A particle and event type identification system using a neural network processor

Abstract

The present invention discloses a particle and event type identification system using a neural network processor, which relates to the field of radiation detection technology. A nuclear pulse signal acquisition module collects the ray energy of a target to be identified, converts it into a nuclear pulse digital waveform and sends it to a CPU processing unit to extract nuclear pulse signal data; the NPU processing unit determines the first type of the target to be identified based on the nuclear pulse signal data using a first identification model deployed in the NPU processing unit; the first type is alpha particles or beta particles; or based on the nuclear pulse signal data, using a second identification model deployed in the NPU processing unit, determines the second type of the target to be identified; the second type is a full energy deposition event or a partial energy deposition event; or based on the nuclear pulse signal data, using a third identification model deployed in the NPU processing unit, determines the third type of the target to be identified; the third type is a neutron or a gamma ray. The present invention realizes online identification of particle and event types.

Description

Technical Field

[0001] The present invention relates to the field of radiation detection technology, and in particular to a particle and event type identification system using a neural network processor. Background Art

[0002] In nuclear physics research and radiation detection, it is of great significance to identify the particle and event types of nuclear reactions. For example, in neutron detection experiments, neutrons and gamma rays are identified, α and β particles are identified in liquid scintillation spectrometers, single-point events corresponding to signal events and multi-point events corresponding to background events are identified in neutrinoless double beta decay experiments, and full energy deposition events in the full energy peak range and partial energy deposition events in the Compton plateau region are identified in gamma ray detection. For this reason, it is necessary to develop a high-performance, general-purpose method that can be used for different application scenarios and can identify particle and event types online. Summary of the invention

[0003] The purpose of the present invention is to provide a particle and event type identification system using a neural network processor, which realizes the online identification of particle and event types.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A particle and event type identification system using a neural network processor includes: a nuclear pulse signal acquisition module and a SOC processor, wherein the SOC processor is equipped with a CPU processing unit and an NPU processing unit; the nuclear pulse signal acquisition module, the CPU processing unit and the NPU processing unit are connected in sequence;

[0006] The nuclear pulse signal acquisition module is used to collect the ray energy of the target to be identified, and convert the ray energy into a nuclear pulse digital waveform, and send the nuclear pulse digital waveform to the CPU processing unit;

[0007] The CPU processing unit is used to extract nuclear pulse signal data from the nuclear pulse digital waveform;

[0008] The NPU processing unit is used for:

[0009] Based on the nuclear pulse signal data, a first type of the target to be identified is determined using a first identification model deployed in the NPU processing unit; the first identification model is obtained by training a neural network using a plurality of nuclear pulse signal data and a corresponding first type, the first type is an alpha particle or a beta particle, and the neural network is a convolutional neural network or a fully connected neural network;

[0010] or based on the nuclear pulse signal data, using a second discrimination model deployed in the NPU processing unit to determine a second type of the target to be discriminated; the second discrimination model is obtained by training a neural network using a plurality of nuclear pulse signal data and corresponding second types, and the second type is a full energy deposition case or a partial energy deposition case;

[0011] Or based on the nuclear pulse signal data, a third identification model deployed in the NPU processing unit is used to determine the third type of the target to be identified; the third identification model is obtained by training a neural network using multiple nuclear pulse signal data and corresponding third types, and the third type is neutrons or gamma rays.

[0012] Optionally, the nuclear pulse signal acquisition module comprises: a radiation detector, an input buffer circuit, a low-noise signal amplification circuit, a single-ended to differential circuit and a high-speed ADC circuit connected in sequence; the high-speed ADC circuit is connected to the SOC processor;

[0013] The radiation detector is used to collect the radiation energy of the target to be identified and convert the radiation energy into an original voltage pulse signal;

[0014] The input buffer circuit is used to amplify the original voltage pulse signal to obtain a voltage pulse signal after primary amplification;

[0015] The low-noise signal amplifier circuit is used to further amplify and reduce noise on the voltage pulse signal after the primary amplification, and obtain the voltage pulse signal after the secondary amplification;

[0016] The single-ended to differential circuit is used to convert the voltage pulse signal after secondary amplification into a differential signal;

[0017] The high-speed ADC circuit is used to perform analog-to-digital conversion on the differential signal to obtain the nuclear pulse digital waveform.

[0018] Optionally, the identification system further includes: a storage module; the storage module is connected to the CPU processing unit;

[0019] The storage module is used to store the nuclear pulse signal data extracted by the CPU processing unit;

[0020] The CPU processing unit is also used to read the nuclear pulse signal data from the storage module when the NPU processing unit is idle, and send the nuclear pulse signal data to the NPU processing unit for identification.

[0021] Optionally, an artificial neural network autoencoder is further deployed in the NPU processing unit; the artificial neural network autoencoder is respectively connected to the CPU processing unit, the first discrimination model, the second discrimination model and the third discrimination model;

[0022] The artificial neural network autoencoder is used to filter and reduce noise on the nuclear pulse signal data, and send the filtered and denoised nuclear pulse signal data to the first discrimination model, the second discrimination model or the third discrimination model for discrimination.

[0023] Optionally, the SOC processor is further provided with a particle event data packet; the particle event data packet is connected to the CPU processing unit;

[0024] The CPU processing unit is further used to send the identification result output by the NPU processing unit to the particle event data packet for temporary storage; the identification result is the first type, the second type or the third type.

[0025] Optionally, the identification system further includes: a host computer; the host computer is connected to the particle event data packet.

[0026] Optionally, the host computer is connected to the particle event data packet via USB 3.0.

[0027] Optionally, the host computer is connected to the particle event data packet via a router and an Ethernet interface in sequence.

[0028] Optionally, the storage module is a DDR memory or an SDRAM memory.

[0029] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0030] The invention discloses a particle and event type identification system using a neural network processor, comprising: a nuclear pulse signal acquisition module and a SOC processor, wherein the SOC processor is equipped with a CPU processing unit and an NPU processing unit; the nuclear pulse signal acquisition module, the CPU processing unit and the NPU processing unit are connected in sequence; the nuclear pulse signal acquisition module is used to acquire the ray energy of a target to be identified, and convert the ray energy into a nuclear pulse digital waveform, and send the nuclear pulse digital waveform to the CPU processing unit; the CPU processing unit is used to extract nuclear pulse signal data from the nuclear pulse digital waveform; the NPU processing unit is used to: based on the nuclear pulse signal data, using a first identification model deployed in the NPU processing unit, determine a first type of the target to be identified; the first type is alpha particles or beta particles; or based on the nuclear pulse signal data, using a second identification model deployed in the NPU processing unit, determine a second type of the target to be identified; the second type is a full energy deposition event or a partial energy deposition event; or based on the nuclear pulse signal data, using a third identification model deployed in the NPU processing unit, determine a third type of the target to be identified; wherein the first identification model, the second identification model and the third identification model are all obtained based on a neural network. The present invention realizes the online identification of particle and event types by adopting a particle and event type identification system of a neural network processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. 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 creative work.

[0032] Figure 1 A schematic diagram of the structure of a particle and event type identification system using a neural network processor provided in an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram for distinguishing alpha particles from beta particles;

[0034] Figure 3 It is a schematic diagram for distinguishing between full energy deposition events and partial energy deposition events;

[0035] Figure 4 Schematic diagram of neutron and gamma ray discrimination;

[0036] Figure 5 Schematic diagram of the convolutional neural network structure;

[0037] Figure 6 Schematic diagram of the channel address interval corresponding to the full / partial energy deposition case in the energy spectrum;

[0038] Figure 7 Schematic diagram of pulse width-normalized pulse amplitude when distinguishing neutrons and gamma rays by charge comparison method;

[0039] Figure 8 Schematic diagram of total charge-count when distinguishing neutrons from gamma rays using the charge comparison method. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments 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 creative work are within the scope of protection of the present invention.

[0041] The purpose of the present invention is to provide a particle and event type identification system using a neural network processor, aiming to achieve online identification of particle and event types.

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Figure 1 The schematic diagram of the structure of the particle and event type identification system using a neural network processor provided by the embodiment of the present invention. Figure 1 As shown, the particle and event type identification system using a neural network processor in this embodiment includes: a nuclear pulse signal acquisition module and a SOC processor, the SOC processor is deployed with a CPU processing unit and an NPU processing unit; the nuclear pulse signal acquisition module, the CPU processing unit and the NPU processing unit are connected in sequence.

[0044] The nuclear pulse signal acquisition module is used to collect the radiation energy of the target to be identified, convert the radiation energy into a nuclear pulse digital waveform, and send the nuclear pulse digital waveform to the CPU processing unit.

[0045] The CPU processing unit is used for extracting nuclear pulse signal data from the nuclear pulse digital waveform.

[0046] The NPU processing unit is used to:

[0047] Based on the nuclear pulse signal data, the first discrimination model (corresponding to Figure 2 The first identification model is obtained by training a neural network using a plurality of nuclear pulse signal data and a corresponding first type, the first type being an alpha particle or a beta particle, and the neural network is a convolutional neural network or a fully connected neural network.

[0048] Or based on the nuclear pulse signal data, using the second discrimination model deployed in the NPU processing unit (corresponding to Figure 3 The artificial neural network model in the embodiment of the present invention) determines the second type of the target to be identified; the second identification model is obtained by training the neural network using multiple nuclear pulse signal data and the corresponding second type, and the second type is a full energy deposition case or a partial energy deposition case.

[0049] Or based on the nuclear pulse signal data, using the third discrimination model deployed in the NPU processing unit (corresponding to Figure 4 The artificial neural network model in the method) determines the third type of the target to be identified; the third identification model is obtained by training the neural network using multiple nuclear pulse signal data and the corresponding third type, and the third type is neutrons or gamma rays.

[0050] Specifically, Figure 1 The artificial neural network model (case identification) corresponds to the first identification model, the second identification model and the third identification model.

[0051] As an optional implementation, the nuclear pulse signal acquisition module includes: a radiation detector, an input buffer circuit, a low-noise signal amplification circuit, a single-ended to differential circuit and a high-speed ADC circuit connected in sequence; the high-speed ADC circuit is connected to the SOC processor.

[0052] The radiation detector is used to collect the radiation energy of the target to be identified and convert the radiation energy into an original voltage pulse signal.

[0053] Specifically, the radiation interacts with the radiation detector to deposit energy in the radiation detector, which is used to convert the energy of the radiation into an original voltage pulse signal so that the pulse signal can be processed by the back-end electronic circuit to extract effective information.

[0054] The input buffer circuit is used to amplify the original voltage pulse signal to obtain an amplified voltage pulse signal.

[0055] Specifically, in radiation detection, the role of the input buffer circuit is to amplify the original voltage pulse signal output by the radiation detector and match it to the input impedance of the subsequent circuit for further processing and analysis. Radiation detectors usually generate some weak electrical signals as a response to radiation. These electrical signals may be very small and need to be amplified for subsequent processing. The input buffer circuit increases the amplitude of the signal and improves the signal-to-noise ratio so that the signal can be reliably processed in the subsequent circuit. In addition, the input buffer circuit can also provide a suitable input impedance to make the connection between the radiation detector and the subsequent circuit more stable. It can reduce the transmission distortion between the radiation detector and the subsequent circuit and reduce the load impact on the radiation detector.

[0056] The low-noise signal amplifier circuit is used to further amplify and reduce the noise of the voltage pulse signal after the primary amplification, and obtain the voltage pulse signal after the secondary amplification.

[0057] Specifically, the function of the low-noise signal amplifier circuit is to amplify the received weak radiation signal (i.e., the voltage pulse signal after amplification) while minimizing the additional circuit noise. The radiation signal received by the radiation detector is usually very weak, so it needs to be amplified before it can be effectively processed and analyzed by the subsequent circuit. The low-noise signal amplifier circuit can minimize the additional noise introduced while amplifying the signal. This ensures that the amplified signal maintains the original signal quality as much as possible, while providing a sufficient signal-to-noise ratio to make subsequent processing more reliable.

[0058] The single-ended to differential circuit is used to convert the voltage pulse signal after secondary amplification into a differential signal.

[0059] Specifically, the function of the single-ended to differential circuit is to convert a single-ended signal (i.e., a voltage pulse signal after secondary amplification) into a differential signal. This conversion circuit plays an important role in radiation detection systems. Radiation detectors usually output a single-ended signal, which corresponds to the characteristics of the radiation source. However, single-ended signals are susceptible to external interference such as power supply noise, electromagnetic interference, etc., which may reduce the quality and reliability of the signal, affect the accuracy of extracting useful information from a single signal and identifying the detected radiation event. By converting a single-ended signal into a differential signal through a single-ended to differential circuit, common-mode noise and interference can be suppressed, thereby improving the system's anti-interference ability and improving the signal-to-noise ratio and reliability of the signal. In addition, the radiation detector may face radiation signals with a high dynamic range, and the single-ended to differential circuit can also increase the dynamic range. By converting to a differential signal, the dynamic range can be expanded and signal distortion can be reduced, thereby better capturing and analyzing radiation events.

[0060] The high-speed ADC circuit is used to perform analog-to-digital conversion on the differential signal to obtain a nuclear pulse digital waveform.

[0061] Specifically, the high-speed ADC circuit performs analog-to-digital conversion on the pulse signal, converts the analog signal into a digital signal, and inputs it to the SOC processor for operation and analysis. In order to obtain more accurate analysis results in the operation and analysis of the SOC processor, the sampling rate and resolution of the ADC should be as high as possible to obtain sufficient sampling points and sampling accuracy.

[0062] As an optional implementation, the identification system further includes: a storage module; the storage module is connected to the CPU processing unit.

[0063] The storage module is used to store the nuclear pulse signal data extracted by the CPU processing unit.

[0064] The CPU processing unit is also used to read the nuclear pulse signal data from the storage module when the NPU processing unit is idle, and send the nuclear pulse signal data to the NPU processing unit for identification.

[0065] As an optional implementation, an artificial neural network autoencoder is also deployed in the NPU processing unit; the artificial neural network autoencoder is connected to the CPU processing unit, the first discrimination model, the second discrimination model and the third discrimination model respectively.

[0066] The artificial neural network autoencoder is used to filter and reduce noise on the nuclear pulse signal data, and send the filtered and denoised nuclear pulse signal data to the first discrimination model, the second discrimination model or the third discrimination model for discrimination.

[0067] Specifically, the artificial neural network autoencoder is an unsupervised neural network model. During the training process, nuclear pulse waveform data superimposed with various noises are input into the input layer. After the hidden layer is operated, the output layer outputs a string of new waveform data. The newly output waveform data is compared with the ideal waveform without superimposed noise. If the difference is large, the weight parameters of the hidden layer are corrected and another round of training is performed. The training is stopped until the waveform data superimposed with noise input in the input layer is close to the ideal waveform data without superimposed noise after the model operation. At this time, the trained neural network autoencoder has the ability to suppress the noise superimposed on the signal while retaining the characteristics of the original nuclear pulse signal. This model is deployed in the NPU processing unit, and the NPU processing unit can use the efficient computing power of the neural network model to realize the function of filtering and denoising the nuclear pulse signal.

[0068] As an optional implementation, the SOC processor is further equipped with a particle event data packet; the particle event data packet is connected to the CPU processing unit.

[0069] The CPU processing unit is further used to send the identification result output by the NPU processing unit to the particle event data packet for temporary storage; the identification result is the first type, the second type or the third type.

[0070] SOC processor: The ARM architecture CPU processing unit in the SOC processor mainly controls data transmission and program operation, while the NPU processing unit mainly deploys artificial neural network models for more complex data operations and reasoning. The SoC processor is also equipped with particle event data packet FIFO, USB3.0 and Ethernet communication modules. The CPU processing unit controls data transmission, data processing and program operation. Two artificial neural network models are deployed in the NPU processing unit, namely the neural network autoencoder for noise reduction filtering and the neural network model for identifying event types. After the digital nuclear pulse signal converted by the high-speed ADC circuit is transmitted to the SOC processor through parallel communication, the CPU processing unit processes the digital nuclear pulse signal, extracts the fixed-length single nuclear pulse signal data, and then caches it in DDR or SDRAM. When the NPU processing unit is in an idle state, under the control of the CPU processing unit, the cached digital nuclear pulse signal is read from DDR or SDRAM and transmitted to the neural network model input structure variable of the NPU processing unit. After the digital nuclear pulse signal is operated by the neural network autoencoder, the noise in the signal is suppressed, the signal-to-noise ratio is improved, and the original characteristics of the pulse waveform are retained. Then the signal is transmitted to the next-level artificial neural network model for identifying the type of case. The result of the case type identification is saved in the output structure variable of the neural network model, and is saved in the particle case data packet FIFO under the control of the CPU processing unit. The data in the FIFO is then transmitted to the host computer on the PC side through USB3.0 or Ethernet communication.

[0071] As an optional implementation, the identification system further includes: a host computer; the host computer is connected to the particle event data packet.

[0072] As an optional implementation, the host computer is connected to the particle event data packet via USB3.0.

[0073] As an optional implementation, the host computer is connected to the particle event data packet via a router and an Ethernet interface in sequence.

[0074] As an optional implementation, the storage module is a DDR memory or an SDRAM memory.

[0075] Specifically, DDR and SDRAM are both synchronous dynamic random access memories used to store programs and data, providing computer operation and data processing capabilities. In the present invention, the storage module is mainly used to cache the fixed-length digital nuclear pulse signal data extracted by the CPU processing unit from the original waveform data transmitted by the high-speed ADC. The purpose of the cache is to solve the problem that when the intensity of the radiation source is large and the counting rate is too high, the NPU processing unit generates another nuclear pulse signal at the same time during the dead time generated by processing a nuclear pulse signal, which cannot be processed by the NPU in time, resulting in the loss of the nuclear pulse signal and the reduction of the pass rate.

[0076] Specifically, in Figure 1 In the system shown, the radiation detector measures the radiation energy of the target to be identified and converts the radiation energy into an original voltage pulse signal. The original voltage pulse signal is processed by the analog front-end circuit and then converted into a digital signal by the high-speed ADC and transmitted to the SOC processor. The CPU processing unit in the SOC processor extracts the nuclear pulse signal data from the original waveform data and caches it in the DDR or SDRAM memory. When the working state of the embedded neural network processor (NPU) processing unit is idle, the CPU processing unit controls the reading of the nuclear pulse signal data from the DDR / SDRAM and transmits it to the NPU processing unit. The artificial neural network autoencoder deployed in the NPU performs filtering and noise reduction, and the trained artificial neural network model is used to identify particles and cases. Finally, the identified case type result is returned to the CPU processing unit. The CPU processing unit buffers the identified case type data in the form of a particle case data packet into the FIFO, and then transmits the particle identification data packet to the host computer through the USB3.0 or Gigabit network interface for further analysis. The artificial neural network model used to identify the particle and case type in the NPU can be changed according to the needs of the application, so the nuclear pulse signal processor based on this hardware architecture has strong functional expansibility and versatility.

[0077] In this system, since the neural network model needs to perform case identification based on the discrete data of the nuclear pulse signal, it is necessary to collect as many sampling points of the nuclear pulse signal as possible to fully obtain the shape characteristics of the nuclear pulse signal, so a high-speed and high-precision analog-to-digital converter (ADC) is required.

[0078] Different types of particles and events have different physical processes of interaction with radiation detectors. The differences in the physical processes of interaction are reflected in the shape of the nuclear pulse signal. Therefore, the type of particles and events interacting with the detector can be inferred from the shape characteristics of the nuclear pulse signal. The present invention introduces three methods for particle and event identification based on artificial neural network technology used in the field of nuclear physics. The model in the present invention is deployed on the NPU hardware platform to achieve online identification of particle and event types. The application scope of the present invention includes but is not limited to the three situations described below. For other applications involving particle and event type identification, after obtaining waveform data and marking label values, the present invention can also be used to achieve online identification of particle and event types. In order to train the various identification models in the present invention, the technical contents involved are as follows:

[0079] 1. Build an artificial neural network model.

[0080] Artificial neural network technology can learn the shape features of nuclear pulse signals more precisely. Applying it in the application of particle and event type identification can greatly improve the accuracy of identification. The discrete data of nuclear pulse signals can be regarded as sequence data, so the present invention designs the following Figure 5 The neural network (ANN) shown is used as a network model for identifying particle and event types. The neural network can perform operations such as convolution on the sequence data corresponding to the nuclear pulse signal data, learn the waveform characteristics of the nuclear pulse signal during the training process, and the trained network model can be used to identify the particle and event type corresponding to the nuclear pulse signal.

[0081] The convolution layer is used to extract features from the waveform of the nuclear pulse signal; batch normalization is used to normalize the nuclear pulse waveform data so that the waveform data is at the same order of magnitude and eliminate the impact of the difference in pulse waveform amplitude on model discrimination; the ReLU activation function is used to increase the nonlinear relationship between the layers of the neural network, while making the activation of the neural network sparse, and can solve the gradient vanishing problem and accelerate the training of the neural network; random inactivation is to randomly return some weights or outputs of the hidden layer to zero, reduce the interdependence between nodes, thereby achieving the regularization of the neural network and preventing overfitting, while reducing the network parameters involved in the calculation and reducing the model capacity; the fully connected layer is used to integrate the calculation results of the previous hidden layer to achieve the final classification task; the softmax activation function is used to convert the output of the network into a probability distribution to obtain the predicted probability of each class.

[0082] 2. Preparation of data set.

[0083] (a) α and β particles.

[0084] The method for distinguishing α and β particles is based on a physical mechanism: α particles are heavy charged particles. Because of their large weight, the interaction process with matter is relatively slow, and they usually lose energy over a long period of time, and their energy deposition location is more concentrated. β particles are light charged particles. Compared with α particles, their energy deposition location is more dispersed, and the duration of energy loss is relatively short. Therefore, the amplitude, duration and shape of the nuclear pulse signal generated by the interaction between α and β particles and the detector are somewhat different. The detector is bombarded with α and β particles respectively, and the nuclear pulse signal data is obtained and marked according to the type of incident particles or rays to complete the preparation of the data set.

[0085] (b) Full / partial energy deposition example.

[0086] The method for distinguishing full energy deposition events from partial energy deposition events is based on a physical mechanism: the full energy deposition event deposits all the gamma ray energy in the detector, and the energy corresponding to the pulse amplitude generated by it is in the full energy peak interval in the energy spectrum, while the partial energy deposition event only deposits part of the gamma ray energy in the detector, and the energy corresponding to the pulse amplitude generated by it is in the Compton plateau interval in the energy spectrum. In the process of making a data set, the event type corresponding to the nuclear pulse signal is judged based on whether the nuclear pulse signal amplitude is in the full energy peak interval or the Compton plateau area, and then the group of waveform data is labeled, such as Figure 6 The waveform data and label values ​​of the pulses constitute the data set of the artificial neural network model used to distinguish full / partial energy deposition events.

[0087] (C) Neutrons and gamma rays.

[0088] When gamma rays interact with scintillator materials, photons are generated, forming fast light scintillation signals. The interaction process of neutrons is relatively complex, and secondary particles such as protons and alpha particles may be produced. The energy release and darkening process of these secondary particles through the interaction with the scintillator is much slower than the light scintillation process of gamma rays. Therefore, the scintillation light signal of neutrons lasts longer than that of gamma rays. The role of gamma rays in scintillators is mainly to produce ionized charges through the ionization effect, while the signal caused by the interaction of neutrons is mainly the light signal generated by the scintillation process. Therefore, the signal shape of neutrons is closer to a long-duration attenuation signal, while the signal shape of gamma rays is usually a spike-like pulse signal. Judgment and identification can be made by comparing the duration of the signal, the shape of the signal, the rise time of the signal and other characteristics. In the process of making a data set, the charge comparison method is first used to identify neutrons and gamma rays, such as Figure 7 and Figure 8 As shown, the nuclear pulse signal is labeled according to the identification result, and the waveform data and label value of the nuclear pulse signal constitute the data set of the artificial neural network model for identifying neutron gamma.

[0089] 3. Train the artificial neural network model.

[0090] The data set is divided into training set, validation set and test set for training neural network model. The model training adopts SGDM deep learning optimizer. The number of small batch data randomly sampled each time is 1024, the maximum number of iterations is 20, and the initial learning rate is 0.001. After each round of training, the data set of the training model is shuffled to prevent model jitter and overfitting during training and improve the quality and prediction performance of the model. The artificial neural network model continuously optimizes its own weights and other parameters during the training process. When the model is based on the input nuclear pulse signal data, after the hidden layer is calculated, the accuracy of the particle and case type reasoning results output in the output layer reaches the preset accuracy or the accuracy cannot be improved for a long period of training time, the training is stopped. At this time, the nuclear pulse signal data is input to the input layer of the trained neural network model, and the output layer of the model outputs the probability value of each particle and case type. The classification with the largest probability value is the prediction result of the model, thereby realizing the identification of particle and case types.

[0091] 4. Model deployment and online analysis.

[0092] Model deployment is to convert the trained artificial neural network model into a model file format that can be recognized and loaded by NPU through a dedicated tool chain after model optimization, model conversion and model quantization. When the ARM architecture CPU runs the program, the model file is loaded to configure the NPU processing engine. After the configuration is completed, when performing data calculations, the NPU processing engine calculates and analyzes the real-time input nuclear pulse signal data according to the structure and operators of the previously trained neural network model, so as to infer the probability value of each particle and event type classification, and take the classification type with the largest inferred probability value as the prediction result, thereby realizing real-time analysis of particle and event types. The nuclear pulse signal processor collects pulse waveform data in real time and analyzes it through the NPU, outputs the probability value of each event type, and finds the corresponding event type according to the number with the largest probability value. The schematic diagram of the whole process is shown below. Figure 2-Figure 4 shown.

[0093] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0094] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the system and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A particle and event type identification system using a neural network processor, characterized in that: The identification system comprises: a nuclear pulse signal acquisition module and a SOC processor, wherein the SOC processor is equipped with a CPU processing unit and an NPU processing unit; the nuclear pulse signal acquisition module, the CPU processing unit and the NPU processing unit are connected in sequence; The nuclear pulse signal acquisition module is used to collect the ray energy of the target to be identified, and convert the ray energy into a nuclear pulse digital waveform, and send the nuclear pulse digital waveform to the CPU processing unit; The CPU processing unit is used to extract nuclear pulse signal data from the nuclear pulse digital waveform; The NPU processing unit is used for: Based on the nuclear pulse signal data, a first type of the target to be identified is determined using a first identification model deployed in the NPU processing unit; the first identification model is obtained by training a neural network using a plurality of nuclear pulse signal data and a corresponding first type, the first type is an alpha particle or a beta particle, and the neural network is a convolutional neural network or a fully connected neural network; or based on the nuclear pulse signal data, using a second discrimination model deployed in the NPU processing unit to determine a second type of the target to be discriminated; the second discrimination model is obtained by training a neural network using a plurality of nuclear pulse signal data and corresponding second types, and the second type is a full energy deposition case or a partial energy deposition case; Or based on the nuclear pulse signal data, a third identification model deployed in the NPU processing unit is used to determine the third type of the target to be identified; the third identification model is obtained by training a neural network using multiple nuclear pulse signal data and corresponding third types, and the third type is neutrons or gamma rays.

2. The particle and event type identification system using a neural network processor according to claim 1, characterized in that: The nuclear pulse signal acquisition module comprises: a radiation detector, an input buffer circuit, a low-noise signal amplification circuit, a single-ended to differential circuit and a high-speed ADC circuit connected in sequence; the high-speed ADC circuit is connected to the SOC processor; The radiation detector is used to collect the radiation energy of the target to be identified and convert the radiation energy into an original voltage pulse signal; The input buffer circuit is used to amplify the original voltage pulse signal to obtain a voltage pulse signal after primary amplification; The low-noise signal amplifier circuit is used to further amplify and reduce noise on the voltage pulse signal after the primary amplification, and obtain the voltage pulse signal after the secondary amplification; The single-ended to differential circuit is used to convert the voltage pulse signal after secondary amplification into a differential signal; The high-speed ADC circuit is used to perform analog-to-digital conversion on the differential signal to obtain the nuclear pulse digital waveform.

3. The particle and event type identification system using a neural network processor according to claim 1, characterized in that: The identification system further includes: a storage module; the storage module is connected to the CPU processing unit; The storage module is used to store the nuclear pulse signal data extracted by the CPU processing unit; The CPU processing unit is also used to read the nuclear pulse signal data from the storage module when the NPU processing unit is idle, and send the nuclear pulse signal data to the NPU processing unit for identification.

4. The particle and event type identification system using a neural network processor according to claim 1, characterized in that: An artificial neural network autoencoder is also deployed in the NPU processing unit; the artificial neural network autoencoder is respectively connected to the CPU processing unit, the first discrimination model, the second discrimination model and the third discrimination model; The artificial neural network autoencoder is used to filter and reduce noise on the nuclear pulse signal data, and send the filtered and denoised nuclear pulse signal data to the first discrimination model, the second discrimination model or the third discrimination model for discrimination.

5. The particle and event type identification system using a neural network processor according to claim 1, characterized in that: The SOC processor is also equipped with a particle event data packet; the particle event data packet is connected to the CPU processing unit; The CPU processing unit is further used to send the identification result output by the NPU processing unit to the particle event data packet for temporary storage; the identification result is the first type, the second type or the third type.

6. The particle and event type identification system using a neural network processor according to claim 5, characterized in that: The identification system also includes: a host computer; the host computer is connected to the particle event data packet.

7. The particle and event type identification system using a neural network processor according to claim 6, characterized in that: The host computer is connected to the particle event data packet via USB3.

0.

8. The particle and event type identification system using a neural network processor according to claim 6, characterized in that: The host computer is connected to the particle event data packet via a router and an Ethernet interface in sequence.

9. The particle and event type identification system using a neural network processor according to claim 3, characterized in that: The storage module is a DDR memory or an SDRAM memory.

Images