly) and limited spatial coverage (less than 5% of the monitored sea area), making it difficult to capture the dynamic process of nuclide diffusion. Second-generation in-situ monitoring buoys achieve continuous monitoring through integrated NaI(Tl) crystal detectors. However, the gamma energy spectra they acquire suffer from severe spectral distortion, resulting in nuclide identification accuracy of less than 75% (based on the IAEA test dataset). This is limited by:

[0004] (1) Water attenuation effect: Gamma rays in seawater undergo a composite attenuation due to Compton scattering and photoelectric effect. The number of photons effectively received by the detector is less than 10% of the theoretical value. Simulation calculations show that the attenuation rate of 60keV photons at a water depth of 3m reaches 92%;

[0005] (2) Insufficient crystal energy resolution: The energy resolution (FWHM) of commercial NaI(Tl) crystals is generally >7% @ 662keV. 134 Cs (604.7keV) and 137 The overlap of the characteristic peak of Cs (661.7keV) is 35%, and the low activity nuclides (such as 131 The characteristic peak of I@364.5keV is easily annihilated in the background fluctuation, resulting in problems such as missed weak peaks, false peak misjudgment, and difficulty in analyzing overlapping peaks during energy spectrum analysis.

[0006] In response to these technical challenges, this paper proposes an innovative solution: an in-situ marine radioactivity monitoring method based on deconvolution spectrum reconstruction and a multi-layer perceptron (MLP) network. This method aims to overcome the limitations of traditional monitoring methods through advanced signal processing techniques and machine learning algorithms, enabling accurate and rapid monitoring of radionuclides in the marine environment, providing strong technical support for marine environmental protection and the safe use of nuclear energy. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention is achieved through the following technical solutions:

[0008] The present invention discloses a method for in-situ marine radioactivity monitoring based on deconvolution spectrum reconstruction and MLP network, which specifically comprises the following steps:

[0009] S1: Construction of the detector full spectrum energy response matrix;

[0010] The peak signal x of the linear combination has statistical fluctuations during the radiation signal collection process and appears as a Gaussian broadened peak. The formation process of the energy spectrum can be described as the convolution of the radionuclide characteristic signal x and the detector response function H. The generation process is expressed as follows:

[0011]

[0012] Where y is the energy spectrum data, x is the original signal, and b is the interference signal;

[0013] The dimension of the interference signal b is N; the dimension of the response matrix H is N×N; the expanded matrix form represents the above formula as follows:

[0014]

[0015] The above formula can be described in matrix form:

[0016] y=H*x+b (3)

[0017] S2: Solve the original signal x based on the Maximum Likelihood Expectation Maximization (MLEM) algorithm;

[0018] The process of deconvolution energy spectrum reconstruction is actually the process of solving the original signal x through matrix operations. However, if x is solved directly, many different functions are required to solve the convolution equation within the error range of the experimental data, and the estimated value of the solution is more sensitive to the errors in the input data. The noise signal in the energy spectrum measurement process will cause huge oscillations in the results. By introducing the regularization method and using prior knowledge to constrain the solution space, the ill-conditioned problem is transformed into a well-conditioned problem, ensuring the stability and accuracy of the solution. The MLEM algorithm is used for energy spectrum reconstruction. This method is based on the maximum likelihood principle and estimates the original signal x by constructing the likelihood function of the objective function. The algorithm gradually approaches the variable value under the maximum likelihood probability in an iterative manner, and its iterative process satisfies the following formula:

[0019]

[0020] The iterative steps are as follows:

[0021] S2.1. Set the initial non-zero solution x = [1, 1, 1..., 1];

[0022] S2.2. Set the number of loop iterations Z = 100;

[0023] S2.3. Calculate the value of x after the nth iteration according to the iterative formula;

[0024] S2.4. The x value obtained after each iteration is used as the excitation operation, that is, x = |x| p , where p = 1.1;

[0025] S2.5. Output the energy spectrum after reaching the cycle number Z, otherwise return to step S2.3.

[0026] S3: Nuclide identification based on the reconstructed energy spectrum based on the MLP network;

[0027] Furthermore, the iterative steps in S2 are as follows:

[0028] S2.1. Set the initial non-zero solution x = [1, 1, 1..., 1];

[0029] S2.2. Set the number of loop iterations Z = 100;

[0030] S2.3. Calculate the value of x after the nth iteration according to the iterative formula;

[0031] S2.4. The x value obtained after each iteration is used as the excitation operation, that is, x = |x| p , where p = 1.1;

[0032] S2.5. Output the energy spectrum after reaching the cycle number Z, otherwise return to step S2.3;

[0033] Furthermore, the nuclide identification steps in S3 are as follows: S3.1: Dataset construction: The Monte Carlo method is used to generate the broadened gamma energy spectrum of the nuclide of interest, and the MELM algorithm is used to reconstruct the spectrum as a training set.

[0034] S3.2. Label Encoding: Seawater contains abundant natural and artificial radionuclides, which form multiple characteristic peaks in the gamma-ray spectrum. To accurately represent the presence of different nuclides in the spectrum, one-hot encoding is used to label the existing nuclides.

[0035] S3.3. Evaluation Criteria: Hamming loss is a metric used to evaluate the performance of classification models, particularly when dealing with multi-label classification problems. It measures the ratio of inconsistencies between predicted labels and true labels. Specifically, Hamming loss calculates the proportion of locations where there is inconsistency between the predicted label and the true label. The smaller the value of Hamming loss, the more consistent the model's predictions are with the true labels. Its value range is [0,1]. When all labels are predicted correctly, the Hamming loss is 0; when all labels are predicted incorrectly, the Hamming loss is 1. The Hamming loss expression is as follows:

[0036]

[0037] The F1 score is the harmonic mean of precision and recall, providing a single metric that comprehensively considers both precision and recall, avoiding the bias that can be introduced by a single metric. When the number of class samples is unbalanced, the F1 score can better reflect the performance of the model. The formula for calculating the F1 score is as follows:

[0038]

[0039] S3.4.MLP Model Construction: The hidden layer uses the ReLU (Rectified Linear Unit) activation function. The model training strategy uses the cross-entropy loss function from the PyTorch framework as the loss function. Finally, the Softmax function is activated to calculate the loss between the model output and the label. Backpropagation is used to adjust the weight parameters of each neuron. By setting a dynamic learning rate, the model can be ensured to converge to the optimal value and accelerate convergence.

[0040]

[0041] Cross entropy loss function: L = -∑y i *log(p i )(9)

[0042] The beneficial effects of the present invention are as follows: in the energy spectrum reconstructed through deconvolution using the MLEM iterative algorithm, discrete characteristic peak signals are convolved into corresponding energy channels, effectively improving the energy resolution of the gamma spectrum and enhancing the accuracy of overlapping peak analysis. Subsequent use of the MLP network for nuclide identification reduces manual intervention and effectively increases the accuracy of nuclide identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A Monte Carlo model for in-situ ocean detection of CeBr3 detectors;

[0044] Figure 2 Flowchart of the method for in-situ radioactivity measurement in the ocean based on MLP network and deconvolution; Figure 3 The MLP network model built;

[0045] Figure 4 To simulate the overlapping peak separation effect of the deconvolution reconstruction algorithm when the energy branching ratio is 1:1;

[0046] Figure 5 To simulate the overlapping peak separation effect of the deconvolution reconstruction algorithm when the energy branching ratio is 7:3;

[0047] Figure 6 Measured gamma energy spectrum and its reconstructed spectrum in in-situ ocean monitoring;

[0048] Figure 7 Laboratory standard 137 Measured γ-ray spectrum of Cs liquid source and its reconstructed spectrum;

[0049] Figure 8 This is the spectrum decomposition result of the MLP algorithm during the in-situ measurement process. DETAILED DESCRIPTION

[0050] The present invention will be further described below with reference to the accompanying drawings and examples.

[0051] The ocean in-situ measurement method based on the MLP network and deconvolution of the present invention comprises the following steps:

[0052] S1: The detector measures the standard radiation source, obtains the crystal FWHM parameters, and uses Figure 1 The model shown,constructs the full spectrum response matrix based on the Monte Carlo method;

[0053] Obtain the standard point source γ energy spectrum, and perform Gaussian fitting to calculate the parameter σ for each characteristic peak based on formula (10). Then, the FWHM value corresponding to each energy point is obtained according to formula (11). Finally, the energy efficiency curve is obtained by fitting formula (12).

[0054]

[0055] FWHM=2.355σ (11)

[0056]

[0057] use Figure 1 The model shown in Figure 1 obtains the energy spectrum of each energy point in the range of 0 to 2048 keV based on the fitting parameters in formula (12), and finally obtains the full spectrum energy response matrix of the detector in the ocean in-situ measurement.

[0058] S2: Use MLEM algorithm to iteratively reconstruct the energy spectrum through deconvolution;

[0059] The iterative process of the MLEM algorithm satisfies the following formula:

[0060]

[0061] The iterative steps are as follows: S2.1: Set the initial non-zero solution x = [1, 1, 1..., 1];

[0062] S2.2: Set the number of loop iterations Z = 100;

[0063] S2.3: Calculate the value of x after the nth iteration according to the iterative formula;

[0064] S2.4: The x value obtained after each iteration is used as an excitation operation, that is, x = |x| p , where p = 1.1;

[0065] S2.5: Output the energy spectrum after reaching the cycle number Z, otherwise return to step 3.

[0066] S3: If Figure 2 As shown in Figure 1, the MLP model is used to perform spectrum decomposition on the reconstructed energy spectrum.

[0067] S3.1: Dataset Construction: The training dataset consists of Geant4 simulation data and actual lead chamber background data measured using a CeBr3 detector. During the simulation data acquisition process, a random shift mechanism for characteristic peaks was introduced (the characteristic peaks of a nuclide were randomly shifted within a ten-channel range around their energy values). The training set was generated by randomly combining the energy spectra of different nuclides to enhance the robustness and generalization of the model.

[0068] S3.2. Label encoding: The present invention adopts the following label encoding scheme: (1) Label vector initialization: Define a label vector with a length equal to the number of radionuclide species, initialized to an all-zero vector, i.e., label = [0, 0, ..., 0], where each position in the vector corresponds to a specific nuclide. (2) Nuclide presence marking: If the i-th nuclide is detected in the energy spectrum, the i-th element of the label vector is set to 1; otherwise, it remains 0. For example, if the second and fifth nuclides are simultaneously present in a certain energy spectrum, its label vector is label = [0, 1, 0, 0, 1, ..., 0].

[0069] S3.3. Hamming loss (Equation 14) and F1-score (Equation 15) are used as evaluation indicators to measure model performance.

[0070]

[0071] S3.4. The MLP model structure constructed by the present invention is as follows Figure 3 As shown in the figure, the network architecture is as follows: reconstructed energy spectrum data is used as the input layer, and feature extraction is performed through two fully connected hidden layers. Each layer uses the ReLU (Rectified Linear Unit) activation function for nonlinear transformation. A batch normalization layer is introduced after the hidden layer to accelerate model convergence, and the ReLU function is used again for activation. The output layer uses the Sigmoid activation function to map the predicted values ​​to the range [0, 1]. For model training, the cross-entropy loss function is used to measure the difference between the predicted output and the true label, and the network parameters are optimized through backpropagation. The learning rate adopts a dynamic adjustment strategy: the initial learning rate is set to 0.001, and a step-by-step decay method is used. After every 20 training epochs, the learning rate is decayed to one-tenth of the current value. This effectively balances the need for rapid convergence in the early stages of model training with the need for parameter fine-tuning in the later stages.

[0072] Example 1:

[0073] based on Figure 1 The model shown, get 137 Cs simulation spectrum. During the simulation, the energy branch is set to 1:1, and 662KeV (characteristic peak 1) is used as the initial energy. The energy spectra of overlapping peaks (characteristic peak 2) with intervals of 20, 30, 40, and 50KeV are obtained respectively. The energy spectrum is reconstructed using the MLEM algorithm. The results are as follows: Figure 4 and as shown in Table 1.

[0074] Table 1 Comparison of peak position and count rate of simulated spectrum and reconstructed spectrum when energy branching ratio is 1:1

[0075]

[0076] from Figure 4 It can be seen that when the distance between the two characteristic peaks is less than or equal to 30 channels, the energy spectrum exhibits a single peak with a right-side tail. At this point, traditional peak-finding algorithms can only identify a single peak, and the characteristic peak area calculated by Gaussian fitting has a large error compared to the simulated value. In the reconstructed spectrum, however, the two peaks are successfully separated, and the peak position error remains within two channels. As the energy interval increases, the characteristic peaks gradually exhibit a distinct double-peak morphology, and the peak position error in the reconstructed spectrum is further reduced to one channel. The count rate of characteristic peak 1 in the reconstructed spectrum is generally higher than the simulated value, with a relative error between the two values ​​ranging from 5% to 8%. The relative error of the count rate of characteristic peak 2 is less than 3%. When the count rates of the two characteristic peaks are added together, the total count rate has a relative error of less than 3.29% compared to the simulated value.

[0077] Example 2:

[0078] based on Figure 1 The model shown, get 137 Cs simulation spectrum. During the simulation, the energy branch is set to 7:3, and 662KeV (characteristic peak 1) is used as the initial energy. The energy spectra of overlapping peaks (characteristic peak 2) with intervals of 20, 30, 40, and 50KeV are obtained respectively. The energy spectrum is reconstructed using the MLEM algorithm. The results are as follows: Figure 5 and as shown in Table 2.

[0079] Table 2 Comparison of peak position and count rate of simulated spectrum and reconstructed spectrum when energy branching ratio is 7:3

[0080]

[0081]

[0082] from Figure 5 It can be seen that when the energy branching ratio is 7:3, the single peak morphology presented at the same energy spacing is narrower and has a right tail at approximately half the peak intensity, significantly increasing the difficulty of separating overlapping peaks. When the spacing between the two characteristic peaks is 20 channels, the peak position of characteristic peak 2 in the reconstructed spectrum differs from the simulated value by 4 channels. This error gradually decreases with increasing energy spacing. The relative error between the count rate of characteristic peak 1 in the reconstructed spectrum and the simulated value is within 4.77%, and the relative error between the count rate of characteristic peak 2 and the simulated value is within 2.82%. The maximum relative error between the total count rate of the two peaks and the simulated value is 3.24%.

[0083] Example 3:

[0084] like Figure 6As shown in the figure, a CeBr3 detector is used to obtain the ocean in-situ γ spectrum, and the spectrum is reconstructed based on the MLEM algorithm. 40 The K activity concentration is 11340.44 Bq·m -3 25L of seawater was collected near the measurement area and measured using a high-purity germanium detector calibrated in the laboratory. 40 The K activity concentration is 10545 Bq·m -3 The calculation error of this method is 7.0%, which is within the acceptable range.

[0085] Implementation 4:

[0086] like Figure 7 As shown, built in the laboratory 137 Cs standard liquid source, using CeBr3 detector to obtain γ spectrum, reconstructed spectrum calculated based on MLEM algorithm 137 The Cs activity concentration is 754.41 Bq m -3 The liquid source is measured using a high-precision, high-purity germanium detector calibrated in the laboratory. 137 The Cs activity concentration is 749.71 Bqm -3 , and the calculation error of this method is only 0.63%.

[0087] Example 5:

[0088] like Figure 8 As shown in the figure, a real-time monitoring experiment was carried out in the field, using a CeBr3 detector to obtain the in-situ γ spectrum of seawater in real time, and the algorithm proposed in this paper was used to carry out real-time monitoring. 40 K and its activity concentration.

[0089] The proposed method demonstrates excellent performance in in-situ marine radioactivity monitoring, specifically: 1. Excellent identification of overlapping peaks. For overlapping peaks with similar energies (greater than 20 keV), traditional peak-finding algorithms are unable to identify double peaks, but this method achieves high identification accuracy. Under varying branching ratios, the single-peak count rate error in the reconstructed spectrum remains generally below 5.6%, and the total count error remains below 3.29%. 2. After reconstructing the energy spectrum using the MLEM algorithm, the discrete full-energy peak information is convolved onto the corresponding single channel address, significantly improving the accuracy of the MLP network in automated radionuclide monitoring.

Claims

1. A method for measuring in-situ radioactivity in the ocean based on MLP network and deconvolution, characterized in that: The following steps are involved: S1: Construction of the detector full spectrum energy response matrix; The peak signal x of the linear combination has statistical fluctuations during the radiation signal collection process and appears as a Gaussian broadened peak. The formation process of the energy spectrum can be described as the convolution of the radionuclide characteristic signal x and the detector response function H. The generation process is expressed as follows: Where y is the energy spectrum data, x is the original signal, and b is the interference signal; The dimension of the interference signal b is N; the dimension of the response matrix H is N×N; the expanded matrix form represents the above formula as follows: The above formula can be described in matrix form: y=H*x+b (3) S2: Solve the original signal x based on the Maximum Likelihood Expectation Maximization (MLEM) algorithm; The deconvolution spectrum reconstruction process is actually the process of solving the original signal x through matrix operations. However, if x is solved directly, many different functions are needed to solve the convolution equation within the error range of the experimental data. The estimated value of the solution is sensitive to errors in the input data. The noise signal in the spectrum measurement process will cause large oscillations in the result. By introducing a regularization method and using prior knowledge to constrain the solution space, the ill-conditioned problem is transformed into a well-conditioned problem, ensuring the stability and accuracy of the solution. The MLEM algorithm is used for energy spectrum reconstruction. This method is based on the maximum likelihood principle and estimates the original signal x by constructing the likelihood function of the objective function. The algorithm gradually approximates the variable value under the maximum likelihood probability through iteration. The iterative process satisfies the following formula: S3: Nuclide identification is performed on the reconstructed energy spectrum based on the MLP network.

2. The method for measuring in-situ radioactivity in the ocean based on MLP network and deconvolution according to claim 1, characterized in that: The iterative steps in S2 are as follows: S2.

1. Set the initial non-zero solution x = [1, 1, 1..., 1]; S2.

2. Set the number of loop iterations Z = 100; S2.

3. Calculate the value of x after the nth iteration according to the iterative formula; S2.

4. The x value obtained after each iteration is used as the excitation operation, that is, x = |x| p , where p = 1.1; S2.

5. Output the energy spectrum after reaching the cycle number Z, otherwise return to step S2.

3.

3. The method for measuring in-situ radioactivity in the ocean based on MLP network and deconvolution according to claim 1, characterized in that: The steps of nuclide identification in S3 are as follows: S3.1: Dataset construction: Generate the broadened gamma spectrum of the nuclide of interest using the Monte Carlo method, and reconstruct the spectrum using the MELM algorithm as a training set; S3.

2. Label Encoding: Seawater contains abundant natural and artificial radionuclides, which form multiple characteristic peaks in the gamma-ray spectrum. To accurately represent the presence of different nuclides in the spectrum, one-hot encoding is used to label the existing nuclides. S3.

3. Evaluation Criteria: Hamming loss is a metric used to evaluate the performance of classification models. It measures the ratio of inconsistencies between predicted labels and true labels, particularly when dealing with multi-label classification problems. Specifically, Hamming loss calculates the proportion of locations where there are inconsistencies between predicted and true labels. A smaller Hamming loss value indicates a higher consistency between the model's predictions and the true labels. Its value range is [0, 1]. When all labels are predicted correctly, the Hamming loss is 0; when all labels are predicted incorrectly, the Hamming loss is 1. The Hamming loss expression is as follows: The F1 score is the harmonic mean of precision and recall, and aims to provide a single metric that comprehensively considers both precision and recall, avoiding the bias that may be caused by a single metric. When the number of class samples is unbalanced, the F1 score can better reflect the performance of the model. The calculation formula for the F1 score is as follows: S3.4.MLP model construction: The hidden layer uses the ReLU (Rectified Linear Unit) activation function. In terms of model training strategy, the cross entropy loss function under the PyTorch framework is selected as the loss function. Finally, it is activated by the Softmax function to calculate the loss between the model output and the label, and backpropagation is used to adjust the weight parameters of each neuron. By setting the dynamic learning rate, it can ensure that the model converges to the optimal value while accelerating the convergence speed. Cross entropy loss function: L = -∑y i *log(p i )(9).

Images