CLYC neutron-gamma ray discrimination detection method and device based on deep learning

Abstract

The invention relates to the technical field of CLYC neutron-gamma rays, and provides a CLYC neutron-gamma ray screening detection method and device based on deep learning, and the method comprises the steps: carrying out the preprocessing of an original pulse waveform outputted by a CLYC detector, and obtaining clean standard waveform data; constructing a one-dimensional convolutional neural network model, respectively capturing the features of the waveform at different time scales, and identifying the most discriminative feature region in the waveform through an attention mechanism; performing adaptive adjustment on the original hyper-parameter of the one-dimensional convolutional neural network model to obtain an optimized hyper-parameter; establishing a direct mapping relation from an original pulse waveform to a neutron-gamma particle type to obtain a trained model; and carrying out compression optimization processing on the model, and deploying the model to an embedded platform to carry out neutron-gamma ray discrimination. According to the method, automatic extraction of deep discrimination features of the CLYC pulse waveform and accurate identification of key feature areas are realized, and the discrimination accuracy and noise interference resistance of neutron-gamma rays are improved.

Description

Technical Field

[0001] The present invention relates to the field of CLYC neutron-gamma ray technology, and in particular to a CLYC neutron-gamma ray discrimination and detection method and device based on deep learning. Background Art

[0002] CLYC (Cesium lithium yttrium chloride, Cs2LiYCl6:Ce 3+ ) Neutron-gamma-ray discrimination detection technology refers to the use of CLYC crystal detectors to distinguish and identify neutrons and gamma rays. As a new type of dual-mode scintillation detector, the CLYC crystal can produce scintillation responses to neutrons and gamma rays at the same time, and can achieve particle type discrimination by analyzing the differences in the time characteristics of their output pulse waveforms. Compared with traditional He-3 proportional counters or plastic scintillator detectors, CLYC detectors have significant advantages such as higher detection efficiency, better energy resolution, and stronger anti-interference capabilities. This technology has important application value in the fields of nuclear safety monitoring, nuclear material identification, nuclear reactor monitoring, radioactive waste detection, and nuclear non-proliferation verification. The CLYC detector can achieve high-precision neutron-gamma-ray discrimination in complex radiation environments, providing reliable technical support for nuclear safety protection and radiation monitoring.

[0003] Existing techniques for CLYC neutron-gamma-ray discrimination primarily rely on traditional signal processing methods based on pulse shape discrimination. While these methods achieve basic distinction between neutrons and gamma rays, in practice, they suffer from insufficient discrimination performance in the low-energy range and poor robustness to noise interference. Furthermore, most methods employ simple threshold comparison or linear discriminant analysis, ignoring the multi-timescale characteristics of the complex luminescence mechanism of CLYC crystals. This makes it difficult to accurately discriminate weak signal events at low count rates in high-precision scenarios such as nuclear safety. In particular, the missed detection rate for thermal neutrons is high in strong gamma-ray backgrounds, and the misjudgment rate is high in complex energy spectrum environments. This makes it impossible to meet the stringent requirements for discrimination accuracy and real-time performance in nuclear safety monitoring. Summary of the Invention

[0004] In view of this, the present invention proposes a CLYC neutron-gamma ray discrimination and detection method and device based on deep learning, which solves the problems of insufficient discrimination performance in the low energy range and poor robustness to noise interference in the existing technology. In addition, most of them adopt simple threshold comparison or linear discriminant analysis, ignoring the multi-time scale characteristics of the complex luminescence mechanism of CLYC crystals, resulting in a high missed detection rate of thermal neutrons in a strong gamma ray background and a high misjudgment rate in a complex energy spectrum environment.

[0005] The technical solution of the present invention is implemented as follows: In a first aspect, the present invention provides a CLYC neutron-gamma ray discrimination detection method based on deep learning, comprising the following steps: Preprocess the original pulse waveform output by the CLYC detector to obtain clean standard waveform data; A one-dimensional convolutional neural network model is constructed, which uses multiple convolution kernels of different scales to extract features from clean standard waveform data, capturing features at different time scales of the waveform and identifying the most discriminative feature regions in the waveform through an attention mechanism. Adopting a particle swarm optimization algorithm to adaptively adjust the original hyperparameters of the one-dimensional convolutional neural network model to obtain optimized hyperparameters; Performing end-to-end deep learning training on the one-dimensional convolutional neural network model based on optimized hyperparameters, establishing a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type, verifying the discrimination performance of the trained model, and obtaining a trained neutron-gamma ray discrimination model; The neutron-gamma ray discrimination model is compressed and optimized, and deployed to an embedded platform for real-time neutron-gamma ray discrimination.

[0006] Based on the above technical solution, preferably, the one-dimensional convolutional neural network model is constructed, and multiple convolution kernels of different scales are used to extract features from the clean standard waveform data, respectively capturing features of the waveform at different time scales, and identifying the most discriminative feature areas in the waveform through the attention mechanism, including: Constructing a multi-scale convolutional feature extraction network and establishing a multi-scale one-dimensional convolutional neural network architecture, the multi-scale one-dimensional convolutional neural network architecture includes a short-time convolution layer, a medium-time convolution layer, and a long-time convolution layer. Convolution kernels of different scales are used to perform parallel feature extraction on the clean standard waveform data to obtain the short-time scale features, medium-time scale features, and long-time scale features of the waveform. At the same time, the energy information and time information of the waveform are extracted to obtain waveform feature maps at multiple scales. An adaptive attention mechanism is integrated. Based on the luminescence mechanism characteristics of CLYC crystals, a query-key-value-value attention algorithm is used to perform weighted fusion of multi-scale feature maps. The energy-time two-dimensional information is combined to identify and highlight the most discriminative time regions and feature components in the waveform for neutron-gamma discrimination, and an enhanced feature representation is obtained after fusion.

[0007] On the basis of the above technical solution, preferably, the short-time convolution layer adopts a one-dimensional convolution operation with a convolution kernel size of 7 and a number of filters of 64, which is used to capture the rapid change characteristics and rising edge characteristics of the waveform, corresponding to the rapid response of the CVL luminescence component of the CLYC crystal; The medium-time convolution layer uses a one-dimensional convolution operation with a convolution kernel size of 15 and 128 filters to capture the medium time scale characteristics of the waveform and the shape characteristics of the pulse body, corresponding to the medium response of the Ce luminescence component; The long-term convolution layer uses a one-dimensional convolution operation with a convolution kernel size of 31 and 256 filters to capture the slow decay characteristics and long-time scale variation patterns of the waveform, corresponding to the slow response of the STE luminescence component; The deposition energy information is extracted from the waveform data through integration operation. Each convolution layer is configured with a batch normalization layer and a ReLU activation function, and downsampling is performed through a maximum pooling layer.

[0008] Based on the above technical solutions, preferably, the query-key-value-value attention algorithm adopts a multi-head attention mechanism to map the feature map of each scale into a query matrix, a key-value matrix and a value matrix respectively, and incorporates energy-time two-dimensional information; based on the different response characteristics of neutrons and gamma rays in CLYC crystals, the attention weight matrix is ​​obtained by calculating the similarity between the query matrix and the key-value matrix, and the attention enhancement features of each scale are obtained by weighted summation of the value matrix; the original features are fused with the attention enhancement features using residual connection and layer normalization technology; the enhanced features of different scales are weightedly combined through a dynamic weight allocation strategy based on particle physics properties to obtain a fused and unified enhanced feature representation.

[0009] On the basis of the above technical solution, preferably, the particle swarm optimization algorithm is used to adaptively adjust the original hyperparameters of the one-dimensional convolutional neural network model to obtain optimized hyperparameters, including: Constructing a particle swarm optimization model and establishing a parameter space mapping relationship of the particle swarm optimization algorithm, wherein the particle swarm optimization model includes a learning rate, a batch size, a random dropout rate, and a search range of a regularization coefficient, initializing the position and velocity vectors of the particle swarm, and constructing a fitness evaluation function, wherein the fitness evaluation function uses the model identification accuracy as a constraint condition; The optimal hyperparameter combination is searched through the iterative process of the particle swarm optimization algorithm. The position and velocity of the particles are updated according to the fitness evaluation results in each iteration until the global optimal solution is converged, and the optimized hyperparameters suitable for the CLYC neutron-gamma discrimination task are obtained.

[0010] Based on the above technical solution, preferably, the method of adaptively adjusting the original hyperparameters of the one-dimensional convolutional neural network model using a particle swarm optimization algorithm to obtain optimized hyperparameters further includes: Perform particle swarm iterative optimization. In each iteration, the one-dimensional convolutional neural network model is trained using the current hyperparameter combination and the fitness score is calculated. The position and velocity of the particles are adjusted according to the standard update rule of the particle swarm optimization algorithm, and the individual optimal position and the global optimal position are updated, gradually converging to the optimal hyperparameter combination. Output optimized hyperparameters, where the optimized hyperparameters are a combination of hyperparameters corresponding to the global optimal position, and the optimized hyperparameters include an optimal learning rate, an optimal batch size, an optimal random dropout rate, and an optimal regularization coefficient.

[0011] On the basis of the above technical solution, preferably, the one-dimensional convolutional neural network model is subjected to end-to-end deep learning training based on optimized hyperparameters, a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type is established, and the discrimination performance of the trained model is verified to obtain a trained neutron-gamma ray discrimination model, including: The one-dimensional convolutional neural network model is configured by optimizing hyperparameters, the cross-entropy loss function and the Adam optimizer are used for model training, and the prediction results are calculated by forward propagation and the network weights are updated by backpropagation to establish a direct mapping relationship from the clean standard waveform to the neutron-gamma particle type label; An independent test data set is used to verify the discrimination performance of the trained model, and the model's evaluation indicators, including accuracy, precision, recall rate, and F1 score, are calculated. The model's discrimination performance in different energy ranges and waveform conditions is analyzed. When the model meets the expected neutron-gamma ray discrimination requirements, the trained neutron-gamma ray discrimination model is output.

[0012] On the basis of the above technical solution, preferably, the one-dimensional convolutional neural network model is subjected to end-to-end deep learning training based on optimized hyperparameters, a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type is established, the discrimination performance of the trained model is verified, and a trained neutron-gamma ray discrimination model is obtained, further comprising: The clean standard waveform data is divided into training, validation, and test sets, and the training data is randomly shuffled and batched. The model training parameters, including the learning rate, batch size, random dropout rate, and regularization coefficient, are set using optimized hyperparameters. The training process is performed in an end-to-end manner, inputting the clean standard waveform data, extracting features through multi-scale convolution and using an attention mechanism to obtain feature representations. Neutron-gamma binary classification results are output through a fully connected classification layer, and the prediction error is calculated using a cross-entropy loss function. The network parameters are updated using the back-propagation algorithm of the Adam optimizer to gradually optimize the model's discrimination performance. The trained model was tested for discrimination performance using an independent test dataset containing neutron and gamma-ray waveform samples across different energy ranges and signal-to-noise ratios. The confusion matrix of the model's predictions was statistically analyzed, and evaluation metrics including accuracy, precision, recall, F1 score, and area under the receiver operating characteristic (ROC) curve were calculated. The model's ability to discriminate between neutrons and gamma rays was analyzed, and its stability and robustness under different conditions were evaluated.

[0013] Based on the above technical solution, preferably, the neutron-gamma ray discrimination model is compressed and optimized, and deployed on an embedded platform for real-time neutron-gamma ray discrimination, including: The neutron-gamma ray discrimination model was lightweighted by using model quantization, parameter pruning, and knowledge distillation to reduce the model's storage space and computational complexity. While maintaining discrimination performance, a compressed and optimized model suitable for deployment on embedded platforms was obtained. The compression optimization model is transplanted to the embedded hardware platform, and model format conversion, runtime environment configuration and hardware acceleration optimization are performed to establish a real-time data acquisition and processing process.

[0014] In a second aspect, the present invention further provides a CLYC neutron-gamma ray discrimination detection device based on deep learning, the device comprising: The data processing module is used to pre-process the original pulse waveform output by the CLYC detector to obtain clean standard waveform data; The feature extraction module is used to build a one-dimensional convolutional neural network model. It uses multiple convolution kernels of different scales to extract features from clean standard waveform data, capturing features at different time scales of the waveform and identifying the most discriminative feature areas in the waveform through an attention mechanism. A parameter optimization module, configured to adaptively adjust the original hyperparameters of the one-dimensional convolutional neural network model using a particle swarm optimization algorithm to obtain optimized hyperparameters; A model training module is used to perform end-to-end deep learning training on the one-dimensional convolutional neural network model based on optimized hyperparameters, establish a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type, verify the discrimination performance of the trained model, and obtain a trained neutron-gamma ray discrimination model; The model deployment module is used to compress and optimize the neutron-gamma ray discrimination model and deploy it to the embedded platform for real-time neutron-gamma ray discrimination.

[0015] The CLYC neutron-gamma ray discrimination detection method and device based on deep learning of the present invention have the following beneficial effects compared with the prior art: (1) By constructing a deep learning model combining a multi-scale one-dimensional convolutional neural network with an attention mechanism, using a particle swarm optimization algorithm for adaptive adjustment of network hyperparameters, establishing a direct mapping relationship based on an end-to-end training method, and implementing model compression optimization and embedded deployment, the automatic extraction of deep-level discriminative features of the CLYC pulse waveform and the precise identification of key feature areas were achieved, thereby improving the neutron-gamma ray discrimination accuracy and noise resistance. (2) By constructing a multi-scale one-dimensional convolutional neural network architecture including short-term, medium-term, and long-term convolutional layers, combined with a query-key-value-value algorithm with a multi-head attention mechanism, based on CLYC crystal CVL, Ce³ + The dynamic weight allocation strategy based on the physical characteristics of the STE luminescent components and the energy-time two-dimensional adaptive attention calculation achieves accurate feature extraction of CLYC pulse waveforms at different time scales and automatic identification of key discrimination areas, thereby improving the accuracy of neutron-gamma ray discrimination. (3) By constructing a particle swarm optimization model with a four-dimensional search space including learning rate, batch size, random inactivation rate and regularization coefficient, and adopting a swarm intelligence search strategy, combined with a fitness evaluation function that takes the discrimination accuracy as a constraint and comprehensively considers the balance between precision and recall rate, the adaptive optimization adjustment of the hyperparameters of the one-dimensional convolutional neural network model is achieved, thereby improving the generalization performance and training stability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 This is a flow chart of the CLYC neutron-gamma ray discrimination detection method based on deep learning of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] See also Figure 1 The present invention provides a CLYC neutron-gamma ray discrimination detection method based on deep learning, comprising the following steps: Preprocess the original pulse waveform output by the CLYC detector to obtain clean standard waveform data; A one-dimensional convolutional neural network model is constructed, which uses multiple convolution kernels of different scales to extract features from clean standard waveform data, capturing features at different time scales of the waveform and identifying the most discriminative feature regions in the waveform through an attention mechanism. Adopting a particle swarm optimization algorithm to adaptively adjust the original hyperparameters of the one-dimensional convolutional neural network model to obtain optimized hyperparameters; Performing end-to-end deep learning training on the one-dimensional convolutional neural network model based on optimized hyperparameters, establishing a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type, verifying the discrimination performance of the trained model, and obtaining a trained neutron-gamma ray discrimination model; The neutron-gamma ray discrimination model is compressed and optimized, and deployed to an embedded platform for real-time neutron-gamma ray discrimination.

[0020] Specifically, this embodiment realizes the automatic extraction of deep-level discrimination features of CLYC pulse waveforms and accurate identification of key feature areas by constructing a deep learning model combining a multi-scale one-dimensional convolutional neural network with an attention mechanism, adopting a particle swarm optimization algorithm for adaptive adjustment of network hyperparameters, establishing a direct mapping relationship based on an end-to-end training method, and implementing model compression optimization and embedded deployment, thereby improving the neutron-gamma ray discrimination accuracy and noise resistance.

[0021] The preprocessing of the original pulse waveform output by the CLYC detector to obtain clean standard waveform data includes: The preprocessing includes high-speed ADC sampling, baseline correction, noise filtering and amplitude normalization. The original pulse waveform output by the CLYC detector is sampled by the high-speed ADC and preliminarily digitized. The sampling frequency is set to no less than 80MSps, the sampling accuracy is set to no less than 14 bits, the sampling length is set to 1024 sampling points, and trigger judgment and effective event screening are performed.

[0022] In a specific embodiment, the high-speed ADC sampling adopts an adaptive sampling strategy, dynamically adjusting the sampling starting point according to the rising edge characteristics of the pulse signal to ensure that the sampling window completely covers the rising edge, peak and attenuation tail of the pulse waveform; the trigger judgment adopts a dual threshold mechanism, setting a low threshold for event detection and a high threshold for noise suppression. When the signal amplitude exceeds the low threshold and exceeds the high threshold within a preset time window, it is determined to be a valid event and sampling is started; the valid event screening eliminates abnormal events such as truncated waveforms and those exceeding the energy range through waveform integrity checks and energy window screening.

[0023] Waveform cleaning and standardization processing: baseline correction, noise filtering and amplitude normalization are performed on the digitized waveform data in sequence to obtain clean standard waveform data with an amplitude range of 0-1.

[0024] In a specific embodiment, the baseline correction adopts a moving median filtering algorithm, selects the first 100 sampling points of the waveform to calculate the baseline value, and performs baseline drift correction on the entire waveform; the noise filtering adopts an adaptive Wiener filtering method, automatically adjusts the filtering parameters according to the signal-to-noise ratio, and sets the filtering cutoff frequency to 40 MHz, thereby maintaining the integrity of the waveform characteristics while suppressing high-frequency noise; the amplitude normalization adopts a maximum value normalization strategy, normalizing the peak amplitude of the corrected waveform to 1, while keeping the relative time characteristics and shape characteristics of the waveform unchanged.

[0025] Specifically, this embodiment achieves high-quality digital acquisition and standardized processing of the CLYC original pulse waveform by adopting a high-speed ADC digital processing with a sampling frequency of not less than 80MSps and a sampling accuracy of not less than 14 bits, combined with an adaptive sampling strategy, a dual-threshold trigger judgment mechanism, a moving median filter baseline correction, an adaptive Wiener filter noise suppression and a maximum normalization processing, effectively suppressing the interference of baseline drift, high-frequency noise and abnormal events, ensuring that the sampling window accurately captures the complete characteristics of the pulse waveform, thereby obtaining clean standard waveform data with an amplitude range between 0-1 and maintaining the integrity of relative time and shape characteristics.

[0026] The one-dimensional convolutional neural network model is constructed, and multiple convolution kernels of different scales are used to extract features from clean standard waveform data, respectively capturing features of waveforms at different time scales, and identifying the most discriminative feature areas in the waveform through an attention mechanism, including: Constructing a multi-scale convolutional feature extraction network and establishing a multi-scale one-dimensional convolutional neural network architecture, the multi-scale one-dimensional convolutional neural network architecture includes a short-time convolution layer, a medium-time convolution layer, and a long-time convolution layer. Convolution kernels of different scales are used to perform parallel feature extraction on the clean standard waveform data to obtain the short-time scale features, medium-time scale features, and long-time scale features of the waveform. At the same time, the energy information and time information of the waveform are extracted to obtain waveform feature maps at multiple scales. An adaptive attention mechanism is integrated. Based on the luminescence mechanism characteristics of CLYC crystals, a query-key-value-value attention algorithm is used to perform weighted fusion of multi-scale feature maps. The energy-time two-dimensional information is combined to identify and highlight the most discriminative time regions and feature components in the waveform for neutron-gamma discrimination, and an enhanced feature representation is obtained after fusion.

[0027] Specifically, this embodiment constructs a multi-scale one-dimensional convolutional neural network architecture including short-term, medium-term, and long-term convolutional layers, combines the query-key-value-value algorithm of the multi-head attention mechanism, and is based on CLYC crystal CVL, Ce³ +The dynamic weight allocation strategy of the physical characteristics of the STE luminescent components and the energy-time two-dimensional adaptive attention calculation realize the accurate feature extraction of CLYC pulse waveform at different time scales and the automatic identification of key discrimination areas, thereby improving the accuracy of neutron-gamma ray discrimination.

[0028] The one-dimensional convolutional neural network model is constructed, and multiple convolution kernels of different scales are used to extract features from clean standard waveform data, respectively capturing features of waveforms at different time scales, and identifying the most discriminative feature areas in the waveform through an attention mechanism. The model also includes: The short-time convolution layer uses a one-dimensional convolution operation with a convolution kernel size of 7 and 64 filters to capture the rapid change characteristics and rising edge characteristics of the waveform, corresponding to the rapid response of the CVL luminescence component of the CLYC crystal; the medium-time convolution layer uses a one-dimensional convolution operation with a convolution kernel size of 15 and 128 filters to capture the medium time scale characteristics and pulse body shape characteristics of the waveform, corresponding to the medium response of the Ce luminescence component; the long-time convolution layer uses a one-dimensional convolution operation with a convolution kernel size of 31 and 256 filters to capture the slow decay characteristics and long time scale change pattern of the waveform, corresponding to the slow response of the STE luminescence component; The deposition energy information is extracted from the waveform data through integration operation. Each convolution layer is configured with a batch normalization layer and a ReLU activation function, and downsampling is performed through a maximum pooling layer. The query-key-value-value attention algorithm adopts a multi-head attention mechanism to map the feature map of each scale into a query matrix, a key-value matrix and a value matrix respectively, and incorporates energy-time two-dimensional information. Based on the different response characteristics of neutrons and gamma rays in CLYC crystals, the attention weight matrix is ​​obtained by calculating the similarity between the query matrix and the key-value matrix, and the attention-enhanced features of each scale are obtained by weighted summation of the value matrix. The original features are fused with the attention-enhanced features using residual connection and layer normalization techniques to prevent gradient vanishing and maintain feature diversity. The enhanced features of different scales are weighted and combined through a dynamic weight allocation strategy based on particle physics properties to obtain a fused and unified enhanced feature representation.

[0029] In a specific embodiment, the calculation formula of the enhanced feature representation is: ; ; ; in, for t Enhanced feature representation of the moment, For the i Dynamic time weight of each scale, For the i The feature vector extracted by the convolutional layer, 、 、 are the outputs of the short-term, medium-term, and long-term convolutional layers, respectively. is the energy-time coupling correction factor, E is the deposition energy of the particle in the CLYC crystal, is the energy threshold for neutron-gamma discrimination, t is the sampling time point. is 2 ns , is the decay time constant of the CVL luminescence component of the CLYC crystal, corresponding to the rapid luminescence process; 50 ns , is CLYC crystal Ce³ + The decay time constant of the luminescent component corresponds to a medium-speed luminescence process; 1 μs , is the decay time constant of the STE luminescence component of the CLYC crystal, corresponding to the slow luminescence process.

[0030] In a specific embodiment, the energy-time two-dimensional adaptive attention calculation formula of the energy-time two-dimensional information is: ; in, is the energy-time two-dimensional adaptive attention weight matrix of energy-time two-dimensional information, is the sigmoid activation function, is the time domain query vector, is the energy domain key value vector, d is the dimension of the attention vector, is the time period coupling coefficient, is the energy logarithmic coupling coefficient, is the characteristic period of the waveform, is the reference energy.

[0031] In a specific embodiment, the calculation formula of the dynamic weight allocation strategy based on particle physical properties is: ; ; ; ; in, For the i The dynamic weight coefficient of each scale, For the i Neutron response learning parameters at each scale, For the i The gamma response learning parameters of the scale, For the i The coupling strength learning parameters of each scale, is the neutron response characteristic function, is the gamma response characteristic function, is the time-energy coupling strength function, is the normalized waveform function, is the characteristic time constant of the slow component, is the characteristic energy parameter, is the peak energy corresponding to the waveform, is the time-energy coupling strength parameter, is an exponential function.

[0032] Specifically, The particle swarm optimization algorithm is used to adaptively adjust the original hyperparameters of the one-dimensional convolutional neural network model to obtain optimized hyperparameters, including: Constructing a particle swarm optimization model and establishing a parameter space mapping relationship of the particle swarm optimization algorithm, wherein the particle swarm optimization model includes a learning rate, a batch size, a random dropout rate, and a search range of a regularization coefficient, initializing the position and velocity vectors of the particle swarm, and constructing a fitness evaluation function, wherein the fitness evaluation function uses the model identification accuracy as a constraint condition; The optimal hyperparameter combination is searched through the iterative process of the particle swarm optimization algorithm. The position and velocity of the particles are updated according to the fitness evaluation results in each iteration until the global optimal solution is converged, and the optimized hyperparameters suitable for the CLYC neutron-gamma discrimination task are obtained.

[0033] Specifically, this embodiment constructs a particle swarm optimization model with a four-dimensional search space including learning rate, batch size, random inactivation rate and regularization coefficient, adopts a swarm intelligence search strategy, and combines it with a fitness evaluation function that takes the discrimination accuracy as a constraint and comprehensively considers the balance between precision and recall rate. It realizes the adaptive optimization and adjustment of the hyperparameters of the one-dimensional convolutional neural network model, thereby improving the generalization performance and training stability of the model.

[0034] The method further includes adaptively adjusting the original hyperparameters of the one-dimensional convolutional neural network model using a particle swarm optimization algorithm to obtain optimized hyperparameters. Perform particle swarm iterative optimization. In each iteration, first use the current hyperparameter combination to train the one-dimensional convolutional neural network model and calculate the fitness score. Then, adjust the position and velocity of the particles according to the standard update rule of the particle swarm optimization algorithm, update the individual optimal position and the global optimal position, and gradually converge to the optimal hyperparameter combination. Set the convergence judgment condition and the maximum number of iterations to 50. When the improvement of the global optimal fitness is less than 0.001 in 10 consecutive iterations or when the maximum number of iterations is reached, the optimization process is terminated. Output the optimized hyperparameters, which are the hyperparameter combination corresponding to the global optimal position. The optimized hyperparameters include the optimal learning rate, the optimal batch size, the optimal random dropout rate, and the optimal regularization coefficient, which are used in the subsequent model training process.

[0035] In a specific embodiment, the hyperparameter search space is set to the range of learning rate search [0.0001, 0.1], the range of batch size search [16, 128], the range of random dropout rate search [0.1, 0.8], and the range of regularization coefficient search [0.0001, 0.01]. Initialize the particle swarm parameters, set the particle swarm size to 20 particles, the position vector of each particle to a 4-dimensional hyperparameter combination, the velocity vector to zero vector, the inertia weight coefficient to 0.9, the cognitive learning factor to 2.0, and the social learning factor to 2.0; Construct a fitness evaluation function, using the model's recognition accuracy on the validation dataset as the main evaluation indicator. While considering the balance between precision and recall, calculate the fitness score of each hyperparameter combination through weighted synthesis, and give priority to hyperparameter configurations with high recognition performance and stable training.

[0036] Specifically, this embodiment constructs a four-dimensional search space including a learning rate [0.0001, 0.1], a batch size [16, 128], a random dropout rate [0.1, 0.8], and a regularization coefficient [0.0001, 0.01]. It adopts a particle swarm parameter configuration of 20 particles, an inertia weight coefficient of 0.9, a cognitive learning factor of 2.0, and a social learning factor of 2.0. In combination with a weighted fitness evaluation function that takes the discrimination accuracy as the main evaluation indicator and comprehensively considers the balance between precision and recall, and sets a dual convergence judgment mechanism of a maximum of 50 iterations and an improvement of less than 0.001 for 10 consecutive times, it achieves precise search and global optimal configuration of the hyperparameters of the one-dimensional convolutional neural network, avoids the blindness and local optimal trap of traditional manual parameter adjustment, improves the stability and convergence efficiency of model training, and ensures that the CLYC neutron-gamma discrimination model obtains the optimal combination of learning rate, batch size, random dropout rate, and regularization coefficient.

[0037] The one-dimensional convolutional neural network model is subjected to end-to-end deep learning training based on optimized hyperparameters, a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type is established, and the discrimination performance of the trained model is verified to obtain a trained neutron-gamma ray discrimination model, including: The one-dimensional convolutional neural network model is configured by optimizing hyperparameters, the cross-entropy loss function and the Adam optimizer are used for model training, and the prediction results are calculated by forward propagation and the network weights are updated by backpropagation to establish a direct mapping relationship from the clean standard waveform to the neutron-gamma particle type label; An independent test data set is used to verify the discrimination performance of the trained model, and the model's evaluation indicators, including accuracy, precision, recall rate, and F1 score, are calculated. The model's discrimination performance in different energy ranges and waveform conditions is analyzed. When the model meets the expected neutron-gamma ray discrimination requirements, the trained neutron-gamma ray discrimination model is output.

[0038] The method further includes performing end-to-end deep learning training on the one-dimensional convolutional neural network model based on optimized hyperparameters, establishing a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type, verifying the discrimination performance of the trained model, and obtaining a trained neutron-gamma ray discrimination model. The clean standard waveform data was divided into training, validation, and test sets in an 8:1:1 ratio. The training data was randomly shuffled and batched, with each batch containing 16-128 waveform samples to ensure a balanced number of neutron and gamma-ray samples. Model training parameters were set using optimized hyperparameters, including learning rate, batch size, random dropout rate, and regularization coefficient. The number of training rounds was set to 100-200, and an early stopping strategy was adopted to prevent overfitting. The training process was executed in an end-to-end manner, with clean standard waveform data as input. Feature representation was obtained through multi-scale convolutional feature extraction and attention mechanism. The neutron-gamma binary classification result was output through a fully connected classification layer. The prediction error was calculated using the cross-entropy loss function. The network parameters were updated through the back-propagation algorithm of the Adam optimizer to gradually optimize the model's discrimination performance. The trained model is tested for discrimination performance using an independent test dataset containing neutron and gamma-ray waveform samples in different energy ranges and under different signal-to-noise ratio conditions to ensure the comprehensiveness and representativeness of the test. The confusion matrix of the model's prediction results is statistically analyzed, and evaluation indicators such as accuracy, precision, recall, F1 score, and area under the ROC curve are calculated. The model's ability to discriminate neutrons and gamma rays is analyzed, and the model's stability and robustness under different conditions are evaluated. The model's discrimination performance differences in different energy ranges are analyzed, and areas and conditions with weak performance are identified. The model structure or training strategy is adjusted based on the test results. When the overall discrimination accuracy of the model reaches above 95% and the recall rates of neutrons and gamma rays both exceed 90%, the model training is confirmed to be complete, and the final neutron-gamma-ray discrimination model is output.

[0039] Specifically, this embodiment adopts an 8:1:1 ratio data partitioning, random shuffling and batch processing training data organization strategy, combined with batch processing including 16-128 samples and 100-200 training cycles. End-to-end deep learning framework, as well as cross-entropy loss function and Adam optimizer's back-propagation algorithm, configures early stopping strategy and overfitting protection mechanism, calculates comprehensive evaluation indicators such as accuracy, precision, recall rate, F1 score and area under the ROC curve based on the confusion matrix, sets performance standards of more than 95% identification accuracy and more than 90% neutron-gamma ray recall rate, realizes the establishment of a direct mapping relationship from clean standard waveform to neutron-gamma particle type label, ensures the stable identification performance of the model in different energy ranges and waveform conditions, overcomes the limitations of traditional methods that require complex feature engineering and manual labeling, and provides an end-to-end deep learning solution for high-precision real-time identification of CLYC detectors in complex radiation environments.

[0040] The neutron-gamma ray discrimination model is compressed and optimized, and deployed on an embedded platform for real-time neutron-gamma ray discrimination, including: The neutron-gamma ray discrimination model is lightweighted, and methods such as model quantization, parameter pruning and knowledge distillation are used to reduce the model's storage space and computational complexity. While maintaining the discrimination performance, a compressed and optimized model suitable for deployment on embedded platforms is obtained.

[0041] In a specific embodiment, the model parameters are converted from 32-bit floating-point numbers to 8-bit integer representations, and the model accuracy is maintained and the model storage space is reduced through quantization-aware training or post-training quantization methods; the importance distribution of weight parameters in the network layer is analyzed, and redundant parameters close to zero and neuronal connections that contribute little to the discrimination performance are removed, and the number of model parameters is reduced through a combination of structured pruning and unstructured pruning.

[0042] A lightweight student network model is constructed, and the trained original model is used as the teacher network. The knowledge of the teacher network is transferred to the student network through soft label training and feature alignment, resulting in a compressed model with more than 70% parameter reduction but less than 3% performance loss.

[0043] The compression optimization model is transplanted to an embedded hardware platform, and model format conversion, runtime environment configuration, and hardware acceleration optimization are performed. A real-time data acquisition and processing process is established to achieve real-time neutron-gamma ray discrimination of the CLYC detector output waveform.

[0044] In a specific embodiment, the compressed and optimized model is converted into a model format supported by the embedded platform, such as TensorFlow Lite, ONNX, or a dedicated hardware format, and the runtime library and dependent environment are configured to ensure the compatibility and stability of the model on the target embedded platform; a complete data processing link is established from CLYC detector signal acquisition to identification result output, and a real-time data buffer and a multi-threaded processing mechanism are set to ensure that the real-time preprocessing, feature extraction, and identification classification of waveform data can be completed within 100 microseconds; the GPU, DSP, or dedicated accelerator of the embedded platform is used to accelerate model inference, and the identification speed is improved through parallel computing and pipeline processing methods to achieve real-time identification performance of processing more than 10,000 waveform samples per second, while monitoring the system power consumption and temperature to ensure long-term stable operation.

[0045] Specifically, this embodiment compresses the model parameters from 32-bit floating point numbers to 8-bit integer representation by adopting quantization-aware training and post-training quantization methods, removes redundant parameters and low-contribution neuron connections by combining structured pruning and unstructured pruning, and uses knowledge distillation technology to construct a lightweight student network with a parameter reduction of more than 70%. It configures the embedded platform runtime environment and hardware acceleration optimization, establishes a complete real-time processing link from signal acquisition to identification result output, and uses GPU, DSP or dedicated AI accelerators for parallel computing and pipeline processing, thereby realizing efficient deployment and real-time operation of deep learning identification models on embedded platforms. While keeping the identification performance loss less than 3%, it significantly reduces the model storage space by 60-80% and the parameter amount by 30-50%, ensuring that the real-time preprocessing, feature extraction and identification classification of waveform data can be completed within 100 microseconds, and achieving real-time identification performance of processing more than 10,000 waveform samples per second.

[0046] The present invention also provides a CLYC neutron-gamma ray discrimination detection device based on deep learning, the device comprising: The data processing module is used to pre-process the original pulse waveform output by the CLYC detector to obtain clean standard waveform data; The feature extraction module is used to build a one-dimensional convolutional neural network model. It uses multiple convolution kernels of different scales to extract features from clean standard waveform data, capturing features at different time scales of the waveform and identifying the most discriminative feature areas in the waveform through an attention mechanism. A parameter optimization module, configured to adaptively adjust the original hyperparameters of the one-dimensional convolutional neural network model using a particle swarm optimization algorithm to obtain optimized hyperparameters; A model training module is used to perform end-to-end deep learning training on the one-dimensional convolutional neural network model based on optimized hyperparameters, establish a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type, verify the discrimination performance of the trained model, and obtain a trained neutron-gamma ray discrimination model; The model deployment module is used to compress and optimize the neutron-gamma ray discrimination model and deploy it to the embedded platform for real-time neutron-gamma ray discrimination.

[0047] Specifically, a deep learning-based CLYC neutron-gamma ray discrimination detection device in this embodiment obtains clean standard waveform data by constructing a data processing module to preprocess the original pulse waveform, constructs a feature extraction module that uses a multi-scale one-dimensional convolutional neural network combined with an attention mechanism to automatically extract deep-level discrimination features of waveforms at different time scales, constructs a parameter optimization module that uses a particle swarm optimization algorithm to achieve adaptive adjustment of network hyperparameters, constructs a model training module that performs end-to-end deep learning training based on optimized hyperparameters to establish a direct mapping relationship, and constructs a model deployment module for model compression optimization and real-time deployment on an embedded platform, ensuring full-process automated processing from original waveform data acquisition to final discrimination result output, improving the neutron-gamma ray discrimination accuracy and processing efficiency, and overcoming the limitations of traditional methods that require complex manual parameter adjustment and feature engineering.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A CLYC neutron-gamma ray discrimination detection method based on deep learning, characterized in that: The following steps are involved: Preprocess the original pulse waveform output by the CLYC detector to obtain clean standard waveform data; A one-dimensional convolutional neural network model is constructed, which uses multiple convolution kernels of different scales to extract features from clean standard waveform data, capturing features at different time scales of the waveform and identifying the most discriminative feature regions in the waveform through an attention mechanism. Adopting a particle swarm optimization algorithm to adaptively adjust the original hyperparameters of the one-dimensional convolutional neural network model to obtain optimized hyperparameters; Performing end-to-end deep learning training on the one-dimensional convolutional neural network model based on optimized hyperparameters, establishing a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type, verifying the discrimination performance of the trained model, and obtaining a trained neutron-gamma ray discrimination model; The neutron-gamma ray discrimination model is compressed and optimized, and deployed to an embedded platform for real-time neutron-gamma ray discrimination.

2. The CLYC neutron-gamma ray discrimination detection method based on deep learning according to claim 1, characterized in that: The one-dimensional convolutional neural network model is constructed, and multiple convolution kernels of different scales are used to extract features from clean standard waveform data, respectively capturing features of waveforms at different time scales, and identifying the most discriminative feature areas in the waveform through an attention mechanism, including: Constructing a multi-scale convolutional feature extraction network and establishing a multi-scale one-dimensional convolutional neural network architecture, the multi-scale one-dimensional convolutional neural network architecture includes a short-time convolution layer, a medium-time convolution layer, and a long-time convolution layer. Convolution kernels of different scales are used to perform parallel feature extraction on the clean standard waveform data to obtain the short-time scale features, medium-time scale features, and long-time scale features of the waveform. At the same time, the energy information and time information of the waveform are extracted to obtain waveform feature maps at multiple scales. An adaptive attention mechanism is integrated. Based on the luminescence mechanism characteristics of CLYC crystals, a query-key-value-value attention algorithm is used to perform weighted fusion of multi-scale feature maps. The energy-time two-dimensional information is combined to identify and highlight the most discriminative time regions and feature components in the waveform for neutron-gamma discrimination, and an enhanced feature representation is obtained after fusion.

3. The CLYC neutron-gamma ray discrimination detection method based on deep learning according to claim 2, characterized in that: The short-time convolution layer uses a one-dimensional convolution operation with a convolution kernel size of 7 and 64 filters to capture the rapid change characteristics and rising edge characteristics of the waveform, corresponding to the rapid response of the CVL luminescence component of the CLYC crystal; The medium-time convolution layer uses a one-dimensional convolution operation with a convolution kernel size of 15 and 128 filters to capture the medium time scale characteristics of the waveform and the shape characteristics of the pulse body, corresponding to the medium response of the Ce luminescence component; The long-term convolution layer uses a one-dimensional convolution operation with a convolution kernel size of 31 and 256 filters to capture the slow decay characteristics and long-time scale variation patterns of the waveform, corresponding to the slow response of the STE luminescence component; The deposition energy information is extracted from the waveform data through integration operation. Each convolution layer is configured with a batch normalization layer and a ReLU activation function, and downsampling is performed through a maximum pooling layer.

4. The CLYC neutron-gamma ray discrimination detection method based on deep learning according to claim 2, characterized in that: The query-key-value-value attention algorithm adopts a multi-head attention mechanism to map the feature map of each scale into a query matrix, a key-value matrix and a value matrix respectively, and incorporates energy-time two-dimensional information. Based on the different response characteristics of neutrons and gamma rays in CLYC crystals, the attention weight matrix is ​​obtained by calculating the similarity between the query matrix and the key-value matrix, and the attention-enhanced features of each scale are obtained by weighted summation of the value matrix. The original features are fused with the attention-enhanced features using residual connection and layer normalization techniques. The enhanced features of different scales are weighted and combined through a dynamic weight allocation strategy based on particle physics properties to obtain a fused and unified enhanced feature representation.

5. The CLYC neutron-gamma ray discrimination detection method based on deep learning according to claim 1, characterized in that: The particle swarm optimization algorithm is used to adaptively adjust the original hyperparameters of the one-dimensional convolutional neural network model to obtain optimized hyperparameters, including: Constructing a particle swarm optimization model and establishing a parameter space mapping relationship of the particle swarm optimization algorithm, wherein the particle swarm optimization model includes a learning rate, a batch size, a random dropout rate, and a search range of a regularization coefficient, initializing the position and velocity vectors of the particle swarm, and constructing a fitness evaluation function, wherein the fitness evaluation function uses the model identification accuracy as a constraint condition; The optimal hyperparameter combination is searched through the iterative process of the particle swarm optimization algorithm. The position and velocity of the particles are updated according to the fitness evaluation results in each iteration until the global optimal solution is converged, and the optimized hyperparameters suitable for the CLYC neutron-gamma discrimination task are obtained.

6. The CLYC neutron-gamma ray discrimination detection method based on deep learning according to claim 5, characterized in that: The method further includes adaptively adjusting the original hyperparameters of the one-dimensional convolutional neural network model using a particle swarm optimization algorithm to obtain optimized hyperparameters. Perform particle swarm iterative optimization. In each iteration, the one-dimensional convolutional neural network model is trained using the current hyperparameter combination and the fitness score is calculated. The position and velocity of the particles are adjusted according to the standard update rule of the particle swarm optimization algorithm, and the individual optimal position and the global optimal position are updated, gradually converging to the optimal hyperparameter combination. Output optimized hyperparameters, where the optimized hyperparameters are a combination of hyperparameters corresponding to the global optimal position, and the optimized hyperparameters include an optimal learning rate, an optimal batch size, an optimal random dropout rate, and an optimal regularization coefficient.

7. The CLYC neutron-gamma ray discrimination detection method based on deep learning according to claim 1, characterized in that: The one-dimensional convolutional neural network model is subjected to end-to-end deep learning training based on optimized hyperparameters, a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type is established, and the discrimination performance of the trained model is verified to obtain a trained neutron-gamma ray discrimination model, including: The one-dimensional convolutional neural network model is configured by optimizing hyperparameters, the cross-entropy loss function and the Adam optimizer are used for model training, and the prediction results are calculated by forward propagation and the network weights are updated by backpropagation to establish a direct mapping relationship from the clean standard waveform to the neutron-gamma particle type label; An independent test data set is used to verify the discrimination performance of the trained model, and the model's evaluation indicators, including accuracy, precision, recall rate, and F1 score, are calculated. The model's discrimination performance in different energy ranges and waveform conditions is analyzed. When the model meets the expected neutron-gamma ray discrimination requirements, the trained neutron-gamma ray discrimination model is output.

8. The CLYC neutron-gamma ray discrimination detection method based on deep learning according to claim 7, characterized in that: The method further includes performing end-to-end deep learning training on the one-dimensional convolutional neural network model based on optimized hyperparameters, establishing a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type, verifying the discrimination performance of the trained model, and obtaining a trained neutron-gamma ray discrimination model. The clean standard waveform data is divided into training, validation, and test sets, and the training data is randomly shuffled and batched. The model training parameters, including the learning rate, batch size, random dropout rate, and regularization coefficient, are set using optimized hyperparameters. The training process is performed in an end-to-end manner, inputting the clean standard waveform data, extracting features through multi-scale convolution and using an attention mechanism to obtain feature representations. Neutron-gamma binary classification results are output through a fully connected classification layer, and the prediction error is calculated using a cross-entropy loss function. The network parameters are updated using the back-propagation algorithm of the Adam optimizer to gradually optimize the model's discrimination performance. The trained model was tested for discrimination performance using an independent test dataset containing neutron and gamma-ray waveform samples across different energy ranges and signal-to-noise ratios. The confusion matrix of the model's predictions was statistically analyzed, and evaluation metrics including accuracy, precision, recall, F1 score, and area under the receiver operating characteristic (ROC) curve were calculated. The model's ability to discriminate between neutrons and gamma rays was analyzed, and its stability and robustness under different conditions were evaluated.

9. The CLYC neutron-gamma ray discrimination detection method based on deep learning according to claim 1, characterized in that: The neutron-gamma ray discrimination model is compressed and optimized, and deployed on an embedded platform for real-time neutron-gamma ray discrimination, including: The neutron-gamma ray discrimination model was lightweighted by using model quantization, parameter pruning, and knowledge distillation to reduce the model's storage space and computational complexity. While maintaining discrimination performance, a compressed and optimized model suitable for deployment on embedded platforms was obtained. The compression optimization model is transplanted to the embedded hardware platform, and model format conversion, runtime environment configuration and hardware acceleration optimization are performed to establish a real-time data acquisition and processing process.

10. A CLYC neutron-gamma ray discrimination detection device based on deep learning, used to perform a CLYC neutron-gamma ray discrimination detection method based on deep learning according to any one of claims 1 to 9, characterized in that: The device comprises: The data processing module is used to pre-process the original pulse waveform output by the CLYC detector to obtain clean standard waveform data; The feature extraction module is used to build a one-dimensional convolutional neural network model. It uses multiple convolution kernels of different scales to extract features from clean standard waveform data, capturing features at different time scales of the waveform and identifying the most discriminative feature areas in the waveform through an attention mechanism. A parameter optimization module, configured to adaptively adjust the original hyperparameters of the one-dimensional convolutional neural network model using a particle swarm optimization algorithm to obtain optimized hyperparameters; A model training module is used to perform end-to-end deep learning training on the one-dimensional convolutional neural network model based on optimized hyperparameters, establish a direct mapping relationship from the original pulse waveform to the neutron-gamma particle type, verify the discrimination performance of the trained model, and obtain a trained neutron-gamma ray discrimination model; The model deployment module is used to compress and optimize the neutron-gamma ray discrimination model and deploy it to the embedded platform for real-time neutron-gamma ray discrimination.

Images