Nuclear pulse signal discrimination method and system

Abstract

The invention discloses a nuclear pulse signal discrimination method and system, and belongs to the field of nuclear radiation detection.The nuclear pulse signal discrimination method comprises the steps that firstly, a nuclear pulse data set is constructed in two modes, one mode is that a simulated nuclear pulse generator is used for generating pulse signals for determining ray types, and the pulse signals are collected through an FPGA + ADC hardware system; secondly, Python software is used for simulating and generating pulses for determining ray types according to actual nuclear pulse signal characteristics; then, a CNN model is constructed based on the LeNet network, and a triple loss function is adopted for model training; and finally, similarity matching is carried out through measurement by means of feature vectors output by the neural network, and discrimination and classification of different types of nuclear pulse signals are realized. According to the method, the nuclear pulse signal features are automatically extracted through the deep metric learning model, the limitation of dependence on artificial feature extraction in a traditional method is overcome, the accuracy of nuclear pulse signal discrimination is improved, pulse signals generated by different types of rays can be effectively distinguished, and an efficient and reliable technical scheme is provided for nuclear radiation detection.

Description

Technical Field

[0001] This application belongs to the field of nuclear radiation detection, and particularly relates to a method and system for discriminating nuclear pulse signals. Background Art

[0002] In the field of nuclear radiation detection, accurately distinguishing pulse signals generated by different types of rays (such as neutrons, gamma rays, etc.) is crucial for applications such as nuclear safety monitoring, radioactive material identification, nuclear physics experiments, and medical diagnosis. Traditional methods for discriminating nuclear pulse signals mainly rely on Pulse Shape Discrimination (PSD) technology. This technology utilizes the different mechanisms of different rays depositing energy in the scintillator, resulting in differences in the shapes of the generated optical pulses for discrimination. Common PSD methods include the charge comparison method, the zero-crossing method, the time-to-digital conversion method, etc.

[0003] However, these traditional methods have the following limitations: 1) sensitive to noise, and baseline drift will seriously affect the discrimination effect; 2) require manual setting of threshold parameters, and the discrimination effect is greatly affected by subjective factors; 3) the discrimination effect significantly decreases when signals overlap or the signal intensity is weak; 4) limited ability to discriminate nuclear pulse signals in complex environments.

[0004] Therefore, there is an urgent need to develop an efficient and accurate method for discriminating nuclear pulse signals, which can overcome the limitations of traditional methods and existing deep learning methods and achieve accurate discrimination of different types of nuclear pulse signals. Summary of the Invention

[0005] The purpose of this application is: to overcome the problems of the prior art, a method and system for discriminating nuclear pulse signals are disclosed. The method of this application realizes efficient and accurate discrimination of different types of nuclear pulse signals by constructing a high-quality nuclear pulse data set, combining an improved CNN model and a triplet loss function, and solves the problems that traditional methods are sensitive to noise, rely on manual setting of thresholds, and existing deep learning methods have insufficient generalization ability and single training objectives.

[0006] On the one hand, the purpose of this application is achieved through the following technical solutions:

[0007] A method for discriminating nuclear pulse signals, the method for discriminating nuclear pulse signals includes:

[0008] S1: Construct a nuclear pulse data set, including:

[0009] S1.1: Use an analog nuclear pulse generator to generate pulse signals of preset ray types according to pulse shape characteristics, and collect the generated pulses;

[0010] S1.2: Simulate and generate pulse signals of preset ray types according to the characteristics of actual nuclear pulse signals;

[0011] S1.3: Combine the data obtained in S1.1 and S1.2 to form a complete training database;

[0012] Step S2: Model construction and training, including:

[0013] S2.1: Construct a CNN model based on the LeNet network architecture;

[0014] S2.2: Use the triplet loss function to train the CNN model so that the convolutional neural network can learn to distinguish the feature representations of different types of nuclear pulse signals;

[0015] Step S3: Inference and verification, including:

[0016] S3.1: Input the nuclear pulse signal to be recognized into the trained CNN model to obtain the corresponding feature vector;

[0017] S3.2: Use a metric method to calculate the similarity between the feature vector of the signal to be recognized and the standard feature vector;

[0018] S3.3: According to the similarity matching result, realize the discrimination and classification of the nuclear pulse signal to be recognized.

[0019] According to a preferred embodiment, in step S1.1, the pulse signals generated by the nuclear pulse generator include neutron pulse signals and gamma pulse signals.

[0020] According to a preferred embodiment, in step S1.1, the generated pulses are collected by the FPGA+ADC hardware acquisition system.

[0021] According to a preferred embodiment, in step S1.2, the pulse signals are simulated and generated by Python software, and the simulated pulse signals contain the characteristics of actual nuclear pulse signals, including amplitude, rise time, fall time, and pulse width parameters.

[0022] According to a preferred embodiment, in S1.3, the obtained data is preprocessed, including:

[0023] a. Normalization processing to unify the amplitude range of all signals;

[0024] b. Denoising processing to remove the high-frequency noise components in the signals;

[0025] c. Sampling alignment to align all pulse signals on the time axis;

[0026] d. Data augmentation, including expanding the dataset size by adding noise and time shift with different degrees.

[0027] According to a preferred embodiment, in step S2.1, the CNN model structure based on LeNet includes:

[0028] a. Input layer, configured to receive the preprocessed nuclear pulse signal;

[0029] b. Convolution module, composed of several convolutional layers, and a ReLU activation function and a pooling layer are respectively provided after each convolutional layer;

[0030] c. Fully connected layer: mapping the features extracted by the convolution module to the feature space;

[0031] d. Output layer: outputting a feature vector representing the features of the nuclear pulse signal.

[0032] According to a preferred embodiment, in step S2.2, the triplet loss function is defined as:

[0033] L = max(0, D(a, p) - D(a, n) + margin)

[0034] where a represents the anchor sample, p represents the positive sample of the same class as the anchor sample, n represents the negative sample of a different class from the anchor sample, D represents the distance function between feature vectors, and margin is the set boundary value used to control the distance difference between positive and negative samples.

[0035] According to a preferred embodiment, in step S3.2, the measurement methods adopted include Euclidean distance, cosine similarity, and Mahalanobis distance.

[0036] According to a preferred embodiment, the nuclear pulse signal discrimination method further includes: S4: optimizing the trained model to improve the accuracy and generalization ability of the model by adjusting the network structure, parameters, and optimization algorithm.

[0037] On the other hand, the present application also discloses:

[0038] A nuclear pulse signal discrimination system, which includes an FPGA platform, on which a CNN model is deployed, and the aforementioned nuclear pulse signal discrimination method is used for online nuclear pulse signal discrimination.

[0039] The main solution of the present application and its various further selection solutions can be freely combined to form multiple solutions, all of which are solutions that can be adopted and claimed by the present application. Those skilled in the art can understand that there are various combinations according to the prior art and common general knowledge after understanding the solution of the present application, and all of them are the technical solutions to be protected by the present application, which will not be enumerated here.

[0040] Advantages of this application:

[0041] 1. By combining physical acquisition with software simulation of the nuclear pulse generator, this application effectively solves the problem of difficult acquisition of labeled data for nuclear pulse signals, providing a high-quality and diverse data source for the training of deep learning models.

[0042] 2. This application uses a CNN model based on the LeNet architecture and combines triplet loss function for training, which can automatically learn the feature representations of different types of nuclear pulse signals, overcoming the limitation of manual feature design in traditional methods.

[0043] 3. Through the metric learning method of the triplet loss function, this application enables the model to learn more discriminative feature representations, improving the discrimination ability for different types of nuclear pulse signals, especially under complex conditions such as low signal-to-noise ratio and signal overlap.

[0044] 4. This application uses the method of feature vector similarity matching for nuclear pulse signal discrimination, which has good scalability and can easily add new types of nuclear pulse signals without retraining the entire network.

[0045] 5. The method of this application can be deployed on the FPGA platform to achieve online real-time nuclear pulse signal discrimination, meeting the requirements of real-time and reliability for nuclear radiation detection systems. Brief Description of the Drawings

[0046] Figure 1 is the overall flowchart of a method for discriminating nuclear pulse signals in this application;

[0047] Figure 2 is a schematic diagram of the dataset construction method in this application;

[0048] Figure 3 is a schematic diagram of training with the triplet loss function in this application;

[0049] Figure 4 is a schematic diagram of feature vector similarity matching in this application;

[0050] Figure 5 is the principle block diagram of the nuclear pulse signal discrimination system in this application. Detailed Embodiments

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0052] Accordingly, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0053] Embodiment 1

[0054] As Figure 1 shown, this embodiment discloses a method for discriminating nuclear pulse signals, including the following steps: Step S1: Construct a nuclear pulse data set; Step S2: Model construction and training; Step S3: Inference verification; Step S4: Optimize the trained CNN model.

[0055] As Figure 2 shown, Step S1 constructs a nuclear pulse data set in two ways. One is to use an artificial nuclear pulse generator combined with a hardware acquisition system to obtain physical pulse data, and the other is to use Python software simulation to generate nuclear pulse signal data. The two methods complement each other to form a complete training database.

[0056] Specifically: Step S1.1: Use an artificial nuclear pulse generator to generate pulse signals of a certain ray type according to the pulse shape characteristics, and collect the generated pulses through an FPGA + ADC hardware acquisition system.

[0057] In this embodiment, an existing artificial nuclear pulse generator in the laboratory is used. This generator can simulate and generate nuclear pulse signals caused by different rays according to preset parameters, including neutron pulse signals and gamma pulse signals. For neutron and gamma pulses, different rise times, fall times, and tail characteristic parameters are set respectively to make the generated pulse signals have typical neutron and gamma ray characteristics.

[0058] The hardware acquisition system consists of an FPGA and a high-speed ADC. In this embodiment, an Xilinx ZYNQ series FPGA chip and an ADC with a 16-bit resolution and a maximum sampling rate of 250 MSPS are used for data acquisition. The working process of the acquisition system is as follows: First, the analog signal generated by the artificial nuclear pulse generator is amplified by a preamplifier and then input to the ADC for analog-to-digital conversion; then, the FPGA reads the digital output of the ADC, caches and preprocesses the collected data; finally, the processed data is transmitted to a computer through a USB or Ethernet interface for storage and analysis.

[0059] In this embodiment, 10,000 samples were collected for neutron and gamma pulses respectively. The sampling rate was set to 100 MSPS, and 1024 points were collected for each pulse, covering the complete shape of the pulse. To increase the diversity of the data, during the collection process, by adjusting the parameters of the pseudo-nuclear pulse generator, the generated pulses had a certain range of variations in terms of amplitude, rise time, and fall time.

[0060] Specifically, step S1.2: Use Python software to simulate and generate pulse signals for determining the type of ray according to the characteristics of the actual nuclear pulse signal.

[0061] To further expand the dataset and enhance the generalization ability of the model, this embodiment uses Python software to perform simulation generation according to the characteristics of the actual nuclear pulse signal. During the simulation process, typical characteristics of the actual nuclear pulse signal are considered, including:

[0062] a. Neutron pulse: Usually has a relatively long tail decay time and can be simulated using a double-exponential function:

[0063]

[0064] where A is the amplitude coefficient, τ1 is the fall time constant, τ2 is the rise time constant, and τ1 >> τ2.

[0065] b. Gamma pulse: Usually has a relatively fast decay characteristic and can be simulated using the following function:

[0066]

[0067] where the values of τ1 and τ2 are smaller compared to the neutron pulse, making the pulse decay faster.

[0068] In the Python simulation, by randomly adjusting the parameters of A, τ1, and τ2, neutron and gamma pulse signals with various different characteristics were generated. In addition, different degrees of Gaussian white noise were added to the simulation signals to simulate the noise interference in the actual signals. In this way, a total of 20,000 simulation pulse signals were generated, including 10,000 neutron pulses and 10,000 gamma pulses.

[0069] S1.3: Combine the data obtained in S1.1 and S1.2 to form a complete training database.

[0070] Combine the data obtained through hardware collection and software simulation to form a complete dataset containing 40,000 nuclear pulse signal samples. Before combination, all data is preprocessed, including:

[0071] a. Normalization processing: Normalize the amplitudes of all pulse signals to the range [0, 1] to eliminate the amplitude differences caused by different amplification factors;

[0072] b. Denoising processing: Use wavelet transform or digital filters to denoise the signals and remove high-frequency noise components;

[0073] c. Sampling alignment: Ensure that all pulse signals are aligned on the time axis, with the pulse rising edge as the reference point for alignment;

[0074] d. Data augmentation: Expand the dataset size by adding different levels of noise, small time shifts, amplitude scaling, etc., to improve the robustness of the model.

[0075] After preprocessing, the dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1, which are used for model training, hyperparameter tuning, and performance evaluation respectively.

[0076] As Figure 3 shown, step S2 constructs a CNN model based on the LeNet network architecture and uses a triplet loss function for training, enabling the network to learn the feature representations for distinguishing different types of nuclear pulse signals.

[0077] Specifically, step S2.1: Construct a CNN model based on the LeNet network architecture.

[0078] LeNet is a classic convolutional neural network structure. In this embodiment, the LeNet network is appropriately modified according to the characteristics of nuclear pulse signals, and the specific structure is as follows:

[0079] a. Input layer: Receive the preprocessed nuclear pulse signals, with an input dimension of 1×1024 (single channel, each pulse contains 1024 sampling points);

[0080] b. First convolutional layer: Use 16 convolutional kernels of size 1×5, with a stride of 1, and an output dimension of 16×1020;

[0081] c. First pooling layer: Use max pooling of size 1×2, with a stride of 2, and an output dimension of 16×510;

[0082] d. Second convolutional layer: Use 32 convolutional kernels of size 1×5, with a stride of 1, and an output dimension of 32×506;

[0083] e. Second pooling layer: Use max pooling of size 1×2, with a stride of 2, and an output dimension of 32×253;

[0084] f. Third convolutional layer: Use 64 convolutional kernels of size 1×5, with a stride of 1, and an output dimension of 64×249;

[0085] g. Third pooling layer: Use max pooling with a size of 1×2, a stride of 2, and an output dimension of 64×124;

[0086] h. First fully connected layer: The input dimension is 64×124, and the output dimension is 256;

[0087] i. Second fully connected layer: The input dimension is 256, and the output dimension is 128. The output of this layer is the feature vector representing the characteristics of the nuclear pulse signal;

[0088] ReLU activation function is used after each convolutional layer and fully connected layer to introduce non-linearity. At the same time, a Dropout layer (dropout rate is 0.5) is added after the fully connected layer to prevent overfitting.

[0089] Step S2.2: Train the model using the triplet loss function.

[0090] The triplet loss function is a loss function commonly used in metric learning. Its core idea is to make the distance between similar samples closer in the feature space and the distance between different samples farther. In this embodiment, a triplet consists of an anchor sample (a), a positive sample (p), and a negative sample (n), where a and p are nuclear pulse signals of the same type (such as both neutron pulses), and a and n are nuclear pulse signals of different types (such as a is a neutron pulse and n is a gamma pulse).

[0091] The triplet loss function is defined as:

[0092] L = max(0, D(a, p) - D(a, n) + margin)

[0093] where D represents the distance function between feature vectors. In this embodiment, the Euclidean distance is used; margin is the set boundary value used to control the distance difference between positive and negative samples, which is set to 0.5 in this embodiment.

[0094] During the training process, 32 anchor samples are randomly selected in each batch, and 1 positive sample and 1 negative sample are selected for each anchor sample to form 32 triplets. By optimizing the triplet loss function, the network learns a discriminative feature representation. The training uses the Adam optimizer, the initial learning rate is set to 0.001, and a learning rate decay strategy is used. The learning rate is halved every 20 epochs. The model training is carried out for a total of 100 epochs, and the model with the best performance on the validation set is selected as the final model.

[0095] As Figure 4 shown, in step S3, the nuclear pulse signal to be recognized is input into the trained CNN model to obtain the feature vector, and the discrimination and classification are realized through similarity matching.

[0096] Step S3.1: Input the nuclear pulse signal to be recognized into the trained CNN model to obtain the corresponding feature vector.

[0097] For the nuclear pulse signal to be recognized, first perform the same preprocessing as the training data, and then input it into the trained CNN model. Obtain the feature vector through forward propagation (i.e., the output of the fully connected layer 2, with a dimension of 128).

[0098] Step S3.2: Use a metric method to calculate the similarity between the feature vector of the signal to be recognized and the standard feature vector.

[0099] To realize the discrimination and classification of nuclear pulse signals, a standard feature vector library needs to be established in advance for each type of nuclear pulse signal. In this embodiment, by performing clustering analysis on the feature vectors of each type of pulse signal in the training set, multiple standard feature vectors are established for neutron pulses and gamma pulses respectively to cover the pulse signals with different features in the same type.

[0100] During the discrimination process, calculate the similarity between the feature vector of the signal to be recognized and each standard feature vector. In this embodiment, the Euclidean distance is used as the similarity metric method, and the calculation formula is:

[0101] D(x,y)=sqrt(sum((x i -y i ) 2 ))

[0102] where x and y represent the feature vector of the signal to be recognized and the standard feature vector respectively, and x i and y i represent the i-th element of the vector respectively.

[0103] In addition to the Euclidean distance, other metric methods such as cosine similarity or Mahalanobis distance can also be used. The cosine similarity calculation formula is:

[0104] η t =η min +0.5×(η max -η min )×(1+cos(t / T×π))

[0105] where x·y represents the dot product of vectors x and y, and ||x|| represents the L2 norm of vector x.

[0106] Step S3.3: According to the similarity matching result, realize the discrimination and classification of the nuclear pulse signal to be recognized.

[0107] According to the calculated similarity, the signal to be recognized is classified into the category corresponding to the standard feature vector with the closest distance (or the highest similarity) to its feature vector. To improve the reliability of discrimination, a similarity threshold can be set, and a discrimination decision is made only when the highest similarity exceeds the threshold; otherwise, the signal is marked as "unknown type".

[0108] In this embodiment, by adjusting the similarity threshold, a trade-off can be made between the discrimination accuracy and the recognition rate. A higher threshold will increase the discrimination accuracy but decrease the recognition rate, and vice versa for a lower threshold. According to the requirements of specific application scenarios, the threshold can be flexibly adjusted.

[0109] The verification results on the test set show that this method achieves an accuracy of 97.8% in the neutron and gamma pulse discrimination task, significantly superior to the traditional pulse shape discrimination method. Especially under the condition of low signal-to-noise ratio (SNR = 3 dB), this method can still maintain an accuracy of 93.5%, while the accuracy of the traditional method drops below 80%.

[0110] Step S4: Optimize the trained CNN model.

[0111] To improve the accuracy and generalization ability of the model, the following optimizations are carried out on the model:

[0112] a. Structural optimization: Optimize the network structure parameters through the grid search method, including the convolution kernel size, the number of convolutional layers, the number of convolution kernels, and the number of neurons in the fully connected layer. After optimization, it is determined that the first convolutional layer uses 16 convolution kernels of size 1×7, the second convolutional layer uses 32 convolution kernels of size 1×5, and the third convolutional layer uses 64 convolution kernels of size 1×3 to better capture the time features of the nuclear pulse signal.

[0113] b. Attention mechanism optimization: Introduce the self-attention mechanism after the convolutional layer, enabling the network to automatically focus on the key time period features in the nuclear pulse signal, especially the features in areas such as the pulse rising edge and the tail decay, to improve the discrimination ability for different types of nuclear pulse signals. The calculation formula of the self-attention mechanism is:

[0114] Attention(Q,K,V)=softmax((Q·K T ) / √d k )·V

[0115] where Q, K, and V respectively represent the query matrix, the key matrix, and the value matrix, which are obtained by linear transformation of the input features, and d k is the dimension of the key vector.

[0116] c. Regularization Optimization: In addition to Dropout, L2 regularization (weight decay coefficient is 0.0001) and Batch Normalization techniques are introduced to effectively suppress overfitting and improve the generalization ability of the model.

[0117] d. Learning Rate Optimization: The cosine annealing learning rate scheduling strategy is adopted to make the learning rate change periodically during training, avoiding the model falling into local optimal solutions. The formula is:

[0118] η t =η min +0.5×(η max -η min )×(1 + cos(t / T×π))

[0119] where η t represents the learning rate of the t-th epoch, η min and η max represent the minimum and maximum learning rates respectively, and T represents the cycle length.

[0120] e. Data Balance Optimization: The weighted sampling strategy is adopted to ensure that different types of nuclear pulse signals are evenly sampled during training, avoiding bias in the model due to sample imbalance.

[0121] f. Triplet Mining Optimization: The Hard Triplet Mining strategy is adopted to preferentially select difficult triplet samples for training, that is, select negative samples close to the anchor sample and positive samples far from the anchor sample to construct triplets, accelerating model convergence and improving discrimination ability. The specific implementation method is that for each anchor sample, select the K farthest samples from the same-class samples as candidate positive samples, select the K nearest samples from different-class samples as candidate negative samples, and then randomly select positive and negative samples from these candidate samples to construct triplets.

[0122] After the above optimizations, the accuracy of the model on the test set is increased from 97.8% to 99.2%, and the accuracy under low signal-to-noise ratio conditions (SNR = 3dB) is also increased from 93.5% to 96.8%, significantly superior to the performance of the model before optimization.

[0123] Embodiment 2

[0124] This embodiment discloses a nuclear pulse signal discrimination system. As Figure 5 shown, the principle block diagram of the nuclear pulse signal discrimination system is disclosed. Based on Embodiment 1, the optimized CNN model is deployed to the FPGA platform, and the nuclear pulse signal discrimination method described in Embodiment 1 is used for online nuclear pulse signal discrimination.

[0125] The process of deploying the optimized CNN model to the FPGA platform includes:

[0126] a. Model quantization: Quantize the optimized CNN model, convert floating-point parameters to fixed-point parameters, and reduce the occupancy of computing resources;

[0127] b. Hardware acceleration: Utilize the parallel computing ability of the FPGA to accelerate the convolution operation and feature extraction process;

[0128] c. Real-time inference: Deploy the quantized model to the FPGA platform to achieve real-time discrimination of nuclear pulse signals. Transmit the discrimination results to the host computer through a high-speed interface (such as PCIe) for display and analysis;

[0129] d. Performance evaluation: Test the performance of the deployed system in an actual nuclear radiation detection environment, and verify its discrimination accuracy and real-time performance under different signal-to-noise ratios and signal intensities.

[0130] Thus, through this nuclear pulse signal discrimination system, online real-time discrimination of nuclear pulse signals is achieved, meeting the requirements of real-time performance and reliability for nuclear radiation detection systems.

[0131] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. A method for discriminating nuclear pulse signals, characterized in that, The nuclear pulse signal discrimination method includes: S1: Construct a nuclear pulse data set, including: S1.1: Use an artificial nuclear pulse generator to generate pulse signals of preset ray types according to pulse shape characteristics, and collect the generated pulses; S1.2: Simulate and generate pulse signals of preset ray types according to the characteristics of actual nuclear pulse signals; S1.3: Combine the data obtained in S1.1 and S1.2 to form a complete training database; Step S2: Model construction and training, including: S2.1: Construct a CNN model based on the LeNet network architecture; S2.2: Use a triplet loss function to train the CNN model so that the convolutional neural network can learn to distinguish the feature representations of different types of nuclear pulse signals; Step S3: Inference and verification, including: S3.1: Input the nuclear pulse signal to be identified into the trained CNN model to obtain the corresponding feature vector; S3.2: Use a metric method to calculate the similarity between the feature vector of the signal to be identified and the standard feature vector; S3.3: According to the similarity matching result, realize the discrimination and classification of the nuclear pulse signal to be identified.

2. The nuclear pulse signal discrimination method according to claim 1, characterized in that In step S1.1, the pulse signals generated by the artificial nuclear pulse generator include neutron pulse signals and gamma pulse signals.

3. The nuclear pulse signal discrimination method according to claim 1 or 2, characterized in that In step S1.1, the generated pulses are collected by an FPGA+ADC hardware acquisition system.

4. The nuclear pulse signal discrimination method according to claim 1, wherein In step S1.2, the pulse signals are simulated and generated by Python software, and the simulated pulse signals contain the characteristics of actual nuclear pulse signals, including amplitude, rise time, fall time, and pulse width parameters.

5. The nuclear pulse signal discrimination method according to claim 1, wherein In S1.3, the obtained data is preprocessed, including: a. Normalization processing to unify the amplitude range of all signals; b. Denoising processing to remove the high-frequency noise components in the signals; c. Sampling alignment to align all pulse signals on the time axis; d. Data augmentation, including expanding the data set scale by adding different levels of noise and time shift.

6. The nuclear pulse signal discrimination method according to claim 1, characterized in that In step S2.1, the CNN model structure constructed based on LeNet includes: a. Input layer, configured to receive the preprocessed nuclear pulse signals; b. Convolution module, composed of several convolutional layers, and a ReLU activation function and a pooling layer are respectively provided after each convolutional layer; c. Fully connected layer: Map the features extracted by the convolution module to the feature space; d. Output layer: Output the feature vector representing the characteristics of the nuclear pulse signal.

7. The nuclear pulse signal discrimination method according to claim 1, wherein In step S2.2, the triplet loss function is defined as: L = max(0, D(a, p) - D(a, n) + margin) where a represents the anchor sample, p represents the positive sample of the same class as the anchor sample, n represents the negative sample of a different class from the anchor sample, D represents the distance function between feature vectors, and margin is the set boundary value used to control the distance difference between positive and negative samples.

8. The nuclear pulse signal discrimination method according to claim 1, characterized in that In step S3.2, the metric methods used include Euclidean distance, cosine similarity, and Mahalanobis distance.

9. The nuclear pulse signal discrimination method according to claim 1, characterized in that The nuclear pulse signal discrimination method further includes: S4: Optimize the trained CNN model to improve the accuracy and generalization ability of the model by adjusting the network structure, parameters, and optimization algorithms.

10. A nuclear pulse signal discrimination system, characterized in that, The nuclear pulse signal discrimination system includes an FPGA platform, on which a CNN model is deployed, and the nuclear pulse signal discrimination method described in any one of claims 1 to 9 is used for on-line nuclear pulse signal discrimination.

Images