A method for uranium mine gamma energy spectrum deconvolution based on PSO optimized ResNet

Abstract

The present application relates to a kind of ResNet based on PSO optimization's uranium mine gamma spectrum deconvolution method, belong to uranium mine gamma spectrum deconvolution technical field.The ResNet1D model in the present application avoids gradient vanishing by residual connection, so that the network can be deeper to extract the subtle features of gamma spectrum, improve the resolution ability to spectral overlap.The ResNet1D model of the present application is robust after actual test, and the anti-noise ability is superior, can extract stable feature mode in noisy environment, more adaptable to field logging noise and instrument drift.Moreover, the present application uses PSO to optimize hyperparameter, improves model performance, avoids the blindness and inefficiency of artificial parameter adjustment.Compared with the traditional method, the present scheme has better performance on mean square error (MSE), coefficient of determination (R2), residual prediction deviation (RPD) and the like, especially suitable for uranium logging such as high noise, spectral line complex application scenario.

Description

Technical Field

[0001] This invention relates to a method for interpreting uranium ore gamma-ray spectra based on PSO-optimized ResNet, belonging to the field of uranium ore gamma-ray spectra interpretation technology. Background Technology

[0002] In uranium exploration and radioactive logging, gamma spectral interpretation is a crucial step in the quantitative identification of radioactive elements such as uranium, thorium, and potassium. Traditional interpretation methods often struggle to handle situations involving high-resolution gamma spectroscopy, low signal-to-noise ratios, overlapping spectral lines, and complex backgrounds.

[0003] Currently, the main methods employed are the empirical band method, the standard spectrum least squares method, statistical and fitting methods, and artificial intelligence and machine learning methods. The empirical band method divides the measured energy spectrum into several energy bands and removes mutual interference using linear algebra. Its drawback is significant stripping errors and poor noise resistance when spectral lines overlap severely. The standard spectrum least squares method uses a known standard response spectrum to establish a system of linear equations to solve for the contribution of each radionuclide. Its drawback is strong dependence on the standard spectrum and susceptibility to instrument drift, environmental noise, and energy resolution, leading to unstable spectral results. The statistical and fitting method estimates the content of each nuclide through probabilistic statistics or nonlinear function fitting. Its drawbacks include high computational complexity, slow convergence, and sensitivity to initial values. The artificial intelligence and machine learning method learns the mapping relationship between the energy spectrum and nuclide content using training samples. Its drawbacks include insufficient model generalization, susceptibility to overfitting, strong reliance on manual feature selection, and poor robustness. Summary of the Invention

[0004] The purpose of this invention is to provide a method for interpreting uranium ore gamma-ray spectra based on PSO-optimized ResNet, in order to solve the technical problems in the background art.

[0005] This invention provides a method for interpreting uranium gamma-ray spectra based on PSO-optimized ResNet, comprising: The raw data of uranium gamma spectrum data is obtained, the raw data of uranium gamma spectrum data is preprocessed and a dataset is constructed. The dataset is divided into training set, validation set and test set according to a preset ratio. A ResNet1D model based on one-dimensional convolution is constructed. The input of the ResNet1D model is uranium gamma spectral data, and the output is the predicted value of nuclide concentration. The PSO optimization algorithm was used to determine the training parameter configuration of the ResNet1D model; Based on the training parameter configuration determined by the PSO optimization algorithm, the ResNet1D model is trained for multiple epochs. At the end of each epoch, the loss between the training set and the validation set is calculated, and the best model with the minimum validation loss is saved. The optimal model is loaded and a test set is used for prediction. The prediction results are evaluated using evaluation metrics. The model whose evaluation results meet the requirements of practical applications is taken as the final model. The final model is then used to perform gamma-ray spectral analysis of uranium ore.

[0006] As a preferred embodiment, preprocessing the raw uranium gamma spectroscopy data and constructing the dataset includes: Remove outliers and missing values ​​from the original data to obtain the dataset; Use standardization or normalization methods to numerically scale the dataset to make all data scale uniformly.

[0007] As a preferred embodiment, after dividing the dataset into training, validation, and test sets according to a preset ratio, the method further includes: The training set data samples are randomly scaled for augmentation, with the scaling factor controlled between 0.9 and 1.1.

[0008] In a preferred embodiment, the ResNet1D model includes an input layer, four residual layers, a global average pooling layer, and a fully connected layer connected in sequence. The input layer's processing of the input sequence of uranium ore gamma spectral data includes: Input sequence for obtaining uranium gamma spectrum data B represents the batch size. L represents the number of channels, and L represents the sequence length. The input sequence The output after convolution, Batch Normalization (BN), and ReLU operations Represented as: Where * denotes convolution; The output after max pooling operation Represented as: The residual layer's output to the input layer The processing steps include: Output of input layer The data passes through four residual layers in sequence, with the output of the previous residual layer serving as the input of the next residual layer. Each residual layer consists of multiple residual blocks (Basic Block D) stacked together. The number of channels in the four residual layers are 16, 32, 64, and 128, respectively. The i-th residual layer contains There are residual blocks, and the number of channels is... The output of the residual layer Represented as: The output of the global average pooling layer for the fourth residual layer The processing procedure is represented as follows: The fully connected layer's output for the global average pooling layer The processing procedure is represented as follows: In a preferred embodiment, the processing of the input sequence by the residual block of the residual layer includes: For the input sequence Perform the first convolution: Where * represents one-dimensional convolution, and 3 is the kernel size; The output of the first convolution Activation is performed using an activation function: Output of the activation function Perform a second convolution: If downsampling is required, then define: otherwise: Final output: .

[0009] As a preferred embodiment, the step of determining the training parameter configuration of the ResNet1D model using the PSO optimization algorithm includes: PSO parameter initialization involves randomly generating a particle swarm in the parameter space. Each particle contains three types of parameters: learning rate, batch size, and momentum. The particle position ranges from 0.000001 to 0.001 for the learning rate, 16 to 128 for the batch size, and 0.1 to 0.99 for the momentum. Fitness calculation involves decoding the hyperparameters corresponding to the particle and applying them to model training, using mean squared error (MSE) as the fitness function; after training the model for one epoch, the validation loss is calculated on the validation set, and the validation loss is used as the fitness value of the particle. The particle swarm is iteratively updated, with particle velocity and position updated based on inertia weight, individual optimal and swarm optimal solutions; each iteration compares and updates the individual optimal and global optimal fitness; this process is repeated 20 times to gradually converge to the global optimal hyperparameters. Speed ​​update is represented as: Location update is represented as: Where w is the inertia weight, c1 and c2 are learning factors, and r1 and r2 are random numbers; Determine the optimal hyperparameters, output the globally optimal parameter combination obtained from PSO, including learning rate, batch size, and momentum, and use it as the training parameter configuration for the final model training.

[0010] As a preferred embodiment, the evaluation metrics used to evaluate the prediction results include mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), coefficient of determination (R²), residual prediction bias (RPD), mean absolute percentage error (MAPE), and symmetric mean absolute percentage error (SMAPE).

[0011] As a preferred embodiment, when training based on the training parameter configuration determined by the PSO optimization algorithm, a stochastic gradient descent optimizer and a learning rate scheduler are combined to train the ResNet1D model for 200 epochs. The loss function is MSELoss (mean squared error), and ReduceLROnPlateau is used. When the loss on the validation set does not decrease, the learning rate is automatically decayed and the best model is saved.

[0012] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: The ResNet1D model in this invention avoids gradient vanishing through residual connections, enabling the network to extract subtle features of the gamma spectrum at a deeper level and improve the ability to distinguish overlapping spectral lines. The input layer of the ResNet1D model utilizes convolutional layers to directly learn features from the energy spectrum data without requiring manual selection of energy bands or features, reducing the influence of human experience on the results. Actual test results of the ResNet1D model demonstrate its robustness and superior noise resistance, enabling it to extract stable feature patterns even in noisy environments and showing greater adaptability to field logging noise and instrument drift. Furthermore, this invention utilizes PSO to optimize hyperparameters, improving model performance. Particle swarm optimization automatically searches for the optimal learning rate, batch size, momentum, and related parameters, avoiding the blindness and inefficiency of manual parameter tuning and achieving better spectral accuracy. Compared to traditional methods, this approach shows better performance in terms of mean squared error (MSE), coefficient of determination (R²), and residual prediction deviation (RPD), making it particularly suitable for high-noise, complex spectral scenarios such as uranium mine logging. Attached Figure Description

[0013] Figure 1The figure shown is a model architecture diagram of a ResNet1D-18 model for a PSO-optimized ResNet-based method for uranium ore gamma spectroscopy interpretation, provided in an embodiment of the present invention. Figure 2 This is a comparison curve of the actual and predicted values ​​of a PSO-optimized ResNet-based gamma-ray spectroscopy method for uranium ore, provided as an embodiment of the present invention. Detailed Implementation

[0014] It should be noted that: The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0015] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0016] In uranium logging projects, the real-time performance and accuracy of uranium gamma-ray spectral analysis are crucial for the rapid identification and analysis of nuclides. Among the prior art known to the inventors of this invention, the shortcomings and causes of existing solutions are as follows: 1. Spectral lines overlap significantly, making precise separation difficult. The characteristic energy peaks of nuclides such as uranium, thorium, and potassium are located close to each other, and the limited energy resolution of the detector leads to peak overlap, resulting in decreased spectral accuracy.

[0017] 2. Noise interference and instrument drift In the field, logging environments are noisy, and instrument gain drift over time causes spectral broadening or shift, rendering traditional linear methods ineffective.

[0018] 3. High dependence on standard response Methods such as least squares and spectrum stripping rely on accurate standard spectra. If the standard spectrum does not match the actual detection environment, the spectral interpretation error will increase significantly.

[0019] 4. Computational complexity and stability issues Nonlinear fitting and statistical methods involve large amounts of computation, are sensitive to initial parameters, and have slow convergence speeds or even fail to converge.

[0020] 5. Traditional machine learning models lack generalization ability. For example, SVM and BP neural networks can only handle a limited number of patterns, have poor adaptability to complex spectral lines or new environments, and are prone to overfitting.

[0021] To address the shortcomings of existing solutions, this embodiment introduces a PSO-optimized ResNet method for uranium ore gamma spectrum analysis. This method first extracts wide-band data from the 1024 acquired gamma data channels, yielding three wide-band data points for each gamma data point. Then, a one-dimensional convolutional version of the ResNet1D-18 model is designed to process the spectral data. The ResNet1D-18 model retains the core idea of ​​ResNet—residual connections—to solve the gradient vanishing / exploding problem in deep network training.

[0022] Combination Figure 1 The ResNet1D-18 model has a structure consisting of an input layer (convolution + BN + ReLU + max pooling) connected in sequence, four residual layers (each layer uses stacked convolutional residual blocks of different depths), a global pooling layer, and a fully connected layer (for classification or regression).

[0023] First, a BasicBlock1D residual block is constructed, in which the convolutional part uses Conv1d convolution with a kernel size of 3. Then, normalization is performed on the time series dimension to alleviate internal covariate bias and accelerate training convergence. In the forward direction, F.relu is used to provide non-linear expressive power. If the number of input and output channels or the stride are different, downsampling is used to adjust the dimensions; otherwise, they are directly added together.

[0024] The formula for BasicBlock1D is shown below, where the input sequence is... Where B is the batch size. L is the number of channels, and L is the sequence length.

[0025] (1) First convolution: Where * denotes one-dimensional convolution.

[0026] (2) Activation function: (3) Second convolution: If downsampling is required, then define: otherwise Final output: The input layer uses Conv1d(7x7, stride=2) to extract preliminary time-series features, equivalent to a 7x7 convolution in Image ResNet. The convolutional layer can learn features directly from the energy spectrum data without requiring manual selection of energy bands or features, reducing the influence of human experience on the results. BatchNorm accelerates and stabilizes training. ReLU non-linear mapping is used. MaxPool1d downsampling reduces feature length and improves computational efficiency.

[0027] In the input layer, the input sequence is : Then, after max pooling: Then, residual layers are stacked, for a total of four layers. Each layer consists of multiple residual blocks (Basic Block D) stacked. Since one-dimensional energy spectrum data is used, the number of channels in each layer is reduced to 16, 32, 64, and 128 respectively. At the same time, stride=2 is used for downsampling to extract high-level features one by one. The _make_layer is used to automatically add downsamples according to the changes in input and output channels to ensure dimensionality matching when the residuals are added.

[0028] For the i-th layer (inclusive) The number of channels is (number of residual blocks). : Finally, AdaptiveAvgPool1d(1) is used to reduce the time dimension to 1, and the output shape becomes [batch, channels, 1]. torch.flatten(x, 1) flattens it into [batch, channels]. Linear fully connected layer, output result.

[0029] The output after all residual layers Fully connected layer: Then I started writing the model training code. First, I stored the wide-band data extracted from the obtained gamma spectrum data into a CSV file. I divided the data into training set, validation set and test set, and used StandardScaler to scale the features to ensure that different features contribute equally to the loss function during training.

[0030] Data augmentation (Feature Scaling) is performed within __getitem__: each sample is randomly multiplied by a scaling factor of [0.9, 1.1] to improve the model's generalization ability.

[0031] Particle Swarm Optimization (PSO) is used to automatically search for optimal hyperparameters (learning rate, batch size, momentum). The formula for PSO is: Speed ​​updates: Location update: Where w is the inertia weight, c1 and c2 are learning factors, and r1 and r2 are random numbers. The fitness function uses the validation set MSE Loss to evaluate the quality of each set of hyperparameters, iteratively updates the settings, and finds the globally optimal combination of hyperparameters.

[0032] After the ResNet1D-18 model is built, the following details the application of the ResNet1D-18 model in uranium ore gamma-ray spectral analysis: S1. Data Acquisition and Preprocessing: Acquire uranium gamma spectral data, remove outliers and missing values ​​from the raw uranium gamma spectral data to obtain a complete dataset; use standardization or normalization methods to numerically scale the data to ensure balanced input dimensions for each channel; divide the dataset into training, validation, and test sets, divide the dataset into test sets and save them independently for final evaluation, and then divide the remaining data into training and validation sets, with the training set accounting for 90% and the validation set accounting for 10%.

[0033] S2. Data Augmentation: Randomly scale the training set data samples to enhance their generalization ability and reduce their dependence on a single sampling condition. The scaling factor is controlled between 0.9 and 1.1.

[0034] S3. Model Construction: A ResNet1D-18 model based on one-dimensional convolution is established, which includes convolutional layers, batch normalization layers, residual blocks, global pooling layers, and fully connected output layers. The input is a gamma energy spectrum sequence, and the output is the predicted value of nuclide concentration. The specific construction process has been described earlier and will not be repeated here.

[0035] S4, PSO parameter initialization: Randomly generate a particle swarm in the parameter space. Each particle contains three types of parameters: learning rate, batch size, and momentum. The particle position ranges are learning rate 0.000001 to 0.001, batch size 16 to 128, and momentum 0.1 to 0.99, respectively.

[0036] S5. Fitness Calculation: Decode the hyperparameters corresponding to the particle and apply them to model training, using mean squared error (MSE) as the fitness function; after training the model for one epoch, calculate the validation loss on the validation set and use the validation loss as the fitness value of the particle.

[0037] S6. Particle swarm iteration update: Update particle velocity and position based on inertia weight, individual optimal and swarm optimal solutions; compare and update individual optimal and global optimal fitness in each iteration; repeat the iteration 20 times to gradually converge to the global optimal hyperparameter.

[0038] S7. Determining the optimal hyperparameters: Output the globally optimal combination of parameters obtained from PSO, including learning rate, batch size, and momentum; use this as the parameter configuration for the final model training.

[0039] S8. Model Training and Validation: Using optimal hyperparameters, a stochastic gradient descent optimizer, and a learning rate scheduler, the ResNet1D model is trained for 200 epochs. At the end of each epoch, the losses on the training and validation sets are calculated, and the model with the minimum validation loss is saved. The optimizer uses SGD (Stochastic Gradient Descent), the loss function uses MSELoss (Mean Squared Error), and ReduceLROnPlateau is used. The learning rate is automatically decayed when the validation set loss does not decrease, and the best model is saved.

[0040] S9. Model Testing and Evaluation: Load the best-performing model from the validation phase and perform predictions using the test set; calculate metrics such as Mean Absolute Error (MAE), Mean Square Error (MSE), Root Mean Square Error (RMSE), Coefficient of Determination (R²), Residual Prediction Deviation (RPD), Mean Absolute Percentage Error (MAPE), and Symmetric Mean Absolute Percentage Error (SMAPE) to comprehensively evaluate the model's accuracy and robustness. The results can be visualized and applied: plot curves comparing actual and predicted values ​​to visually demonstrate the model's performance in uranium mine gamma-ray spectral analysis; the final model can be used for rapid spectral analysis and quantitative analysis in uranium mine well logging.

[0041] The actual evaluation results are as follows: Test set evaluation results for target 1: Mean Absolute Error (MAE): 3.783137559890747; Mean Square Error (MSE): 27.68726348876953; Root Mean Square Error (RMSE): 5.261868953704834; Coefficient of Determination (R²): 0.9960412383079529; Residual Prediction Deviation (RPD): 15.893583297729492. This method achieves high prediction accuracy, with a mean absolute error (MAE) of 3.78, a high coefficient of determination (R²) of 0.996, and a residual prediction deviation (RPD) of 15.89. These indicators show that the model's prediction of uranium content is highly close to the true value, with minimal error, far superior to traditional linear spectral stripping or least squares methods. Figure 2 As can be seen from the comparison curves, the predicted values ​​(orange dashed line) almost completely overlap with the actual values ​​(blue solid line), especially in the main peak region and small spectral peaks, where they can be accurately fitted. This indicates that the proposed method still possesses good spectral interpretation capabilities even when dealing with overlapping spectral lines and complex energy distributions. The model can stably extract key features in high-noise and complex spectral line environments, demonstrating good noise resistance. This is closely related to the automatic learning of spectral line features by the one-dimensional convolutional network and the residual structure ensuring the transfer of deep features.

[0042] Particle Swarm Optimization (PSO) automatically searches for optimal hyperparameters, achieving a balance between training efficiency and prediction accuracy, avoiding the blindness and inefficiency of manual hyperparameter tuning. It is more suitable for the complex environment of uranium mine logging in the field.

[0043] The intelligent spectral analysis method in this embodiment can achieve automated nuclide identification, which greatly improves the efficiency and accuracy of the analysis process, especially in field monitoring and emergency response, and can quickly provide reliable analysis results.

[0044] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0045] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0046] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0047] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0048] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for interpreting uranium ore gamma-ray spectra based on PSO-optimized ResNet, characterized in that, include: The raw data of uranium gamma spectrum data is obtained, the raw data of uranium gamma spectrum data is preprocessed and a dataset is constructed. The dataset is divided into training set, validation set and test set according to a preset ratio. A ResNet1D model based on one-dimensional convolution is constructed. The input of the ResNet1D model is uranium gamma spectrum data, and the output is the predicted value of nuclide concentration. The PSO optimization algorithm was used to determine the training parameter configuration of the ResNet1D model; Based on the training parameter configuration determined by the PSO optimization algorithm, the ResNet1D model is trained for multiple epochs. At the end of each epoch, the loss between the training set and the validation set is calculated, and the best model with the minimum validation loss is saved. Load the best model and use the test set to make predictions. Use evaluation metrics to evaluate the prediction results. Use the model whose evaluation results meet the requirements of practical applications as the final model. Use the final model to perform gamma spectral analysis of uranium ore. The ResNet1D model consists of an input layer, four residual layers, a global average pooling layer, and a fully connected layer connected in sequence. The ResNet1D model's processing of the input sequence of uranium gamma spectral data includes: Input sequence for obtaining uranium gamma spectrum data B represents the batch size. L represents the number of channels, and L represents the sequence length. The input sequence The output after convolution, Batch Normalization (BN), and ReLU operations Represented as: Where * denotes convolution; The output after max pooling operation Represented as: The residual layer's output to the input layer The processing steps include: Output of input layer The data passes through four residual layers in sequence, with the output of the previous residual layer serving as the input of the next residual layer. Each residual layer consists of multiple residual blocks (Basic Block D) stacked together. The number of channels in the four residual layers are 16, 32, 64, and 128, respectively. The i-th residual layer contains There are residual blocks, and the number of channels is... The output of the residual layer Represented as: The output of the global average pooling layer for the fourth residual layer The processing procedure is represented as follows: The fully connected layer's output for the global average pooling layer The processing procedure is represented as follows: 。 2. The method for resolving uranium ore gamma-ray spectra based on PSO-optimized ResNet according to claim 1, characterized in that, The raw data of uranium gamma spectroscopy were preprocessed and a dataset was constructed, including: Remove outliers and missing values ​​from the original data to obtain the dataset; Use standardization or normalization methods to numerically scale the dataset to make all data scale uniformly.

3. The method for interpreting uranium ore gamma-ray spectra based on PSO-optimized ResNet according to claim 1, characterized in that, After dividing the dataset into training, validation, and test sets according to a preset ratio, the following steps are also included: The training set data samples are randomly scaled for augmentation, with the scaling factor controlled between 0.9 and 1.

1.

4. The method for interpreting uranium ore gamma-ray spectra based on PSO-optimized ResNet according to claim 1, characterized in that, The processing of the input sequence by the residual blocks of the residual layer includes: For the input sequence Perform the first convolution: Where * represents one-dimensional convolution, and 3 is the kernel size; The output of the first convolution Activation is performed using an activation function: The output of the activation function Perform a second convolution: If downsampling is required, then define: otherwise: Final output: 。 5. The method for interpreting uranium ore gamma-ray spectra based on PSO-optimized ResNet according to claim 1, characterized in that, The steps for determining the training parameter configuration of the ResNet1D model using the PSO optimization algorithm include: PSO parameter initialization involves randomly generating a particle swarm in the parameter space. Each particle contains three types of parameters: learning rate, batch size, and momentum. The particle position ranges from 0.000001 to 0.001 for the learning rate, 16 to 128 for the batch size, and 0.1 to 0.99 for the momentum. Fitness calculation involves decoding the hyperparameters corresponding to the particle and applying them to model training, using mean squared error (MSE) as the fitness function; after training the model for one epoch, the validation loss is calculated on the validation set, and the validation loss is used as the fitness value of the particle. The particle swarm is iteratively updated, with particle velocity and position updated based on inertia weight, individual optimal and swarm optimal solutions; each iteration compares and updates the individual optimal and global optimal fitness; this process is repeated 20 times to gradually converge to the global optimal hyperparameters. Speed ​​update is represented as: Location update is represented as: Where w is the inertia weight, c1 and c2 are learning factors, and r1 and r2 are random numbers; Determine the optimal hyperparameters, output the globally optimal parameter combination obtained from PSO, including learning rate, batch size, and momentum, and use it as the training parameter configuration for the final model training.

6. The method for interpreting uranium ore gamma-ray spectra based on PSO-optimized ResNet according to claim 1, characterized in that, The evaluation indicators used to assess the prediction results include mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), coefficient of determination (R²), residual prediction bias (RPD), mean absolute percentage error (MAPE), and symmetric mean absolute percentage error (SMAPE).

7. The method for interpreting uranium ore gamma-ray spectra based on PSO-optimized ResNet according to claim 1, characterized in that, When training based on the training parameter configuration determined by the PSO optimization algorithm, a stochastic gradient descent optimizer and a learning rate scheduler are combined to train the ResNet1D model for 200 epochs. The loss function is MSELoss (mean squared error), and ReduceLROnPlateau is used. When the loss on the validation set does not decrease, the learning rate is automatically decayed and the best model is saved.

Images