A particle energy deposition position identification system and method based on neural network

Abstract

The invention discloses a particle energy deposition position identification system and method based on a neural network, and relates to the field of deposition position identification; a signal acquisition module acquires the radiation ray energy of a target particle, and converts the radiation ray energy into a voltage pulse signal, and performs a preprocessing operation on the voltage pulse signal to obtain a digital signal; an extraction module performs data splicing and waveform extraction on the digital signal to obtain extracted data including a first path of extracted data and a second path of extracted data; a neural network autoencoder performs noise filtering processing on the first path of extracted data to obtain a filtered signal; the neural network autoencoder is obtained by performing unsupervised training on a first neural network; a position identification module performs position identification processing on the filtered signal to obtain deposition position information; a neural network model is deployed in the position identification module; the neural network model is obtained by training a second neural network; the invention can realize the identification of the deposition position in real time and accurately.

Description

Technical Field

[0001] The present invention relates to the field of deposition position identification, and in particular to a particle energy deposition position identification system and method based on a neural network. Background Art

[0002] The energy deposition position is of great significance for applications such as particle trajectory reconstruction, particle type identification, ballistic defect correction, and radiation imaging. The traditional method of identifying the energy deposition position is to divide the detector into multiple areas and roughly distinguish the segmented intervals corresponding to the energy deposition position through multiple electrode readout signals. The detector manufacturing process is complex and expensive; another method is to use a pulse shape identification algorithm to roughly estimate the energy deposition position based on the nuclear pulse signal characteristics of multiple readout electrodes. This method is difficult to achieve real-time position identification as the detector volume increases, the number of electrodes and the amount of data processed increase. Therefore, it is necessary to develop a method that is simple and low-cost, and can be used for different types of detectors to achieve online identification of the radiation energy deposition position. Summary of the invention

[0003] The purpose of the present invention is to provide a particle energy deposition position identification system and method based on a neural network, which can realize the identification of the deposition position in real time and accurately.

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

[0005] A particle energy deposition position identification system based on a neural network, the system comprising: a signal acquisition module, an extraction module and a processor connected in sequence;

[0006] The processor includes an NPU processing unit, in which a neural network autoencoder and a position identification module are deployed; the neural network autoencoder is connected to the position identification module; the neural network autoencoder is obtained by performing unsupervised training on a first neural network; a neural network model is deployed in the position identification module; the neural network model is obtained by training a second neural network;

[0007] The signal acquisition module is used to acquire the radiation energy of the target particles, convert the radiation energy into a voltage pulse signal, and perform a preprocessing operation on the voltage pulse signal to obtain a digital signal;

[0008] The extraction module is used to perform data splicing and waveform extraction on the digital signal to obtain extracted data; the extracted data includes: first-path extracted data and second-path extracted data;

[0009] The neural network autoencoder is used to perform noise filtering on the first channel of extracted data to obtain a filtered signal;

[0010] The position identification module is used to perform position identification processing on the filtered signal to obtain deposition position information.

[0011] Optionally, the signal acquisition module comprises: a radiation detector, a preamplifier, a single-ended to differential circuit module and an analog-to-digital converter connected in sequence; the analog-to-digital converter is also connected to the extraction module;

[0012] The radiation detector is used to obtain the radiation energy of the target particles and convert the radiation energy into a voltage pulse signal;

[0013] The preamplifier is used to filter and amplify the voltage pulse signal to obtain an amplified pulse signal;

[0014] The single-ended to differential circuit module is used to convert the amplified pulse signal into a differential signal;

[0015] The analog-to-digital converter is used to perform analog-to-digital conversion on the differential signal to obtain a digital signal.

[0016] Optionally, the processor further includes: a CPU processing unit;

[0017] The CPU processing unit is connected to the extraction module and the position identification module respectively;

[0018] The extraction module is also used to extract features from the second extraction data using a digital algorithm to obtain feature data; the digital algorithm includes: a trapezoidal shaping algorithm;

[0019] The CPU processing unit is used to collect the feature data and the deposition position information.

[0020] Optionally, the extraction module is obtained by matching and combining an adder, a subtractor, an accumulator and a register.

[0021] Optionally, the system further comprises: a host computer;

[0022] The host computer is connected to the CPU processing unit; the CPU processing unit is also used to transmit the feature data and the deposition position information to the host computer.

[0023] A particle energy deposition position identification method based on a neural network, the method is implemented by using the particle energy deposition position identification system based on a neural network as described above; the method comprises:

[0024] Acquiring the radiation energy of the target particles, converting the radiation energy into a voltage pulse signal, and performing a preprocessing operation on the voltage pulse signal to obtain a digital signal;

[0025] Perform data splicing and waveform extraction on the digital signal to obtain extracted data; the extracted data includes: first-channel extracted data and second-channel extracted data;

[0026] Based on a neural network autoencoder, the first channel of extracted data is subjected to noise filtering processing to obtain a filtered signal; the neural network autoencoder is obtained by performing unsupervised training on the first neural network;

[0027] The filtered signal is subjected to position discrimination processing based on a position discrimination module to obtain deposition position information; a neural network model is deployed in the position discrimination module; and the neural network model is obtained by training a second neural network.

[0028] Optionally, the method for determining the neural network autoencoder specifically includes:

[0029] Acquire training data; the training data includes: training extraction data and preset target signal data;

[0030] Build the first neural network;

[0031] Inputting the training data into the first neural network, training the parameters of the first neural network with the goal of minimizing the first loss function, and obtaining the trained first neural network; the parameters of the first neural network include: weights; the first loss function is determined according to the gap between the output of the first neural network and the preset target signal data;

[0032] The trained first neural network is determined as a neural network autoencoder.

[0033] Optionally, the method for determining the neural network model specifically includes:

[0034] Acquire a data set; the data set includes: deposition position information and filtered signals;

[0035] Dividing the data set into a training set, a validation set, and a test set;

[0036] Constructing a second neural network; the second neural network includes an input layer, a hidden layer and an output layer connected to each other; the hidden layer includes: a convolutional layer, a batch normalization layer, a pooling layer, a random dropout layer and a fully connected layer connected in sequence;

[0037] Transmitting the training set to the hidden layer via the input layer to obtain integrated data;

[0038] The integrated data is transformed by a softmax activation function at the output layer to obtain probability distribution data; the probability distribution data is used to determine the deposition position information;

[0039] The training process of the second neural network specifically includes:

[0040] The parameters of the second neural network are trained according to the training set to meet the set number of iterations or the minimum second loss function, so as to obtain a trained second neural network; the second loss function is determined according to the mapping error between the output result of the output layer and the deposition position information in the training set; the parameters of the second neural network include: weights and offsets;

[0041] At any number of iterations, using the verification set to verify the second neural network corresponding to the number of iterations;

[0042] Determining the trained second neural network as the neural network model, and verifying the neural network model using a test set;

[0043] In the hidden layer, it specifically includes:

[0044] Using the convolutional layer to extract the waveform features of the nuclear pulse signal of the training set to obtain waveform feature data;

[0045] Using a batch normalization layer to normalize the waveform feature data to obtain normalized feature data;

[0046] The pooling layer is used to perform data volume fitting processing on the normalized feature data to obtain fitting feature data; the data volume fitting processing includes: data volume and parameter quantity reduction processing;

[0047] Regularizing the fitting feature data using a random dropout layer to obtain regularized data;

[0048] The regularized processed data is integrated using a fully connected layer to obtain integrated data.

[0049] Optionally, obtain a data set, including:

[0050] Perform finite element division on the detector to obtain two-dimensional plane map information;

[0051] Performing spatial division and numbering on the two-dimensional plan view information to obtain position number information;

[0052] A particle emission source is used to irradiate a detector with known position number information to obtain particle data; the particle data includes: radiation ray energy, pulse waveform data and deposition position information in the detector; the deposition position information is determined based on the position number information;

[0053] The particle data is determined as the data set.

[0054] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0055] The present invention provides a particle energy deposition position discrimination system and method based on a neural network, wherein a signal acquisition module obtains the radiation energy of a target particle, and converts the radiation energy into a voltage pulse signal, performs a preprocessing operation on the voltage pulse signal, and obtains a digital signal; an extraction module performs data splicing and waveform extraction on the digital signal to obtain extracted data; the extracted data includes: first-channel extracted data and second-channel extracted data; a neural network autoencoder performs noise filtering on the first-channel extracted data to obtain a filtered signal; the neural network autoencoder is obtained by performing unsupervised training on the first neural network; a position discrimination module performs position discrimination processing on the filtered signal to obtain deposition position information; a neural network model is deployed in the position discrimination module; the neural network model is obtained by training the second neural network; the present invention overcomes the problems of high cost, difficulty in maintenance, and spatial resolution being limited by segmentation size and number of electrodes in the traditional method, which require regional segmentation and a large number of electronic devices for data processing when using multiple electrodes. Therefore, the present invention can realize the discrimination of deposition positions in real time and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] 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 labor.

[0057] Figure 1 A structural diagram of a particle energy deposition position discrimination system based on a neural network provided in an embodiment of the present invention;

[0058] Figure 2 A hardware architecture diagram of a particle energy deposition position discrimination system based on a neural network provided in an embodiment of the present invention;

[0059] Figure 3 A schematic diagram of a process of performing signal conversion using a trapezoidal shaping algorithm provided by an embodiment of the present invention;

[0060] Figure 4 A schematic diagram of the structure of a neural network model provided by an embodiment of the present invention;

[0061] Figure 5 A schematic diagram of a coaxial detector provided by an embodiment of the present invention;

[0062] Figure 6 A schematic diagram of a well-type detector provided by an embodiment of the present invention;

[0063] Figure 7 A schematic diagram of a planar detector provided by an embodiment of the present invention;

[0064] Figure 8 A schematic diagram of the spatial division of a coaxial detector provided in an embodiment of the present invention;

[0065] Fig. 9 A schematic diagram of the spatial division of a well-type detector provided by an embodiment of the present invention;

[0066] Fig.10 A schematic diagram of the spatial division of a planar detector provided by an embodiment of the present invention;

[0067] Fig.11 A top view of a radiation detector provided by an embodiment of the present invention after spatial division;

[0068] Fig.12 A schematic diagram of a location for detection provided by an embodiment of the present invention;

[0069] Fig.13 A schematic diagram of a mapping relationship between a first form of signal data and corresponding deposition position information provided by an embodiment of the present invention;

[0070] Fig.14 A schematic diagram of a mapping relationship between the second form of signal data and the corresponding deposition position information provided by an embodiment of the present invention;

[0071] Fig.15 A schematic diagram of a mapping relationship between a third form of signal data and corresponding deposition position information provided by an embodiment of the present invention;

[0072] Fig.16 A schematic diagram of a process for predicting deposition position information of a coaxial detector;

[0073] Fig.17 A schematic diagram of a process for predicting deposition position information of a well-type detector;

[0074] Fig.18 Schematic diagram of the process for predicting the deposition position information of a planar detector.

[0075] Explanation of symbols:

[0076] Signal acquisition module-1, extraction module-2, processor-3. DETAILED DESCRIPTION

[0077] 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.

[0078] The difference in the shape of two nuclear pulse signals is related to the distribution of energy deposition positions. When the energy deposition positions are close, the difference in their signal shapes is also small. The conventional pulse shape identification PSD technology has low sensitivity to the shape characteristics of nuclear pulse signals and is difficult to accurately identify subtle changes in signals. Therefore, it cannot be applied to the identification of energy deposition positions. With the development of computer technology, artificial neural network technology and SOC chips with neural network processors NPU have also developed rapidly. Deploying artificial neural network models in embedded systems for radiation event information extraction is the development trend of future nuclear technology. Based on this background, a method for nuclear pulse signal feature learning and energy deposition position identification based on artificial neural network technology is proposed, and a digital nuclear pulse signal processor based on neural network processor architecture is developed. The use of artificial neural network models deployed in NPU to perform online identification of energy deposition positions is very meaningful for the development of nuclear technology.

[0079] The purpose of the present invention is to provide a particle energy deposition position identification system and method based on a neural network, which can realize the identification of the deposition position in real time and accurately.

[0080] 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.

[0081] Example 1

[0082] like Figure 1 As shown, an embodiment of the present invention provides a particle energy deposition position identification system based on a neural network, the system comprising: a signal acquisition module 1, an extraction module 2 and a processor 3 connected in sequence.

[0083] Processor 3 includes an NPU processing unit, in which a neural network autoencoder and a position identification module are deployed; the neural network autoencoder and the position identification module are connected; the neural network autoencoder is obtained by unsupervised training of the first neural network; a neural network model is deployed in the position identification module; the neural network model is obtained by training the second neural network.

[0084] The signal acquisition module 1 is used to acquire the radiation energy of the target particles, convert the radiation energy into a voltage pulse signal, and perform preprocessing operations on the voltage pulse signal to obtain a digital signal.

[0085] The extraction module 2 is used to perform data splicing and waveform extraction on the digital signal to obtain extracted data. The extracted data includes: first-channel extracted data and second-channel extracted data.

[0086] The neural network autoencoder is used to perform noise filtering on the first channel of extracted data to obtain a filtered signal.

[0087] The position discrimination module is used to perform position discrimination processing on the filtered signal to obtain deposition position information.

[0088] Specifically, the signal acquisition module 1 includes: a radiation detector, a preamplifier, a single-ended to differential circuit module and an analog-to-digital converter which are connected in sequence; the analog-to-digital converter is also connected to the extraction module.

[0089] The radiation detector is used to obtain the radiation energy of the target particles and convert the radiation energy into a voltage pulse signal.

[0090] The preamplifier is used to filter and amplify the voltage pulse signal to obtain an amplified pulse signal.

[0091] The single-ended to differential circuit module is used to convert the amplified pulse signal into a differential signal.

[0092] The analog-to-digital converter is used to perform analog-to-digital conversion on the differential signal to obtain a digital signal.

[0093] In one embodiment, the processor 3 also includes: a CPU processing unit; the CPU processing unit is connected to the extraction module and the position identification module respectively; the extraction module 2 is also used to use a digital algorithm to extract features from the second extraction data to obtain feature data; the digital algorithm includes: a trapezoidal forming algorithm; the CPU processing unit is used to collect feature data and deposition position information.

[0094] The extraction module 2 is obtained by matching and combining an adder, a subtractor, an accumulator and a register.

[0095] As an optional implementation, the system further includes: a host computer; the host computer is connected to the CPU processing unit; the CPU processing unit is also used to transmit the characteristic data and deposition position information to the host computer.

[0096] Figure 2The system hardware design block diagram provided by the present invention includes a SoC processor, i.e., a processor 3, an FPGA circuit, i.e., an extraction module 2, a preamplifier, a single-ended to differential circuit module, and a high-speed ADC circuit, i.e., an analog-to-digital converter. In the system, the nuclear pulse signal measured and acquired by the radiation detector is processed by the preamplifier and the single-ended to differential circuit, and then converted into a digital signal by the high-speed ADC and transmitted to the FPGA. The digital nuclear pulse signal, i.e., the digital signal, is divided into two paths after data splicing and waveform extraction in the FPGA, one path is transmitted to the digital algorithm of the later stage in the FPGA for nuclear pulse signal feature extraction, and the other path is transmitted to the SoC processor, and the artificial neural network of the NPU processing unit deployed in the SoC processor, i.e., the artificial neural network in the autoencoder and the position discrimination module, i.e., the unsupervised neural network model deployed in the autoencoder and the neural network model deployed in the position discrimination module are used to filter, reduce noise and identify the position of the nuclear pulse signal. Finally, the position number data identified by the artificial neural network model and the feature parameter data extracted by the digital algorithm in the FPGA are transmitted to the CPU processing unit in the SoC processor. Under the control of the CPU processing unit, the data is packaged and transmitted to the PC host computer through USB3.0 or Ethernet communication.

[0097] In this system, since the neural network model needs to perform position discrimination 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. Since the artificial neural network model needs to perform reasoning based on the characteristics of the waveform on the nuclear pulse signal to achieve position discrimination, it is necessary to retain the waveform characteristics of the original signal as much as possible before the analog-to-digital conversion of the nuclear pulse signal. In order to achieve this goal, in the process of designing the analog circuit, the analog circuit is simplified as much as possible before the high-speed ADC circuit to avoid being affected by the gain-bandwidth product of the operational amplifier, which causes a certain degree of distortion of the nuclear pulse signal, thereby affecting the accuracy of discrimination. Therefore, in the present invention, the nuclear pulse signal output from the radiation detector directly enters the high-speed ADC circuit for analog-to-digital conversion after passing through the preamplifier and the single-ended to differential circuit module.

[0098] Specifically, the description of each component of the hardware architecture in this system is as follows:

[0099] Radiation detector: The interaction between rays and radiation detectors deposits energy in the detector. The radiation detector is used to convert the energy of radiation rays into a voltage pulse signal so that the pulse signal can be processed by the back-end electronic circuit to extract effective information.

[0100] Preamplifier: The preamplifier mainly filters and amplifies the nuclear pulse signal output by the radiation detector. Since the original nuclear pulse signal output from the radiation detector has a small amplitude, it is easily affected by noise and interference from the outside world during the transmission process. Therefore, the preamplifier should be designed as close to the radiation detector as possible.

[0101] Single-ended to differential circuit module: The function of the single-ended to differential circuit is to convert the single-ended signal output by the radiation detector into a differential signal. This conversion circuit plays an important role in the radiation detection system. The radiation detector usually outputs a single-ended signal, which corresponds to the characteristics of the radiation source. However, the single-ended signal is susceptible to external interference such as power supply noise, electromagnetic interference, etc. These interferences 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 the 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. 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, so as to better capture and analyze radiation events.

[0102] High-speed ADC circuit, i.e., analog-to-digital converter (ADC): performs analog-to-digital conversion on pulse signals, converts analog signals into digital signals, inputs them into FPGA for data splicing, and performs calculations and analysis in the FPGA's post-digital algorithm module and SoC processor. In order to obtain more accurate analysis results in the calculation and analysis of FPGA and SoC processor, the sampling rate and resolution of ADC should be as high as possible to obtain sufficient sampling points and sampling accuracy.

[0103] SoC processor, i.e. processor: Two artificial neural network models are mainly deployed in the circuit of the NPU processing unit, namely the neural network autoencoder for noise reduction filtering and the convolutional neural network model for identifying the energy deposition position of the ray. The digital nuclear pulse signal output by the high-speed ADC is transmitted to the circuit 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. After that, the signal is divided into two paths, one is transmitted to the next-level artificial neural network model for identifying the energy deposition position information, and the other is sent to the FPGA circuit through parallel transmission for further pulse waveform information extraction. The neural network autoencoder is an unsupervised neural network model. During the training process, it inputs nuclear pulse waveform data superimposed with various noises into the input layer, and outputs a string of new waveform data at the output layer after the operation of the hidden layer. 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 at the input layer is close to the ideal waveform data without superimposed noise after the operation of the model. 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. Deploying this model in the NPU can realize the function of filtering and denoising the nuclear pulse signal input to the NPU.

[0104] FPGA circuit, that is, the extraction module: In certain measurement environments, some other types of nuclear pulse waveform information, such as the amplitude of the nuclear pulse signal and the duration of the pulse front, can also be extracted through the digital algorithm deployed in the FPGA. After filtering and shaping the nuclear pulse signal using some classic digital shaping algorithms, such as the trapezoidal shaping algorithm, the triangle shaping algorithm, and the Gaussian shaping algorithm, its amplitude information can be extracted; the frontier time information of the nuclear pulse signal can be extracted using the constant ratio timing digital algorithm. These are some common digital algorithms. The nuclear pulse waveform data after filtering and denoising by the neural network autoencoder in the NPU circuit is transmitted to the FIFO of the FPGA circuit through parallel communication and other communication methods for caching, and the digital algorithm in the FPGA reads the data from the FIFO for parameter extraction.

[0105] Working principle of the autoencoder based on artificial neural network: The artificial neural network autoencoder is a kind of neural network model, which consists of an encoder and a decoder in the hidden layer. Through training, the artificial neural network autoencoder can learn the multi-level feature expression of the data and capture higher-level abstract features. The input core pulse signal is composed of an ideal signal and noise superimposed on the ideal signal. The goal of the neural network autoencoder is to process the input core pulse signal to obtain the target signal, so that the target signal is as close to the ideal signal as possible. Therefore, during the training process of the autoencoder, the autoencoder continuously optimizes its own weight parameters so that the ideal signal in the input core pulse signal is completely preserved. The noise superimposed on the ideal signal will increase the difference between the target signal and the ideal signal, causing the loss function to increase. Therefore, during the training process, in order to minimize the loss function, these noises will be filtered out. When the loss function reaches the minimum value, the target signal output by the autoencoder is closest to the ideal signal, and the noise superimposed on the ideal signal is suppressed to the greatest extent. When the neural network autoencoder is working, the original nuclear pulse waveform data is input into the input layer, the trained neural network autoencoder operates on the pulse waveform data, and the encoder of the hidden layer compresses the data. During the compression process, the noise is filtered out, while the waveform characteristics of the nuclear pulse signal are retained. The waveform data is then restored through the decoder. The restored waveform is the waveform after suppressing the noise while retaining the original signal characteristics.

[0106] Implementation of digital algorithms: Using digital algorithms deployed in FPGAs to extract characteristic parameters from nuclear pulse signals is a conventional technical method for nuclear pulse signal processors. The most common digital algorithm is to use a trapezoidal forming algorithm to transform the double exponential decay signal of a nuclear pulse into a trapezoidal shape, and then obtain the amplitude at the flat top of the trapezoid. This amplitude reflects the amplitude of the corresponding nuclear pulse signal. There are a series of advantages to using a trapezoidal forming algorithm to extract the amplitude of a nuclear pulse signal. For example, the nuclear pulse signal is also filtered and denoised during the trapezoidal forming process, and the true amplitude of the accumulated pulse signal is restored. The use of symmetrical zero-area trapezoidal forming also has the ability to resist baseline drift, which can more accurately extract the amplitude of the nuclear pulse signal. The schematic diagram of the nuclear pulse signal conversion process using the trapezoidal forming algorithm is shown below. Figure 3 shown. Figure 3 The conversion process of the nuclear pulse signal can be realized in the FPGA through the combination of hardware logic resources such as adders, subtractors, accumulators, and registers.

[0107] The specific workflow of the system is as follows:

[0108] In this system, the nuclear pulse signal measured by the radiation detector is processed by the preamplifier and the single-ended to differential circuit, and then converted into a digital signal by the high-speed ADC and transmitted to the FPGA. The data splicing module in the FPGA splices the bit data transmitted by the high-speed ADC into half-byte data and passes it to the waveform extraction module of the next level. The nuclear pulse signal of a fixed time length is extracted from the original waveform data. After that, the nuclear pulse signal is divided into two paths. One path is transmitted to the digital algorithm module of the FPGA back-end for nuclear pulse signal feature extraction. The extracted nuclear pulse signal feature parameter data is transmitted to the memory space opened up by the ARM architecture CPU processing unit of the SoC processor through parallel communication, and the other path of nuclear pulse signal data is transmitted to the So In the C processor, under the control of the CPU processing unit, the nuclear pulse signal data is loaded into the input structure variable of the artificial neural network model deployed in the NPU processing unit, and the nuclear pulse signal is denoised and filtered by the artificial neural network autoencoder, and the artificial neural network model used for position identification analyzes and infers the probability of each energy deposition position based on the nuclear pulse signal data, that is, each position number has a probability value, and the inference result is saved in the model output structure variable, and then the CPU processing unit in the SoC processor post-processes the output structure variable to extract the position number with the largest probability value. Finally, the CPU processing unit transmits the position number data to the host computer on the PC through USB3.0 or Ethernet communication.

[0109] Example 2

[0110] The embodiment of the present invention provides a particle energy deposition position identification method based on a neural network, and the method is implemented by using the particle energy deposition position identification system in Example 1; the method comprises:

[0111] Step 100: Acquire the radiation energy of the target particle, convert the radiation energy into a voltage pulse signal, and perform a preprocessing operation on the voltage pulse signal to obtain a digital signal.

[0112] Step 200: perform data splicing and waveform extraction on the digital signal to obtain extracted data. The extracted data includes: first-channel extracted data and second-channel extracted data.

[0113] Step 300: Perform noise filtering on the first path of extracted data based on a neural network autoencoder to obtain a filtered signal. The neural network autoencoder is obtained by performing unsupervised training on the first neural network.

[0114] Step 400: Performing position discrimination processing on the filtered signal based on the position discrimination module to obtain deposition position information; a neural network model is deployed in the position discrimination module; and the neural network model is obtained by training the second neural network.

[0115] The method for determining the neural network autoencoder specifically includes:

[0116] Acquire training data; the training data includes: training extraction data and preset target signal data.

[0117] Construct a first neural network; input training data into the first neural network, train the parameters of the first neural network with the goal of minimizing the first loss function, and obtain the trained first neural network; the parameters of the first neural network include: weights; the first loss function is determined based on the gap between the output of the first neural network and the preset target signal data.

[0118] The trained first neural network is determined as a neural network autoencoder.

[0119] In addition, the method for determining the neural network model specifically includes:

[0120] A data set is obtained; the data set includes: deposition location information and filtered signals.

[0121] Divide the dataset into training, validation, and test sets.

[0122] A second neural network is constructed; the second neural network includes an input layer, a hidden layer, and an output layer connected to each other; the hidden layer includes: a convolutional layer, a batch normalization layer, a pooling layer, a random dropout layer, and a fully connected layer connected in sequence.

[0123] The training set is transferred to the hidden layer through the input layer to obtain integrated data.

[0124] The integrated data is transformed with the softmax activation function in the output layer to obtain probability distribution data; the probability distribution data is used to determine the deposition location information.

[0125] The training process of the second neural network specifically includes:

[0126] With the goal of satisfying the set number of iterations or minimizing the second loss function, the parameters of the second neural network are trained according to the training set to obtain a trained second neural network; the second loss function is determined according to the mapping error between the output result of the output layer and the deposition position information in the training set; the parameters of the second neural network include: weights and offsets.

[0127] At any number of iterations, the verification set is used to verify the second neural network corresponding to the number of iterations.

[0128] The trained second neural network is determined as the neural network model, and the neural network model is verified using a test set.

[0129] In the hidden layer, it specifically includes:

[0130] The convolution layer is used to extract the waveform features of the nuclear pulse signal of the training set to obtain the waveform feature data.

[0131] The batch normalization layer is used to normalize the waveform feature data to obtain normalized feature data.

[0132] The pooling layer is used to perform data volume fitting processing on the normalized feature data to obtain fitted feature data; the data volume fitting processing includes: reducing the data volume and the number of parameters.

[0133] A random dropout layer is used to regularize the fitted feature data to obtain regularized data.

[0134] A fully connected layer is used to integrate the regularized data to obtain integrated data.

[0135] In one embodiment, obtaining a data set specifically includes:

[0136] The detector is divided by finite elements to obtain two-dimensional plane map information; the two-dimensional plane map information is spatially divided and numbered to obtain position number information; a particle emission source is used to irradiate the detector with known position number information to obtain particle data; the particle data includes: radiation ray energy, pulse waveform data and deposition position information in the detector; the deposition position information is determined based on the position number information.

[0137] Identify the particle data as a data set.

[0138] In practical applications, this method is based on a physical mechanism: the nuclear pulse signals generated by energy deposition at different positions in the detector have their own shape characteristics. The position of energy deposition can be inferred based on the shape characteristics of the nuclear pulse signals, thereby realizing the identification of the energy deposition position.

[0139] This method proposes a method for learning nuclear pulse signal features and identifying energy deposition positions based on artificial neural network technology. By training the artificial neural network model, the model learns the shape features of nuclear pulse signals corresponding to energy deposition at different positions. The model saves the mapping relationship between the learned shape features of nuclear pulse signals and the energy deposition positions in the model in the form of weight parameters. For a medium-sized model, a large number of weight parameters of the model can enable it to characterize more detailed features on the nuclear pulse signal, thereby achieving higher-precision position identification. In order to achieve this goal, neural network model construction, data set preparation, model training, model deployment and other work were carried out, and the neural network model deployed in the nuclear pulse signal processor was used to perform online identification of the ray energy deposition position.

[0140] 1. Model construction.

[0141] Nuclear pulse signal data can be regarded as one-dimensional sequence data, so convolutional neural network (CNN) is used to calculate and analyze nuclear pulse signal data to identify the location of energy deposition of radiation in the detector. A convolutional neural network model is built, and the structure of the model is as follows Figure 4 As shown. The convolution layer is used to extract features from the nuclear pulse signal waveform, and the batch normalization layer is used to normalize the nuclear pulse waveform data so that the waveform data is at the same level, eliminating the impact of the pulse waveform amplitude difference caused by the difference in ray deposition energy on model identification; the maximum pooling layer is used to reduce the amount of data and the number of parameters, extract useful feature data, and reduce overfitting to a certain extent; random inactivation is to randomly return some weights or outputs of the hidden layer to zero, reduce the mutual dependence between nodes, thereby realizing the regularization of the neural network and preventing overfitting, and at the same time can reduce the network parameters involved in the calculation and reduce 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, and 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.

[0142] 2. Preparation of data set.

[0143] The key to model training lies in the preparation of data sets. The overall idea is as follows: divide the three-dimensional volume of the detector into finite elements and number each finite element. Use the finite element number as the classification label of the nuclear pulse signal, and obtain the nuclear pulse signal data generated by the energy deposition position at the corresponding finite element through experiments. A set of nuclear pulse signal data and its corresponding finite element number constitute a set of data sets. After a large number of data sets are obtained in the experiment, the data sets are divided into training sets, validation sets, and test sets. This method realizes online identification of particle energy deposition positions through the following five steps.

[0144] a. The present invention mainly studies three types of detectors: coaxial type, well type and planar type. The detector structure is as follows Figure 5-Figure 7 As shown. Data sets for realizing position discrimination of the three detectors are respectively produced. The method proposed in the present invention is also applicable to detectors of other shapes.

[0145] b. Perform finite element division on the detector. Establish a cylindrical coordinate system, which consists of three coordinate parameters: radius r, polar angle θ, and height z. Since the detectors of the three shapes are symmetrical along the axis, the influence of the lattice direction on the carrier drift track is ignored. The change of polar angle θ will not affect the shape of the nuclear pulse signal. Therefore, the two parameters r and z are used to characterize the location of energy deposition. Then simplify the detector structure into a two-dimensional plane diagram, and then divide and number the two-dimensional plane diagram in space after simplification, such as Figure 8-Figure 11 shown.

[0146] c. Place collimator A with a small hole above the main detector, the collimator is parallel to the XY plane, and the radiation source is placed above the collimator. The gamma rays enter the main detector through the small hole of the collimator along the negative direction of the Z axis. Place another collimator B with a small hole on the side of the main detector. The small hole of collimator B falls on the plane formed by the small hole of collimator A and the axis of the main detector. A coincidence detector is placed on the side close to the small hole of collimator A and on the other side of collimator B. It is used to measure the gamma rays that escape from the main detector and pass through the small hole of collimator B. The positional relationship between the radiation source, collimator, main detector and coincidence detector is shown in the figure. Fig.12 shown.

[0147] d. The gamma rays emitted by the radiation source enter the main detector through the pinhole of collimator A, and the radius r of the energy deposition position is determined. The rays undergo Compton scattering at a height z in the main detector, depositing energy. The gamma rays with the remaining energy escape from the main detector after being deflected by a certain angle. The pinhole of collimator B selects the gamma rays escaping from the main detector with a deflection angle of 90°. These rays are detected by the coincidence detector, and the height z of the energy deposition position is determined. A coincidence event is detected. The nuclear pulse signal of the coincidence event output by the main detector is the nuclear pulse signal at the energy deposition position with a radius r and a height z. The spatial number corresponding to this position is the classification label of the nuclear pulse signal. Figure 13-Figure 15 It can be seen that the shapes of nuclear pulse signals corresponding to different numbered positions are somewhat different. Conversely, the energy deposition position can be inferred based on the characteristics of the nuclear pulse signals.

[0148] e. Adjust the positions of collimators A and B, and obtain a large amount of nuclear pulse signal data and corresponding classification labels for each number from the main detector electronics system in an experimental environment. The waveform data of the nuclear pulse signal and the corresponding labels constitute the data set for training the neural network model.

[0149] 3. Model training.

[0150] 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 512, the maximum number of iterations is 50, and the initial learning rate is 0.002. 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 neural network model continuously optimizes its own weights and other parameters during the training process, so that the model can accurately output the corresponding position code in the output layer after the hidden layer operation according to the input nuclear pulse signal data. When the accuracy rate is improved to a certain level, the training is stopped. At this time, the trained neural network model can output the corresponding spatial position number according to the input nuclear pulse signal data to realize the identification of the gamma ray energy deposition position.

[0151] 4. Model deployment and online analysis.

[0152] 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, the NPU processing engine performs data calculations according to the structure and operators of the previously trained neural network model to calculate and analyze the real-time input nuclear pulse signal data, thereby inferring the probability value of each position number of the energy deposition of the gamma ray corresponding to the nuclear pulse signal in the detector, and taking the position number with the largest probability value as the inference result, to achieve real-time analysis of the ray energy deposition position. The nuclear pulse signal processor collects pulse waveform data in real time and analyzes it through the NPU, outputs the position number, and obtains the detector area corresponding to the number according to the position number. The schematic diagram of the whole process is shown below. Figure 16-18 As shown, the neural network model deployed in the NPU infers the location number of the energy deposition based on the input nuclear pulse signal data, and the number maps the spatial area within the detector.

[0153] The present invention designs a nuclear pulse signal processor hardware architecture based on NPU for online identification of particle energy deposition positions. By deploying an artificial neural network model for identifying energy deposition positions in the NPU, online identification of particle energy deposition positions is achieved, overcoming the shortcoming that the traditional FPGA-based nuclear pulse signal processor hardware architecture cannot achieve online identification of particle energy deposition positions on a single board.

[0154] In addition, the present invention also proposes an energy deposition position identification method based on artificial neural network technology. By training the artificial neural network model, a single model can predict the energy deposition position of the ray in the entire detector according to the characteristics of the nuclear pulse signal. This overcomes the problems of the traditional method of needing to segment the detector and the high cost and difficulty in maintenance of a large number of electronic devices for data processing when using multiple electrodes, as well as the spatial resolution being limited by the segmentation size and the number of electrodes. At the same time, this method will also promote the miniaturization and industrialization of instruments used for energy deposition position identification.

[0155] 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.

[0156] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea 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 energy deposition position identification system based on neural network, characterized in that: The system comprises: a signal acquisition module, an extraction module and a processor connected in sequence; The processor includes an NPU processing unit, in which a neural network autoencoder and a position identification module are deployed; the neural network autoencoder is connected to the position identification module; the neural network autoencoder is obtained by performing unsupervised training on a first neural network; a neural network model is deployed in the position identification module; the neural network model is obtained by training a second neural network; The signal acquisition module is used to acquire the radiation energy of the target particles, convert the radiation energy into a voltage pulse signal, and perform a preprocessing operation on the voltage pulse signal to obtain a digital signal; The extraction module is used to perform data splicing and waveform extraction on the digital signal to obtain extracted data; the extracted data includes: first-path extracted data and second-path extracted data; The neural network autoencoder is used to perform noise filtering on the first channel of extracted data to obtain a filtered signal; The position identification module is used to perform position identification processing on the filtered signal to obtain deposition position information.

2. The particle energy deposition position identification system based on neural network according to claim 1 is characterized in that: The signal acquisition module comprises: a radiation detector, a preamplifier, a single-ended to differential circuit module and an analog-to-digital converter connected in sequence; the analog-to-digital converter is also connected to the extraction module; The radiation detector is used to obtain the radiation energy of the target particles and convert the radiation energy into a voltage pulse signal; The preamplifier is used to filter and amplify the voltage pulse signal to obtain an amplified pulse signal; The single-ended to differential circuit module is used to convert the amplified pulse signal into a differential signal; The analog-to-digital converter is used to perform analog-to-digital conversion on the differential signal to obtain a digital signal.

3. The particle energy deposition position identification system based on neural network according to claim 1 is characterized in that: The processor also includes: a CPU processing unit; The CPU processing unit is connected to the extraction module and the position identification module respectively; The extraction module is also used to extract features from the second extraction data using a digital algorithm to obtain feature data; the digital algorithm includes: a trapezoidal shaping algorithm; The CPU processing unit is used to collect the feature data and the deposition position information.

4. The particle energy deposition position identification system based on neural network according to claim 3 is characterized in that: The extraction module is obtained by matching and combining an adder, a subtractor, an accumulator and a register.

5. The particle energy deposition position identification system based on neural network according to claim 3 is characterized in that: The system also includes: a host computer; The host computer is connected to the CPU processing unit; the CPU processing unit is also used to transmit the feature data and the deposition position information to the host computer.

6. A particle energy deposition position identification method based on a neural network, characterized in that: The method is implemented by using the particle energy deposition position identification system based on a neural network as described in any one of claims 1 to 5; the method comprises: Acquiring the radiation energy of the target particles, converting the radiation energy into a voltage pulse signal, and performing a preprocessing operation on the voltage pulse signal to obtain a digital signal; Perform data splicing and waveform extraction on the digital signal to obtain extracted data; the extracted data includes: first-channel extracted data and second-channel extracted data; Based on a neural network autoencoder, the first channel of extracted data is subjected to noise filtering processing to obtain a filtered signal; the neural network autoencoder is obtained by performing unsupervised training on the first neural network; The filtered signal is subjected to position discrimination processing based on a position discrimination module to obtain deposition position information; a neural network model is deployed in the position discrimination module; and the neural network model is obtained by training a second neural network.

7. The particle energy deposition position identification method based on neural network according to claim 6 is characterized in that: The method for determining the neural network autoencoder specifically includes: Acquire training data; the training data includes: training extraction data and preset target signal data; Build the first neural network; Inputting the training data into the first neural network, training the parameters of the first neural network with the goal of minimizing the first loss function, and obtaining the trained first neural network; the parameters of the first neural network include: weights; the first loss function is determined according to the gap between the output of the first neural network and the preset target signal data; The trained first neural network is determined as a neural network autoencoder.

8. The particle energy deposition position identification method based on neural network according to claim 6 is characterized in that: The method for determining the neural network model specifically includes: Acquire a data set; the data set includes: deposition position information and filtered signals; Dividing the data set into a training set, a validation set, and a test set; Constructing a second neural network; the second neural network includes an input layer, a hidden layer and an output layer connected to each other; the hidden layer includes: a convolutional layer, a batch normalization layer, a pooling layer, a random dropout layer and a fully connected layer connected in sequence; Transmitting the training set to the hidden layer via the input layer to obtain integrated data; The integrated data is transformed by a softmax activation function at the output layer to obtain probability distribution data; the probability distribution data is used to determine the deposition position information; The training process of the second neural network specifically includes: The parameters of the second neural network are trained according to the training set to meet the set number of iterations or the minimum second loss function, so as to obtain a trained second neural network; the second loss function is determined according to the mapping error between the output result of the output layer and the deposition position information in the training set; the parameters of the second neural network include: weights and offsets; At any number of iterations, using the verification set to verify the second neural network corresponding to the number of iterations; Determining the trained second neural network as the neural network model, and verifying the neural network model using a test set; In the hidden layer, it specifically includes: Using the convolutional layer to extract the waveform features of the nuclear pulse signal of the training set to obtain waveform feature data; Using a batch normalization layer to normalize the waveform feature data to obtain normalized feature data; The pooling layer is used to perform data volume fitting processing on the normalized feature data to obtain fitting feature data; the data volume fitting processing includes: data volume and parameter quantity reduction processing; Regularizing the fitting feature data using a random dropout layer to obtain regularized data; The regularized processed data is integrated using a fully connected layer to obtain integrated data.

9. The particle energy deposition position identification method based on neural network according to claim 8, characterized in that: Get the dataset, including: Perform finite element division on the detector to obtain two-dimensional plane map information; Performing spatial division and numbering on the two-dimensional plan view information to obtain position number information; A particle emission source is used to irradiate a detector with known position number information to obtain particle data; the particle data includes: radiation ray energy, pulse waveform data and deposition position information in the detector; the deposition position information is determined based on the position number information; The particle data is determined as the data set.

Images