A particle identification method and system based on a one-dimensional convolutional neural network

Abstract

The present invention relates to the field of intelligent detection of space environment, and particularly to a particle discrimination method and system based on a one-dimensional convolutional neural network. The method of the present invention includes: preprocessing the waveform data of various particles in space collected to obtain the time-domain waveform data and frequency-domain waveform data of various particles; simultaneously inputting the time-domain waveform data and frequency-domain waveform data of various particles into a pre-established and trained particle discrimination model to obtain a particle discrimination result; the particle discrimination model is established based on a one-dimensional convolutional neural network. The system of the present invention includes a data preprocessing module and a particle discrimination module; the data preprocessing module preprocesses the waveform data of various particles in space to obtain time-domain waveform normalized data and frequency-domain waveform normalized data; the particle discrimination module simultaneously inputs the time-domain waveform data and frequency-domain waveform data into a pre-established and trained particle discrimination model to obtain a particle discrimination result.

Description

Technical Field

[0001] The present invention relates to the field of intelligent detection of the space environment, and particularly to a particle discrimination method and system based on a one-dimensional convolutional neural network. Background Art

[0002] Charged particles in space include galactic cosmic rays, solar cosmic rays, solar wind, Van Allen radiation belts, etc. Galactic cosmic rays are composed of high-energy charged particle fluxes from all directions in the galaxy, and the main components are protons, alpha particles, and other nuclear components; solar cosmic rays are high-energy charged particle fluxes emitted during solar flares, mainly composed of protons; solar wind is a plasma flux continuously ejected by the sun, mainly composed of hydrogen plasma, including a small amount of other components; the Van Allen radiation belt, also known as the Earth radiation belt, is located near the Earth and is a radiation belt captured by the Earth's magnetic field, divided into an inner radiation belt and an outer radiation belt, mainly composed of protons and electrons; in addition, neutrons and gamma rays also widely exist in the universe. Facing various charged particles and neutral particles in space, only by distinguishing them can we further understand the characteristics and distributions of various particles. Therefore, the space environment particle discrimination technology is crucial.

[0003] At present, on-board particle discrimination mainly includes several common methods:

[0004] The first method is the ΔE-E detector telescope method. The main principle is to use one or more solid-state detectors (SSDs) to form a stack. If the first sensor is thin enough, assuming the thickness is x, the energy deposited after the particle penetrates The second (or multiple) sensors are thick enough to completely deposit the remaining energy of the particle, and then the total energy E of the particle can be measured. By the ΔE vs E two-dimensional spectrum, the particle types can be distinguished. The detector telescope method is simple and reliable and is widely used in actual engineering applications. However, it requires the particle to lose very little energy in the transmission detector, which is less applicable to heavier particles, and its correction will also cause trouble.

[0005] The second method is the electrostatic analysis time-of-flight method (ESA-TOF). The electrostatic analyzer ESA scans the voltage V to screen the energy-to-charge ratio E / q of the incident ions, and then combines with the time-of-flight method to finally obtain the mass-to-charge ratio of the ions. The electrostatic analyzer needle is often used to test low-energy charged particles or neutral atoms. However, due to the limitation of the space size of on-board equipment, for medium- and high-energy particles, it is necessary to greatly increase the physical structure to achieve detection, which cannot be realized in orbit.

[0006] The third method is the time-of-flight energy method (TOF×E), and its basic principle is based on the particle energy formula That is, by measuring only the particle energy E and velocity v (distance l and time t), the particle mass can be obtained, thereby determining the type. TOF×E has a relatively good discrimination effect on light particles, especially low-energy light particles. As the particle mass increases, the mass difference becomes smaller and smaller, making it more difficult to distinguish. In addition, for isobars, that is, nuclides with the same mass number but different atomic numbers, the time-of-flight method cannot be used for discrimination and must be combined with other methods to enable reliable discrimination.

[0007] The fourth method is the rising-edge energy method (PSA×E), which uses high-speed digital acquisition technology to collect the rising edge of the particle waveform and create a two-dimensional spectrogram of energy E and rising-edge time. Although this method attempts to use waveform information, only a small part of the data is used, and all the information of the waveform is not utilized, which may cause deviation of the results due to accidental measurement errors.

[0008] In addition, the above four methods all require downloading the on-orbit scientific data to the ground for particle discrimination analysis work, which is not applicable to satellite resources with limited resources, especially in the field of deep space exploration. Summary of the Invention

[0009] The purpose of the present invention is to solve the problems existing in the prior art, and to invent a particle discrimination method with full waveform analysis and on-orbit processing capabilities, further improving the detection accuracy and reducing the on-orbit resource occupancy rate.

[0010] To achieve the above object, the present invention is realized through the following technical solutions.

[0011] The present invention proposes a particle discrimination method based on a one-dimensional convolutional neural network, and the method includes:

[0012] Preprocess the waveform data of various particles in space collected to obtain the time-domain waveform data and frequency-domain waveform data of various particles;

[0013] Input the time-domain waveform data and frequency-domain waveform data of various particles into a pre-established and trained particle discrimination model at the same time to obtain a particle discrimination result; the particle discrimination model is established based on a one-dimensional convolutional neural network.

[0014] As an improvement of the above technical solution, the various particles include: protons, electrons, heavy particles, neutrons, and gamma rays.

[0015] As an improvement of the above technical solution, the preprocessing includes: time-domain data normalization and frequency-domain conversion and normalization.

[0016] As one of the improvements to the above technical solution, the particle discrimination model includes 3M convolutional neural networks arranged in parallel, where M is a natural number. Each convolutional neural network includes: an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer; every three convolutional neural networks form a group. In the same group, the convolutional kernel sizes of the convolutional neural networks are different, and other parameters between different groups of convolutional neural networks are different. The other parameters include: multi-scale data sets, weight initialization, activation functions, the number of network layers, the number of iterative training times, the learning rate, and the batch size;

[0017] Among them, each input layer includes: a time-domain waveform data input channel and a frequency-domain waveform data input channel arranged in parallel; the convolutional layer gradually screens the features of the input layer data through a multi-channel convolutional kernel, and performs dimensionality reduction and non-linearity on the output data through the pooling layer and relu calculation. Finally, the judgment result is output through the fully connected layer;

[0018] The judgment result is the final classification result determined by the voting method of taking N from 3M;

[0019] As one of the improvements to the above technical solution, the method further includes: training the particle discrimination model; the training process includes:

[0020] Preprocessing the waveform data of various particles obtained from the calibration test and on-orbit detection to obtain time-domain data and frequency-domain data;

[0021] Constructing a training data set and a test data set based on the obtained time-domain data and frequency-domain data;

[0022] Inputting the time-domain data and frequency-domain data of the training data set into the particle discrimination model for training;

[0023] And inputting the time-domain data and frequency-domain data of the test data set into the particle discrimination model for testing to determine the correctness of each layer of convolutional kernels and bias parameters, and obtaining a trained particle discrimination model.

[0024] As one of the improvements to the above technical solution, during the training process, the loss function is calculated through the Cross Entropy Loss(·) function, the reverse gradient is calculated through backward(·), and the training parameters are continuously iterated.

[0025] The present invention also proposes a particle discrimination system based on a one-dimensional convolutional neural network. The system includes: a data preprocessing module and a particle discrimination module; among them,

[0026] The data preprocessing module is used to preprocess the waveform data of various particles in space to obtain time-domain waveform normalized data and frequency-domain waveform normalized data;

[0027] The particle discrimination module is used to input the time-domain waveform data and the frequency-domain waveform data into a pre-established and trained particle discrimination model simultaneously to obtain a particle discrimination result; the particle discrimination model is established based on a one-dimensional convolutional neural network.

[0028] As an improvement of the above technical solution, the system is implemented based on FPGA.

[0029] As an improvement of the above technical solution, the system further includes a data acquisition module, which is used to acquire the waveform data of various particles through on-orbit detection or a ground accelerator and transmit it to the data preprocessing module.

[0030] The advantages of the present invention compared with the prior art are as follows:

[0031] 1. The convolutional neural network is good at separating different features in the data and finally realizing classification;

[0032] 2. The convolutional neural network analyzes the data through the overall waveform, and its data information is richer, which is expected to improve the accuracy;

[0033] 3. The convolutional neural network has good scalability and portability, and can be perfectly integrated and improved with traditional methods;

[0034] 4. The convolutional neural network can further improve the model accuracy by updating the model parameters on orbit;

[0035] 5. FPGA itself has high parallelism, low power consumption, and reconfigurability, and is the best platform for on-board real-time processing. Description of the Drawings

[0036] Figure 1 It is the architecture diagram of the on-board particle discrimination FPGA real-time processing platform based on a one-dimensional convolutional neural network of the present invention;

[0037] Figure 2 It is the systolic array computing architecture diagram. Detailed Embodiments

[0038] The pulse waveforms generated by particles in the sensor have certain differences, and the explanations are as follows:

[0039] For charged particles, when light ions and heavy ions with the same energy enter the detector from the rear end, heavy ions have a shorter range and higher ionization density. The electric field strength gradually increases from the rear end to the front end. Heavy ions have a short range, and their Bragg peak falls in the low electric field region. The low electric field strength and high energy loss rate on the ion track make the plasma erosion time of heavy ions longer. In charge carrier transport, electron mobility is faster and hole mobility is slower. When incident from the rear end, due to the short range of heavy ions, the average distance of hole migration increases, which increases the charge carrier transport time. The above two effects work together, and when the incident energy is constant, the charge collection time changes with the change of particle nuclear charge number and mass number. The current pulses generated by heavier particles have a longer duration, lower amplitude, longer charge rise time, and longer zero crossing time. Therefore, the waveforms generated by different particles after deposition in the sensor will have certain differences. For neutral components, taking neutron gamma as an example, after the organic scintillator is irradiated by neutrons or gamma rays, it produces recoil protons and secondary electrons respectively to emit light. The intensity of the light increases quickly to the maximum, while the decay is relatively slow, which should be similar to exponential decay. Its luminescence decay time contains two components, fast and slow. When neutrons or gamma rays irradiate into the same scintillator, the ionization density formed in the scintillator is different, resulting in different intensity ratios of the fast and slow components of the luminescence decay time. The share of the fast component in the proton-excited fluorescence generated by the interaction of neutrons and scintillators is lower than that of the electron-excited fluorescence generated by the interaction of gamma rays and scintillators, while the share of the slow component is higher than that of gamma rays. It is this factor that makes the decay time of the fluorescence generated by the interaction of neutrons and gamma rays with scintillators different, which in turn leads to different pulse shapes. In summary, the differences in waveform data generated by particle deposition in the sensor provide us with the possibility of identification using convolutional neural networks.

[0040] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0041] Example 1

[0042] Embodiment 1 of the present invention proposes a particle identification method based on a one-dimensional convolutional neural network, the method comprising:

[0043] First, preprocess the waveform data of various particles obtained from calibration tests and on-orbit detections (including protons, electrons, heavy particles, neutrons, gamma rays, etc.), including time-domain data normalization, frequency-domain conversion and normalization, and group the waveform data with different scales and resolutions to achieve data augmentation of the input dataset of the convolutional neural network. The processed time-frequency domain data is divided into a training dataset and a test dataset. The time-domain data and frequency-domain data of the training dataset are input into a one-dimensional convolutional neural network software architecture through two channels of the model for training, and the test dataset is used as a verification means for the training results. Among them, the establishment of the one-dimensional convolutional neural combination network for particle discrimination specifically includes: The basic architecture of the one-dimensional convolutional neural network is a two-channel input layer (one channel is time-domain waveform data, and the other channel is frequency-domain data) - convolutional layer 1 - pooling layer 1 - convolutional layer 2 - pooling layer 2 - convolutional layer 3 - fully connected layer - output layer. Considering the diversity of particle waveform data and the unevenness of the rising and falling edge features, different receptive fields (i.e., different types of convolutional kernels) are needed to improve the feature extraction ability. Therefore, three different forms of convolutional kernels are selected, namely 1*3 convolutional kernel, 1*5 dilated convolutional kernel (with an interval of 2), and 1*5 convolutional kernel, and three convolutional neural networks are trained respectively. To further improve the network discrimination accuracy, M groups are set, and each group inputs waveform data of different scales for parallel training, and finally the final particle discrimination result is obtained in the form of taking N from 3M. Through the software training of the one-dimensional convolutional neural network, the convolutional kernels and bias parameters of each layer can be obtained and used as the input parameters of the one-dimensional convolutional neural network in the FPGA platform architecture. In the actual application process of the FPGA platform, the waveform data of particle pulse signals is obtained in a high-speed acquisition manner and input into the normalization module and FFT module under the hardware platform for data preprocessing, and forward inference is realized by means of the one-dimensional convolutional neural network transplanted on the FPGA platform to obtain the particle type. While realizing forward inference, the FPGA platform stores the particle waveform data with labels and downloads it to realize the expansion and iteration of the training dataset, further improving the accuracy of the model parameters, and updating the FPGA platform parameters through data injection and other means to realize a complete closed-loop of on-orbit model update.

[0044] The key designs in this application include the following points:

[0045] 1. Use the feature data of the time-frequency domain two channels as training data, and expand the dataset through on-orbit downlink data and accelerator measured data to achieve the purpose of data augmentation.

[0046] Traditional particle waveform analysis methods usually use time-domain or frequency-domain data for feature extraction and discrimination, but the comprehensive discrimination method of time-domain + frequency-domain has not been used yet. Time-domain and frequency-domain information almost cover all the content of particle waveform information. With the powerful feature extraction ability of the convolutional neural network, the particle types can be more effectively discriminated. In addition, by summarizing and collecting the data transmitted from orbit and the measured data of the accelerator, the dataset of particle waveforms under various working conditions can be maximally expanded, enhancing the generalization characteristics of the entire network architecture.

[0047] 2. Use the CNN combined network to discriminate particle types.

[0048] In traditional methods, only single or partial information of particles is used to discriminate the types, and the waveform information cannot be fully utilized. Moreover, the convolutional neural network can automatically screen and locate various features in the waveform without the need for manual extraction of feature information, which is very convenient for data processing. In addition, to further improve the accuracy of particle discrimination, the form of algorithm integration and combined network is adopted. The same group of networks use different convolutional kernels for feature extraction respectively, and different groups of networks use different hyperparameters for model training. Finally, the final result is determined by voting. This method uses the dataset after data augmentation to train the combined architecture of the convolutional neural network for particle waveform recognition through different receptive fields (convolutional kernels) and different hyperparameters (mainly multi-scale waveform data), which can realize the integration of macroscopic large-scale feature discrimination and microscopic small-scale refined feature discrimination of particle waveforms, and has stronger discrimination ability than a single training network.

[0049] 3. Use FPGA to implement an on-orbit particle discrimination real-time processing platform.

[0050] Traditional methods often download the on-orbit waveform data or information such as energy and speed to the ground for processing. This processing method has two problems. One is that the real-time performance is poor and it cannot make timely and effective judgments. The other is that the amount of waveform data to be downloaded will be extremely large, which is not suitable for satellite resources, especially deep space exploration missions, where resources are extremely scarce. In addition, there are some on-orbit real-time processing methods and techniques, but they only utilize a small part of the waveform or particle information, such as the rising edge, falling edge, or zero-crossing time. The FPGA on-orbit real-time processing platform based on the one-dimensional convolutional neural network can effectively solve the problems of real-time performance, resource occupancy rate, and particle information utilization rate.

[0051] Example 2

[0052] Embodiment 2 of the present invention proposes a particle discrimination system based on a one-dimensional convolutional neural network. The required system includes a software preprocessing module for particle waveform data, a software training module for the one-dimensional convolutional neural network, an FPGA acquisition, storage, and preprocessing module for particle waveform data, an FPGA forward inference module for particle waveform data, and a training set expansion module for the one-dimensional convolutional neural network.

[0053] 1) Software preprocessing module for particle waveform data

[0054] The main function of this module is to process the waveform data of various particles obtained from on-orbit or ground accelerators in the time domain and frequency domain. First, the time-domain waveform is directly normalized. In the frequency-domain analysis, the amplitude-frequency data is obtained by Fourier transform. Considering the accuracy of the particle waveform and the convenience of the FFT operation, the number of waveform input data is 128, and the number of frequency-domain data is also 128. After conversion, the frequency-domain data is also normalized. At the same time, the data such as the rising edge, falling edge, and full waveform of the waveform are grouped. After the above operations, the respective time-domain and frequency-domain data are divided into training data and test data. The training data is mainly used to complete the training of the weights and biases of the entire network structure, and the test data is mainly used to complete the test and verification of the training results. The output format of the training and test data is excel, where the first N data are the model input data, and the (N + 1)-th data is the model label data used to match the classification results.

[0055] 2) Software training module for the one-dimensional convolutional neural network

[0056] This module mainly includes several parts, namely the CNN combination network module, the training module, the test verification module, and the weight bias export and fixed-point number conversion module.

[0057] The CNN combination network module includes 3M convolutional neural networks, where M can be any natural number. Then, for each group, 3 fixed convolutional kernels are selected, and the final classification result is determined by the voting method of 3M taking N. In addition, N should be at least greater than half of 3M for the voting to be effective. Every 3 convolutional neural networks form a group. In the same group, the convolutional kernel sizes of the convolutional neural networks are different, and the other hyperparameters of the convolutional neural networks in different groups are different except for the convolutional kernel size:

[0058] The CNN combined network module in this embodiment contains three groups of combined networks. Each combined network includes 3 convolutional neural networks. The basic composition of each network includes a model input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The input layer is dual-channel time-domain and frequency-domain normalized data. The convolutional layer gradually filters the data features of the input layer through multi-channel convolutional kernels, and performs dimensionality reduction and non-linearity on the output data through the pooling layer and relu calculation to enhance the non-linear fitting ability of the entire system. Then, the judgment result is output through the fully connected layer. Finally, the three groups of combined networks determine the final classification result through a 9-out-of-7 voting method.

[0059] When constructing a neural network, several key parameters need to be determined, namely the number of input data, the number of network layers, the number of convolutional kernel channels in each convolutional layer, the convolutional kernel size, the calculation step size, the number of padding data, and the pooling data size. To improve the flexibility of the software, other parameters except the number of network layers are used as variables for input, which is convenient for modification. This platform is constructed with three convolutional layers as the basic prototype, that is, the network architecture is input layer - convolutional layer 1 - pooling layer 1 - convolutional layer 2 - pooling layer 2 - convolutional layer 3 - fully connected layer - output layer (which can be adjusted according to the actual number of categories). To ensure that the dimensions of the input and output data are consistent, the three separate CNNs in each CNN combined network respectively use a convolutional kernel of 1*3 with a padding data of 1; a dilated convolutional kernel of 1*5 (interval 2) with a padding data of 2; a convolutional kernel of 1*5 with a padding data of 2. The step size of the above three networks is 1. Between the three groups of combined networks, multi-scale waveform data is used, that is, 128 data points of low-resolution full waveform, 128 data points of high-resolution rising edge, 128 data points of falling edge, and their corresponding frequency-domain data.

[0060] Training module

[0061] The training module includes forward propagation calculation, loss function calculation, and backpropagation gradient calculation.

[0062] Relying on the architecture of the convolutional neural network, forward propagation, loss function calculation, and backpropagation gradient calculation are respectively performed on the training data. Forward propagation is a prerequisite for training parameters, which can calculate the deviation between the classification result and the actual inference; the loss function calculation is to understand the deviation between the label data and the calculated data, and use this deviation to continuously iterate and update the weight parameters through backpropagation gradient calculation to achieve the final training purpose. During the actual operation process, first determine the learning rate lr, batch size batch_size, and the number of epochs Epoch for dataset training. The above parameters can be modified by manually changing variables. Subsequently, calculate the loss function through the CrossEntropyLoss() function, perform backpropagation gradient calculation through backward(), and continuously iterate the training parameters.

[0063] Test and verification module

[0064] The test and verification module is used to analyze the accuracy and ability of the model to achieve classification and identification in the current training state. By inputting the pre-processed dual-channel test data set, the accuracy of the inference result can be obtained, so as to understand the effectiveness of the model training and provide data support for determining the number of iterations. In the actual operation process, the key variables include training accuracy, loss function value, and test accuracy. The calculation method of the training accuracy train_corrects_output is the number of train_corrects where the forward inference result of the training data matches the label divided by the total number of training data train_num, that is

[0065] train_corrects_output = train_corrects / train_num

[0066] The calculation result of the loss function value average_loss is the sum of the accumulated loss function train_loss divided by the total number of training data train_num, that is

[0067] average_loss = train_loss / train_num

[0068] The calculation result of the test accuracy test_corrects_output is the number of test_corrects where the forward inference result of the test data matches the label divided by the total number of test data test_num, that is

[0069] test_corrects_output = test_corrects / test_num

[0070] In addition to calculating the intermediate variables, in order to more intuitively understand the parameter change trend of the entire training process, the training accuracy, loss function value, and test accuracy are plotted as curves, and each point on the curve is the result of one Epoch calculation.

[0071] Weight Bias Export and Fixed-Point Conversion Module

[0072] The weight bias export and fixed-point conversion module outputs the weights and biases in fixed-point form after training and writes them into the ROM in the FPGA in the form of a COE file. In actual operation, the numerictype() function of matlab is used to implement the conversion of all weights, biases, and input data, and they are arranged and output in sequence according to the storage architecture.

[0073] 3) Particle Waveform Data FPGA Acquisition, Storage and Preprocessing Module

[0074] The FPGA high-speed acquisition module is responsible for acquiring the particle pulse waveform signals amplified by the preamplifier, with an acquisition rate of 200M. The storage module is responsible for storing 128 complete data points with different resolutions and different waveform positions. During the storage process, cumulative calculations are performed. After completion, the 128 data points are normalized. At the same time, FFT processing and normalization operations are performed on the waveform data to realize the data preparation for the dual-channel input layer of the convolutional neural network.

[0075] 4) FPGA Forward Inference Module for Particle Waveform Data

[0076] A key link in building a convolutional neural network with FPGA is to implement multi-channel operations in the convolutional layer through a parallel computing structure. The pseudo-code diagram of the convolutional calculation is as follows:

[0077]

[0078] Among them, Loop-1 refers to the calculation of each element within the convolutional kernel, Loop-2 refers to the calculation of the convolutional kernels corresponding to different input channels of a single filter, Loop-3 refers to the sliding calculation of a single filter on one-dimensional data, and Loop-4 refers to the calculation of multiple filters. The current parallel computing strategy is full parallelism between convolutional kernels, full parallelism of input channels, parallel operation of a single convolutional kernel, and serial operation of the sliding window.

[0079] According to the above parallel strategy, the systolic array is the best computing means. Taking the single-channel input with a 5-element convolutional kernel and 5 input data as an example, the systolic array calculation process is constructed as Figure 2 shown. The following five rectangles are computing units PE, including five convolutional kernel elements W0 - W4. Three input data are input sequentially from left to right according to the clock beats. The calculation process is shown in Table 1, and finally the result of the convolutional calculation can be obtained.

[0080] Table 1

[0081]

[0082]

[0083]

[0084] Convolution calculation is a major consumer of chip resources and time resources. Quantitative analysis of computing resources and understanding time delay are one of the key steps in building the platform. Since convolution operations are mainly multiplication and addition operations and have high parallel requirements, dedicated DSP modules within the FPGA are used for calculation when constructing the arithmetic unit (PE unit). To improve the inference speed, the cache module needs to use the internal RAM of the FPGA. According to the actual architecture of the convolutional neural network, taking a 5-element convolution kernel and 128 input data as an example, the occupancy of the DSP module in the actual construction process is shown in Table 2, and the time resources are shown in Table 3.

[0085] Table 2

[0086]

[0087] Table 3

[0088]

[0089]

[0090] Through the above steps, the calculated values of different output nodes of different models of the combined network can be obtained. By simple size comparison and voting analysis, the forward inference process of the particles can be realized, and finally the particle types can be determined.

[0091] 5) One-dimensional convolutional neural network training set expansion module

[0092] With the accumulation of on-orbit and accelerator calibration data, the training sample set can be continuously expanded and re-learned to improve the accuracy and precision of discrimination. The waveform data obtained on orbit can be regularly downloaded to the ground as a means of expanding the sample set, re-training the model and updating the parameters. Finally, the parameters of each layer of the FPGA platform are updated by data injection to achieve a closed-loop of space-ground model update.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the embodiments, those of ordinary skill in the art should understand that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A particle discrimination method based on a one-dimensional convolutional neural network, the method comprising: Preprocessing the waveform data generated by various particles incident on a detector in a collected space to obtain the time-domain waveform data and frequency-domain waveform data of various particles; Simultaneously inputting the time-domain waveform data and frequency-domain waveform data of various particles into a pre-established and trained particle discrimination model to obtain a particle discrimination result; the particle discrimination model is established based on a one-dimensional convolutional neural network; Wherein, the one-dimensional convolutional neural network includes M groups, and each group inputs waveform data of different scales for parallel training, and the waveform data of different scales includes full waveform data, rising edge data, falling edge data, and their corresponding frequency-domain data.

2. The particle discrimination method based on a one-dimensional convolutional neural network according to claim 1, wherein The various particles include: protons, electrons, heavy particles, neutrons, and gamma rays.

3. The particle discrimination method based on a one-dimensional convolutional neural network according to claim 1, characterized in that, The preprocessing includes: time-domain data normalization and frequency-domain conversion and normalization.

4. The particle discrimination method based on a one-dimensional convolutional neural network according to claim 1, wherein The particle discrimination model includes 3M parallel convolutional neural networks, M is a natural number, and each convolutional neural network includes: an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer; every 3 convolutional neural networks are in a group, and the convolutional kernel sizes of the convolutional neural networks in the same group are different, and other parameters between different groups of convolutional neural networks are different, and the other parameters include: multi-scale data sets, weight initialization, activation functions, the number of network layers, the number of iterative training times, the learning rate, and the batch size; Wherein, the input layer includes in parallel: a time-domain waveform data input channel and a frequency-domain waveform data input channel; the convolutional layer gradually screens the features of the input layer data through a multi-channel convolutional kernel, and performs dimensionality reduction and non-linearity on the output data through the pooling layer and relu calculation, and finally outputs a judgment result through the fully connected layer; The determination result is the final classification result determined by the voting method of taking N out of 3M.

5. The particle discrimination method based on a one-dimensional convolutional neural network according to claim 1, characterized in that The method further includes: training the particle discrimination model; the training process includes: Preprocessing the waveform data of various particles obtained from calibration tests and on-orbit detections to obtain time-domain data and frequency-domain data; Constructing a training data set and a test data set based on the obtained time-domain data and frequency-domain data; Inputting the time-domain data and frequency-domain data of the training data set into the particle discrimination model for training; And inputting the time-domain data and frequency-domain data of the test data set into the particle discrimination model for testing to determine the correctness of each layer of convolutional kernel and bias parameters, and obtaining a trained particle discrimination model.

6. The particle discrimination method based on a one-dimensional convolutional neural network according to claim 5, characterized in that, During the training process, the loss function is calculated through the Cross Entropy Loss(·) function, the backward gradient calculation is performed through backward(·), and the training parameters are continuously iterated.

7. A particle discrimination system based on a one-dimensional convolutional neural network, characterized in that, The system includes: a data preprocessing module and a particle discrimination module; wherein, The data preprocessing module is used for preprocessing the waveform data generated by various particles incident on a detector in a space to obtain time-domain waveform normalized data and frequency-domain waveform normalized data; The particle discrimination module is used to input the time-domain waveform data and the frequency-domain waveform data into a pre-established and trained particle discrimination model simultaneously to obtain a particle discrimination result; the particle discrimination model is established based on a one-dimensional convolutional neural network; wherein, the one-dimensional convolutional neural network includes M groups, and each group inputs waveform data of different scales for parallel training, and the waveform data of different scales includes full waveform data, rising edge data, falling edge data and their corresponding frequency-domain data.

8. The particle discrimination system based on a one-dimensional convolutional neural network according to claim 7, wherein The system is implemented based on FPGA.

9. The particle discrimination system based on a one-dimensional convolutional neural network according to claim 8, wherein The system further includes a data acquisition module, which is used to acquire waveform data of various particles during on-orbit detection or through a ground accelerator and transmit it to the data preprocessing module.

Images