A nuclear pulse signal processing architecture and method based on neural network autoencoder

Abstract

The present invention discloses a nuclear pulse signal processing architecture and method based on a neural network autoencoder, and relates to the field of nuclear pulse signal filtering technology. The architecture includes a nuclear pulse signal preprocessing module, a SoC processor, and a ZYNQ programmable system on chip that are connected in sequence; the trained neural network autoencoder model is integrated in the SoC processor to filter and denoise the original signal in sequence, and then the parameter information of the nuclear pulse signal after filtering and denoising is extracted. The present invention can improve the filtering and denoising effect of the nuclear pulse signal by integrating the trained neural network autoencoder model in the SoC processor, while reducing the distortion of the nuclear pulse signal after filtering and denoising.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear pulse signal filtering, and in particular to a nuclear pulse signal processing architecture and method based on a neural network autoencoder. Background Art

[0002] In radiation detection, noise can affect threshold triggering, pulse signal shape discrimination, and the energy resolution of the system. Traditional methods mainly suppress system noise through analog or digital CR-RC (CR-RCm filter), trapezoidal shaping, and DPLMS optimal filtering. However, these methods not only have general noise suppression capabilities, but also distort the signal to varying degrees. Therefore, it is necessary to develop a nuclear pulse signal filtering and denoising method that can retain the subtle characteristics of the nuclear pulse signal waveform, minimize signal distortion, and simultaneously suppress the noise to the greatest extent online. Summary of the invention

[0003] The purpose of the present invention is to provide a nuclear pulse signal processing architecture and method based on a neural network autoencoder, which can improve the filtering and denoising effect of the nuclear pulse signal and reduce the distortion of the nuclear pulse signal after filtering and denoising processing.

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

[0005] A nuclear pulse signal processing architecture based on a neural network autoencoder, comprising: a nuclear pulse signal preprocessing module, a SoC (System on Chip) processor and a ZYNQ programmable system on chip connected in sequence;

[0006] The ZYNQ (Zynq-7000 All Programmable SoC, fully programmable system on chip) programmable system on chip is connected to the signal acquisition terminal;

[0007] The nuclear pulse signal preprocessing module is used to perform filtering processing, amplification processing and voltage bias processing on the original signal in sequence. When the nuclear pulse signal exists in the original signal after the amplification processing, the original signal after the voltage bias processing is subjected to analog-to-digital conversion processing to obtain the nuclear pulse digital signal to be denoised; the original signal is a baseline signal randomly superimposed with the nuclear pulse signal;

[0008] The SoC processor is integrated with a trained neural network autoencoder model; the SoC processor is used to load the core pulse digital signal to be denoised into an input structure variable, input the input structure variable into the trained neural network autoencoder model to obtain the core pulse signal after filtering and denoising, cache the core pulse signal after filtering and denoising in an output structure variable, and transmit the core pulse signal after filtering and denoising in the output structure variable to the ZYNQ programmable system on chip; the trained neural network autoencoder model is obtained by self-supervising the neural network autoencoder model using the noisy core pulse digital signal and the noise-free core pulse digital signal;

[0009] The ZYNQ programmable system-on-chip integrates a variety of digital algorithms; the ZYNQ programmable system-on-chip is used to use the multiple digital algorithms to extract parameter information of the nuclear pulse signal after filtering and noise reduction processing, store the parameter information and transmit the parameter information to the signal acquisition terminal.

[0010] Optionally, the digitization algorithm includes a shaping algorithm, an amplitude extraction algorithm, a constant ratio timing digitization algorithm and a time extraction algorithm.

[0011] Optionally, the nuclear pulse signal preprocessing module includes: a gain adjustment circuit, a DC offset adjustment circuit, a threshold trigger circuit and a high-speed ADC (Analog-to-Digital Converter) circuit;

[0012] The gain adjustment circuit is connected to the DC offset adjustment circuit and the threshold trigger circuit respectively; the DC offset adjustment circuit and the threshold trigger circuit are both connected to the high-speed ADC circuit; the high-speed ADC circuit is connected to the SoC processor;

[0013] The gain adjustment circuit is used to perform filtering and amplification processing on the original signal in sequence to obtain the original signal after amplification processing;

[0014] The DC offset adjustment circuit is used to perform voltage bias processing on the original signal after amplification processing to obtain the original signal after voltage bias processing;

[0015] The threshold trigger circuit is used to generate a trigger signal according to the original signal after amplification processing; the threshold trigger circuit is used to output a high level as a trigger signal when the signal amplitude of the original signal after amplification processing reaches the amplitude threshold, and output a low level as a trigger signal when the signal amplitude of the original signal after amplification processing is less than the amplitude threshold;

[0016] The high-speed ADC circuit is used to perform analog-to-digital conversion processing on the original signal after voltage bias processing when the trigger signal is at a high level, so as to obtain the core pulse digital signal to be de-noised.

[0017] Optionally, the nuclear pulse signal preprocessing module further includes: a radiation detector;

[0018] The radiation detector is connected to the gain adjustment circuit; the radiation detector is used to obtain the original signal.

[0019] Optionally, the SoC processor includes: a first CPU (Central Processing Unit) processing unit and an NPU (embedded neural-network processor, neural-networkprocess units) processing unit;

[0020] The first CPU processing unit is respectively connected to the nuclear pulse signal preprocessing module, the NPU processing unit and the ZYNQ programmable system on chip;

[0021] The first CPU processing unit is used to load the core pulse digital signal to be denoised into an input structure variable;

[0022] The NPU processing unit is integrated with a trained neural network autoencoder model; the NPU processing unit is used to input the input structure variable into the trained neural network autoencoder model to obtain a nuclear pulse signal after filtering and denoising, and cache the nuclear pulse signal after filtering and denoising in an output structure variable;

[0023] The first CPU processing unit is also used to transmit the nuclear pulse signal after filtering and noise reduction processing in the output structure variable to the ZYNQ programmable chip system.

[0024] Optionally, the ZYNQ programmable system-on-chip includes: a programmable logic unit and a second CPU processing unit;

[0025] The programmable logic unit is connected to the SoC processor and the second CPU processing unit respectively; the second CPU processing unit is also connected to the signal acquisition terminal;

[0026] The programmable logic unit integrates a plurality of digital algorithms; the programmable logic unit is used to extract parameter information of the nuclear pulse signal after filtering and noise reduction processing by using the plurality of digital algorithms;

[0027] The second CPU processing unit is used to store the parameter information and transmit the parameter information to the signal acquisition terminal.

[0028] Optionally, the neural network autoencoder model is a fully connected neural network or a convolutional neural network.

[0029] A nuclear pulse signal processing method based on a neural network autoencoder, comprising:

[0030] Acquire an original signal; the original signal is a baseline signal randomly superimposed with a nuclear pulse signal;

[0031] The original signal is subjected to filtering, amplification and voltage bias processing in sequence;

[0032] When a nuclear pulse signal exists in the original signal after the amplification process, the original signal after the voltage bias process is subjected to analog-to-digital conversion to obtain a nuclear pulse digital signal to be subjected to noise reduction;

[0033] Loading the core pulse digital signal to be denoised into an input structure variable;

[0034] Inputting the input structure variable into the trained neural network autoencoder model to obtain an output structure variable; the output structure variable includes the core pulse signal after filtering and noise reduction processing; the trained neural network autoencoder model is obtained by self-supervising the neural network autoencoder model using the noisy core pulse digital signal and the noise-free core pulse digital signal;

[0035] Utilize a variety of digital algorithms to extract parameter information of nuclear pulse signals after filtering and noise reduction;

[0036] The parameter information is stored and transmitted.

[0037] Optionally, before obtaining the original signal, the method further includes:

[0038] Constructing a neural network autoencoder model; the neural network autoencoder model is a fully connected neural network or a convolutional neural network;

[0039] The noisy kernel pulse digital signal is input into the neural network autoencoder model to obtain the denoised kernel pulse digital signal;

[0040] Determine a loss function value according to the denoised core pulse digital signal and the noise-free core pulse digital signal;

[0041] According to the loss function value, the parameters of the neural network autoencoder model are adjusted and the step of "inputting the noisy core pulse digital signal into the neural network autoencoder model to obtain a denoised core pulse digital signal" is returned until the loss function values ​​of a preset number of consecutive iterations are less than the loss function value threshold, thereby obtaining a trained neural network autoencoder model.

[0042] Optionally, the neural network autoencoder model includes an input layer, multiple encoding layers, multiple decoding layers and an output layer connected in sequence;

[0043] The multiple coding layers include a plurality of coding layers connected in sequence;

[0044] The multiple decoding layers include a plurality of decoding layers connected in sequence;

[0045] The number of the encoding layers is equal to the number of the decoding layers.

[0046] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0047] The present invention provides a nuclear pulse signal processing architecture and method based on a neural network autoencoder. By deploying the neural network autoencoder in an NPU, medium and large-scale neural network autoencoders can be deployed in embedded devices, thereby maximizing the filtering and noise reduction capabilities of the neural network autoencoder, overcoming the shortcomings of the traditional nuclear pulse signal processor hardware architecture based on FPGA (Field-Programmable Gate Array, field programmable gate array) such as low neural network model computing efficiency, high power consumption, and the inability to deploy medium and large-scale neural network autoencoders due to logic resource limitations, which limits the noise suppression capability. The neural network autoencoder deployed in the NPU is used to process the input nuclear pulse signal, while achieving a better noise suppression effect than the traditional digital algorithm, while retaining the original waveform characteristics of the nuclear pulse signal, overcoming the shortcomings of the traditional analog method or digital filtering algorithm that has a general noise suppression capability and causes nuclear pulse signal distortion during the filtering process. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] 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.

[0049] Figure 1 This is a schematic diagram of the structure of a nuclear pulse signal processing architecture based on a neural network autoencoder in Example 1 of the present invention;

[0050] Figure 2 This is a schematic diagram of the working process of the SoC processor in Example 1 of the present invention;

[0051] Figure 3 Schematic diagram of the model structure of the fully connected neural network stacked denoising autoencoder in Example 2 of the present invention;

[0052] Figure 4 Schematic diagram of the model structure of a one-dimensional convolutional neural network autoencoder in Example 2 of the present invention;

[0053] Figure 5 It is a schematic diagram of a nuclear pulse signal superimposed with the first type of noise in Embodiment 2 of the present invention being restored to an ideal signal after being processed by an artificial neural network autoencoder;

[0054] Figure 6 It is a schematic diagram of a nuclear pulse signal superimposed with the second type of noise in Embodiment 2 of the present invention being restored to an ideal signal after being processed by an artificial neural network autoencoder;

[0055] Figure 7 This is a schematic diagram of restoring a nuclear pulse signal superimposed with the third type of noise into an ideal signal after being processed by an artificial neural network autoencoder in Example 2 of the present invention. DETAILED DESCRIPTION

[0056] 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.

[0057] The purpose of the present invention is to provide a nuclear pulse signal processing architecture and method based on a neural network autoencoder, which can improve the filtering and denoising effect of the nuclear pulse signal and reduce the distortion of the nuclear pulse signal after filtering and denoising processing.

[0058] 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.

[0059] Example 1

[0060] like Figure 1 As shown, this embodiment provides a nuclear pulse signal processing architecture based on a neural network autoencoder, including: a nuclear pulse signal preprocessing module, a SoC processor and a ZYNQ programmable system on chip connected in sequence.

[0061] The ZYNQ programmable system on chip is connected to the signal acquisition terminal. The nuclear pulse signal preprocessing module is used to perform filtering, amplification and voltage bias processing on the original signal in sequence. When there is a nuclear pulse signal in the original signal after amplification, the original signal after voltage bias processing is subjected to analog-to-digital conversion processing to obtain a nuclear pulse digital signal to be denoised. The original signal is a baseline signal randomly superimposed with a nuclear pulse signal. The SoC processor integrates a trained neural network autoencoder model. The SoC processor is used to load the nuclear pulse digital signal to be denoised into an input structure variable, input the input structure variable into the trained neural network autoencoder model to obtain a nuclear pulse signal after filtering and denoising, cache the nuclear pulse signal after filtering and denoising in an output structure variable, and transmit the nuclear pulse signal after filtering and denoising in the output structure variable to the ZYNQ programmable system on chip. The trained neural network autoencoder model is obtained by self-supervising the neural network autoencoder model using a noisy nuclear pulse digital signal and a noise-free nuclear pulse digital signal. The neural network autoencoder model is a fully connected neural network or a convolutional neural network. The ZYNQ programmable system-on-chip integrates a variety of digital algorithms. The ZYNQ programmable system-on-chip is used to extract parameter information of the nuclear pulse signal after filtering and noise reduction processing using a variety of digital algorithms, store the parameter information and transmit the parameter information to the signal acquisition terminal. The digital algorithms include shaping algorithm, amplitude extraction algorithm, constant ratio timing digital algorithm and time extraction algorithm.

[0062] Specifically, the nuclear pulse signal preprocessing module includes: a gain adjustment circuit, a DC offset adjustment circuit, a threshold trigger circuit and a high-speed ADC circuit. The gain adjustment circuit is connected to the DC offset adjustment circuit and the threshold trigger circuit respectively. The DC offset adjustment circuit and the threshold trigger circuit are both connected to the high-speed ADC circuit. The high-speed ADC circuit is connected to the SoC processor. The gain adjustment circuit is used to filter and amplify the original signal in sequence to obtain the original signal after amplification. The DC offset adjustment circuit is used to perform voltage bias processing on the original signal after amplification to obtain the original signal after voltage bias processing. The threshold trigger circuit is used to generate a trigger signal according to the original signal after amplification. The threshold trigger circuit is used to output a high level as a trigger signal when the signal amplitude of the original signal after amplification reaches the amplitude threshold, and output a low level as a trigger signal when the signal amplitude of the original signal after amplification is less than the amplitude threshold. The high-speed ADC circuit is used to perform analog-to-digital conversion processing on the original signal after voltage bias processing when the trigger signal is high, so as to obtain the nuclear pulse digital signal to be denoised.

[0063] In addition, the nuclear pulse signal preprocessing module also includes: a radiation detector. The radiation detector is connected to the gain adjustment circuit. The radiation detector is used to obtain the original signal.

[0064] The SoC processor includes: a first CPU processing unit and an NPU processing unit. The first CPU processing unit is respectively connected to the nuclear pulse signal preprocessing module, the NPU processing unit and the ZYNQ programmable system-on-chip; the first CPU processing unit is used to load the nuclear pulse digital signal to be denoised into the input structure variable. The NPU processing unit is integrated with a trained neural network autoencoder model. The NPU processing unit is used to input the input structure variable into the trained neural network autoencoder model to obtain the nuclear pulse signal after filtering and denoising, and cache the nuclear pulse signal after filtering and denoising in the output structure variable. The first CPU processing unit is also used to transmit the nuclear pulse signal after filtering and denoising in the output structure variable to the ZYNQ programmable system-on-chip, and its workflow is as follows: Figure 2 shown.

[0065] The ZYNQ programmable system-on-chip includes: a programmable logic unit and a second CPU processing unit. The programmable logic unit is connected to the SoC processor and the second CPU processing unit respectively. The second CPU processing unit is also connected to the signal acquisition terminal. The programmable logic unit integrates a variety of digital algorithms. The programmable logic unit is used to extract parameter information of the nuclear pulse signal after filtering and noise reduction processing using a variety of digital algorithms. The second CPU processing unit is used to store the parameter information and transmit the parameter information to the signal acquisition terminal.

[0066] The neural network autoencoder in artificial neural network technology can be used to filter and reduce noise of nuclear pulse signals while retaining the original characteristics of the signal. With the development of computer technology, SOC chips with neural network processors (NPUs) have also developed rapidly and are applied in various industries. The emergence and development of chips with this architecture make it possible to deploy neural network autoencoders in embedded systems to achieve online filtering and noise reduction of nuclear pulse signals. Based on this background, a digital nuclear pulse signal processor based on the neural network processor architecture was developed. The use of the neural network autoencoder deployed in the NPU for signal restoration and noise filtering is of great significance to the development of nuclear technology. Figure 1 The nuclear pulse signal processing architecture based on the neural network autoencoder provided in this embodiment includes a SoC processor, a ZYNQ programmable system on chip, a gain adjustment circuit, a threshold trigger circuit, a DC offset adjustment circuit and a high-speed ADC circuit.

[0067] In this system, the nuclear pulse signal obtained by the radiation detector is processed by the gain adjustment circuit and the DC offset adjustment circuit. Under the control of the threshold trigger circuit, the nuclear pulse signal interval is converted into a digital signal through the high-speed ADC circuit and transmitted to the SoC processor. Under the control of the CPU processing unit of the ARM architecture, the nuclear pulse signal data is loaded into the input structure variable of the neural network model of the NPU processing unit. The nuclear pulse signal data in the input structure variable is filtered, denoised, and baseline restored by the artificial neural network autoencoder deployed in the NPU. After being processed by the neural network autoencoder, the noise of the nuclear pulse signal is greatly suppressed, and the signal-to-noise ratio of the signal is improved. At the same time, the signal is basically not distorted, and the subtle features of the original signal are retained. The calculation result of the artificial neural network autoencoder is saved in the output structure variable. The CPU processing unit transmits the nuclear pulse signal data in the output structure variable to the programmable logic unit (PL) of the ZYNQ programmable system-on-chip through parallel communication. A FIFO (First-In, First-Out buffer) is designed in the programmable logic unit of ZYNQ. The buffer is used to cache the nuclear pulse signal data transmitted from the SoC processor. Furthermore, the shaping algorithm, amplitude extraction module, constant ratio timing digitization algorithm, time extraction algorithm, etc. are deployed in the programmable logic unit to extract the parameter information of the nuclear pulse signal. The extracted parameter information is transmitted to the memory variable opened by the CPU processing unit through the AXI (Advanced eXtensible Interface) bus between the programmable logic unit inside the ZYNQ chip and the CPU processing unit of the ARM architecture. Under the control of the CPU, the data in the memory variable is transmitted to the data acquisition terminal through Ethernet communication and the controller. After the operation and processing of the neural network autoencoder deployed in the NPU processing unit in the SoC processor, the noise superimposed on the nuclear pulse signal is greatly suppressed, thereby improving the system energy resolution, and at the same time, the extraction of the nuclear pulse signal characteristics will be more accurate.

[0068] 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.

[0069] Gain adjustment circuit: The gain adjustment circuit filters and amplifies the nuclear pulse signal output from the radiation detector, improves the signal-to-noise ratio of the nuclear pulse signal, and makes the amplitude of the pulse signal as close as possible to the full scale of the subsequent high-speed ADC circuit, thereby improving the measurement accuracy and reducing the quantization error. The output signal of the gain adjustment circuit is divided into two paths, one is transmitted to the threshold trigger circuit, and the other is transmitted to the DC offset adjustment circuit.

[0070] DC offset adjustment circuit: Since the nuclear pulse signal output by the detector is a unipolar signal, and the input end of the high-speed ADC circuit can usually input a bipolar signal, directly inputting the nuclear pulse signal to the high-speed ADC will result in half of the range not being used, thereby wasting the sampling accuracy of the high-speed ADC and increasing the quantization error. Therefore, a DC offset adjustment circuit is required to bias the nuclear pulse signal, and at the same time cooperate with the gain adjustment circuit to make the nuclear pulse signal as close to its range as possible when input to the high-speed ADC circuit to reduce the quantization error. The output signal of the DC offset adjustment circuit is transmitted to the high-speed ADC circuit of the next stage.

[0071] Threshold trigger circuit: The rays emitted by nuclear radiation are random, so the nuclear pulse signal generated by the interaction between the rays and the radiation detector is randomly superimposed on the baseline. Before processing and analyzing the nuclear pulse signal, it is necessary to extract the nuclear pulse signal interval in the original signal output by the radiation detector, and the original baseline without the superimposed nuclear pulse signal is a useless signal and is removed without processing to reduce the workload of the back-end circuit. This function is realized through the threshold trigger circuit. A fixed threshold is set in the threshold trigger circuit. When there is only a baseline signal in the original signal, its signal amplitude is not enough to trigger the threshold of the circuit, so that the threshold trigger circuit outputs a low-level signal. When there is a nuclear pulse signal in the original signal, the nuclear pulse signal triggers the threshold in the circuit, and the threshold trigger circuit outputs a high-level signal of a fixed length of time. The high-level signal is input into the high-speed ADC circuit as a control signal for high-speed ADC sampling.

[0072] High-speed ADC circuit: The main function of the high-speed ADC circuit is to turn on or off ADC sampling according to the trigger signal transmitted by the analog front-end circuit of the previous level, and convert the input analog nuclear pulse signal into a nuclear pulse digital signal. When the trigger signal output by the analog front-end circuit is high, the high-speed ADC circuit starts to perform analog-to-digital conversion on the nuclear pulse signal. Since the trigger signal is a high-level signal with a fixed time length, the analog signal sampled and converted is a nuclear pulse signal of a fixed length, and the number of sampling points of the output digital signal is fixed. When the trigger signal is low, the high-speed ADC circuit stops converting the input signal. Under the control of the trigger signal, the high-speed ADC circuit only performs analog-to-digital conversion on the waveform of the nuclear pulse signal interval, and transmits the waveform data of the nuclear pulse signal to the subsequent circuit, thereby improving the working efficiency of the system.

[0073] SoC processor: The ARM architecture CPU processing unit in the SoC processor mainly controls data transmission, data processing and program operation, while the NPU processing unit mainly deploys an artificial neural network model for more complex data operations and reasoning. A neural network autoencoder for noise reduction filtering is deployed in the NPU processing unit. After the digital nuclear pulse signal converted by the high-speed ADC circuit is transmitted to the SoC processor through parallel communication, under the control of the CPU processing unit, the nuclear pulse signal data is 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. The digital nuclear pulse signal after filtering and noise reduction is stored in the output structure variable of the neural network model. The CPU processing unit obtains the processed digital nuclear pulse signal from the output structure variable, and then transmits the nuclear pulse signal data to the ZYNQ programmable system-on-chip through parallel communication for subsequent extraction of effective information of the nuclear pulse signal.

[0074] The 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, a string of new waveform data is output at the output 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 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. By deploying this model in the NPU, the efficient computing power of the NPU processing unit for the neural network model can be used to realize the function of filtering and denoising the nuclear pulse signal.

[0075] ZYNQ programmable system-on-chip: ZYNQ programmable system-on-chip is mainly divided into CPU processing unit (PS) and programmable logic unit (PL). According to different application scenarios, some effective information of nuclear pulse signals, such as the amplitude of nuclear pulse signals and the duration of pulse front edge, can be extracted through the digital algorithm deployed in the programmable logic unit. After filtering and shaping the nuclear pulse signal using some classic digital shaping algorithms, such as trapezoidal shaping algorithm and Gaussian shaping algorithm, its amplitude information can be extracted; the front edge time information of the nuclear pulse signal can be extracted using the constant ratio timing digital algorithm. These are some common digital algorithms. The FIFO in the programmable logic unit is used to cache the nuclear pulse signal data transmitted from the SoC processor. The digital algorithm in the programmable logic unit reads the nuclear pulse signal data from the FIFO for calculation to realize parameter extraction. The extracted nuclear pulse signal parameter data is transmitted to the memory space opened up by the CPU processing unit through the AXI bus, and then the data is sent to the data acquisition terminal for analysis through Ethernet communication and the controller.

[0076] The connection relationship, signal flow and collaborative working process of the nuclear pulse signal processor are as follows:

[0077] The radiation rays deposit energy in the radiation detector, and the ray energy is converted into a voltage pulse signal and output from the radiation detector. It enters the gain adjustment circuit through the coaxial line to amplify the signal. The amplified signal is then divided into two paths. One path is transmitted to the threshold trigger circuit through the PCB trace, and the corresponding threshold trigger signal is obtained according to the nuclear pulse signal. The other path is transmitted to the DC offset adjustment circuit, and the nuclear pulse signal is converted into a bipolar signal by means of DC offset, so that the nuclear pulse signal input to the high-speed ADC circuit can be as close to the full scale of the ADC as possible, thereby reducing the quantization error. Under the control of the threshold trigger signal, the high-speed ADC performs analog-to-digital conversion on the nuclear pulse signal transmitted from the output end of the DC offset circuit, converts it into a digital signal and transmits it to the SoC processor through the signal line of the PCB board. Under the control of the CPU processing unit of the ARM architecture, the digital nuclear pulse signal is stored in the input structure variable of the neural network autoencoder deployed in the NPU processing unit. After the operation and processing of the artificial neural network autoencoder, the digital nuclear pulse signal is filtered and denoised. After noise suppression, the nuclear pulse signal with a high signal-to-noise ratio is stored in the output structure variable. The CPU processing unit transmits the valid data of the output structure variable to the ZYNQ programmable chip system through parallel communication, and caches it in the FIFO of the programmable logic unit. The digital algorithms deployed in the programmable logic unit, such as the shaping algorithm and the constant ratio timing digital algorithm, read the nuclear pulse signal data from the FIFO to extract the nuclear pulse signal features, and then transmit the extracted feature parameter data to the memory space opened up by the CPU processing unit through the AXI bus. Then, under the control of the CPU processing unit, the data is sent to the data acquisition terminal for analysis through Ethernet communication and the controller.

[0078] Example 2

[0079] This embodiment provides a nuclear pulse signal processing method based on a neural network autoencoder, comprising:

[0080] Step 201: Acquire an original signal. The original signal is a baseline signal randomly superimposed with a nuclear pulse signal.

[0081] Step 202: performing filtering processing, amplification processing and voltage bias processing on the original signal in sequence.

[0082] Step 203: When there is a nuclear pulse signal in the original signal after the amplification processing, the original signal after the voltage bias processing is subjected to analog-to-digital conversion processing to obtain a nuclear pulse digital signal to be subjected to noise reduction.

[0083] Step 204: Load the core pulse digital signal to be denoised into the input structure variable.

[0084] Step 205: Input the input structure variable into the trained neural network autoencoder model to obtain the output structure variable. The output structure variable includes the kernel pulse signal after filtering and noise reduction. The trained neural network autoencoder model is obtained by self-supervising the neural network autoencoder model using the noisy kernel pulse digital signal and the noise-free kernel pulse digital signal.

[0085] Step 206: Utilize a variety of digital algorithms to extract parameter information of the nuclear pulse signal after filtering and noise reduction processing.

[0086] Step 207: Store and transmit parameter information.

[0087] Before step 201, the method further includes:

[0088] Step 208: Construct a neural network autoencoder model. The neural network autoencoder model is a fully connected neural network or a convolutional neural network.

[0089] Step 207: Input the noisy core pulse digital signal into the neural network autoencoder model to obtain a denoised core pulse digital signal.

[0090] Step 2010: Determine the loss function value according to the denoised core pulse digital signal and the noise-free core pulse digital signal.

[0091] Step 2011: According to the loss function value, adjust the parameters of the neural network autoencoder model and return to step 207 until the loss function values ​​of the preset number of consecutive iterations are all less than the loss function value threshold, thereby obtaining the trained neural network autoencoder model.

[0092] Specifically, the neural network autoencoder model includes an input layer, a plurality of encoding layers, a plurality of decoding layers and an output layer connected in sequence. The plurality of encoding layers includes a plurality of encoding layers connected in sequence. The plurality of decoding layers includes a plurality of decoding layers connected in sequence. The number of encoding layers is equal to the number of decoding layers.

[0093] Nuclear pulse signals are random and have short waveform duration. System noise can cause false triggering of nuclear pulse signals. When noise is superimposed on nuclear pulse signals, it will also affect the subsequent extraction of waveform feature information. The artificial neural network autoencoder learns the characteristics of the nuclear pulse signal waveform during the training process, retains the waveform details of the nuclear pulse signal when applied, and suppresses noise at the same time.

[0094] The principle of noise suppression lies in the training process of the autoencoder. During the training process, the autoencoder optimizes its own parameters by minimizing the reconstruction error (that is, the difference between the reconstructed output and the original input). When the input data contains noise, the reconstruction error increases, and the autoencoder will try to reduce this error by adjusting the weight parameters.

[0095] 1. The structure of the neural network autoencoder.

[0096] The denoising autoencoder is a two-layer neural network consisting of an encoder and a decoder. Multiple denoising autoencoder layers are stacked together to form a deeper model. The output of each layer of the encoder is used as the input of the next layer, thereby gradually reducing the dimension of the data. Through multi-layer stacking and training, the stacked denoising autoencoder can learn the multi-level feature expression of the data and capture higher-level abstract features. Nuclear pulse signal data can be regarded as one-dimensional sequence data. In addition to using a fully connected neural network to construct a stacked denoising autoencoder (such as Figure 3 As shown), you can also use convolutional neural networks to build (as shown Figure 4 shown).

[0097] 2. The principle of filtering and denoising nuclear pulse signals by neural network autoencoder.

[0098] like Figure 5-Figure 7 As shown in the figure, the input core pulse signal is composed of an ideal signal and noise superimposed on the ideal signal. The goal of the autoencoder is to obtain the target signal after processing the input core pulse 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.

[0099] There are many types of noise, such as white noise, low-frequency noise, and mutation noise. The nuclear pulse signal superimposed with these types of noise needs to be used as a training set. The autoencoder is trained and its weight parameters are adjusted according to the training set to identify and suppress these types of noise. Therefore, during the training process, as many types of noise as possible should be superimposed on the input nuclear pulse signal, so that the autoencoder after training has stronger robustness and is more adaptable to noise suppression in complex environments.

[0100] During the training process, the goal of the denoising autoencoder is to minimize the reconstruction error between the input data and the reconstructed data. To achieve this goal, the autoencoder needs to learn the features and structures in the kernel pulse signal to reconstruct the original signal as accurately as possible; when the input data contains noise, the reconstruction error will increase, so the autoencoder is also forced to learn strategies to compensate or filter out the noise during the training process, and reduce this error by adjusting the weight parameters. The end-to-end model training is performed through the back-propagation algorithm, and the appropriate optimization algorithm is used to optimize the entire model.

[0101] by Figure 5-Figure 7 Taking the autoencoder structure shown in the figure as an example, when the input is x, the encoder compresses the data in the following way:

[0102] h=σ(Wx+b0).

[0103] Among them, h is the hidden layer output, W is the weight matrix, b0 is the bias, and σ is the activation function.

[0104] The decoder reconstructs the data as follows:

[0105] X=σ(W T h+c0).

[0106] Where X is the decoder output, W T is the transposed matrix of W, and c0 is the bias.

[0107] Taking the mean square error as the loss function, the calculation formula of the loss function L is:

[0108]

[0109] The reverse gradient descent algorithm is used to continuously optimize W, b0, and c0 to minimize the loss function. In the training process of the autoencoder, the influence of signal and noise on the weights is complementary. By introducing noise, the autoencoder can learn strategies for suppressing and filtering noise; and by reconstructing the signal, the autoencoder can learn strategies for retaining and restoring the signal waveform. These two aspects interact with each other and jointly affect the adjustment process of the autoencoder network weights, enabling it to better learn the key features of the input data while filtering out the interference of noise.

[0110] 3. Deployment of neural network autoencoder and online filtering and denoising of nuclear pulse signals.

[0111] The deployment of the neural network autoencoder is to convert the trained artificial neural network model into a model file format that can be recognized and loaded by the 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 autoencoder to calculate and analyze the real-time input nuclear pulse signal data, thereby realizing online filtering and denoising of the detected nuclear pulse data.

[0112] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0113] The principles and implementation methods of the present invention are described herein using specific examples. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those of ordinary skill 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 nuclear pulse signal processing architecture based on a neural network autoencoder, characterized in that: include: A nuclear pulse signal preprocessing module, a SoC processor, and a ZYNQ programmable system-on-chip connected in sequence; The ZYNQ programmable system on chip is connected to a signal acquisition terminal; The nuclear pulse signal preprocessing module is used to perform filtering processing, amplification processing and voltage bias processing on the original signal in sequence. When the nuclear pulse signal exists in the original signal after the amplification processing, the original signal after the voltage bias processing is subjected to analog-to-digital conversion processing to obtain the nuclear pulse digital signal to be denoised; the original signal is a baseline signal randomly superimposed with the nuclear pulse signal; The SoC processor is integrated with a trained neural network autoencoder model; the SoC processor is used to load the core pulse digital signal to be denoised into an input structure variable, input the input structure variable into the trained neural network autoencoder model to obtain the core pulse signal after filtering and denoising, cache the core pulse signal after filtering and denoising in an output structure variable, and transmit the core pulse signal after filtering and denoising in the output structure variable to the ZYNQ programmable system on chip; the trained neural network autoencoder model is obtained by self-supervising the neural network autoencoder model using the noisy core pulse digital signal and the noise-free core pulse digital signal; The SoC processor includes: a first CPU processing unit and an NPU processing unit; The first CPU processing unit is respectively connected to the nuclear pulse signal preprocessing module, the NPU processing unit and the ZYNQ programmable chip system; the first CPU processing unit is used to load the nuclear pulse digital signal to be denoised into an input structure variable; the NPU processing unit is integrated with a trained neural network autoencoder model; the NPU processing unit is used to input the input structure variable into the trained neural network autoencoder model to obtain the nuclear pulse signal after filtering and denoising, and cache the nuclear pulse signal after filtering and denoising in the output structure variable; the first CPU processing unit is also used to transmit the nuclear pulse signal after filtering and denoising in the output structure variable to the ZYNQ programmable chip system; The ZYNQ programmable system-on-chip integrates a variety of digital algorithms; the ZYNQ programmable system-on-chip is used to use the multiple digital algorithms to extract parameter information of the nuclear pulse signal after filtering and noise reduction processing, store the parameter information and transmit the parameter information to the signal acquisition terminal.

2. A nuclear pulse signal processing architecture based on a neural network autoencoder according to claim 1, characterized in that: The digitization algorithm includes a shaping algorithm, an amplitude extraction algorithm, a constant ratio timing digitization algorithm and a time extraction algorithm.

3. The nuclear pulse signal processing architecture based on a neural network autoencoder according to claim 1, characterized in that: The nuclear pulse signal preprocessing module includes: a gain adjustment circuit, a DC offset adjustment circuit, a threshold trigger circuit and a high-speed ADC circuit; The gain adjustment circuit is connected to the DC offset adjustment circuit and the threshold trigger circuit respectively; the DC offset adjustment circuit and the threshold trigger circuit are both connected to the high-speed ADC circuit; the high-speed ADC circuit is connected to the SoC processor; The gain adjustment circuit is used to perform filtering and amplification processing on the original signal in sequence to obtain the original signal after amplification processing; The DC offset adjustment circuit is used to perform voltage bias processing on the original signal after amplification processing to obtain the original signal after voltage bias processing; The threshold trigger circuit is used to generate a trigger signal according to the original signal after amplification processing; the threshold trigger circuit is used to output a high level as a trigger signal when the signal amplitude of the original signal after amplification processing reaches the amplitude threshold, and output a low level as a trigger signal when the signal amplitude of the original signal after amplification processing is less than the amplitude threshold; The high-speed ADC circuit is used to perform analog-to-digital conversion processing on the original signal after voltage bias processing when the trigger signal is at a high level, so as to obtain the core pulse digital signal to be de-noised.

4. A nuclear pulse signal processing architecture based on a neural network autoencoder according to claim 3, characterized in that: The nuclear pulse signal preprocessing module also includes: a radiation detector; The radiation detector is connected to the gain adjustment circuit; the radiation detector is used to obtain the original signal.

5. The nuclear pulse signal processing architecture based on a neural network autoencoder according to claim 2, characterized in that: The ZYNQ programmable system-on-chip includes: a programmable logic unit and a second CPU processing unit; The programmable logic unit is connected to the SoC processor and the second CPU processing unit respectively; the second CPU processing unit is also connected to the signal acquisition terminal; The programmable logic unit integrates a plurality of digital algorithms; the programmable logic unit is used to extract parameter information of the nuclear pulse signal after filtering and noise reduction processing by using the plurality of digital algorithms; The second CPU processing unit is used to store the parameter information and transmit the parameter information to the signal acquisition terminal.

6. A nuclear pulse signal processing architecture based on a neural network autoencoder according to claim 2, characterized in that: The neural network autoencoder model is a fully connected neural network or a convolutional neural network.

7. A method for processing nuclear pulse signals based on a neural network autoencoder, characterized in that: The nuclear pulse signal processing method based on the neural network autoencoder is applied to the nuclear pulse signal processing architecture based on the neural network autoencoder as claimed in any one of claims 1 to 6, and the nuclear pulse signal processing method based on the neural network autoencoder includes: Acquire an original signal; the original signal is a baseline signal randomly superimposed with a nuclear pulse signal; The original signal is subjected to filtering, amplification and voltage bias processing in sequence; When there is a nuclear pulse signal in the original signal after the amplification process, the original signal after the voltage bias process is subjected to analog-to-digital conversion to obtain a nuclear pulse digital signal to be subjected to noise reduction; Loading the core pulse digital signal to be denoised into an input structure variable; Inputting the input structure variable into the trained neural network autoencoder model to obtain an output structure variable; the output structure variable includes the core pulse signal after filtering and noise reduction processing; the trained neural network autoencoder model is obtained by self-supervising the neural network autoencoder model using the noisy core pulse digital signal and the noise-free core pulse digital signal; Utilize a variety of digital algorithms to extract parameter information of nuclear pulse signals after filtering and noise reduction; The parameter information is stored and transmitted.

8. The method for processing nuclear pulse signals based on a neural network autoencoder according to claim 7, characterized in that: Before obtaining the original signal, the method further includes: Constructing a neural network autoencoder model; the neural network autoencoder model is a fully connected neural network or a convolutional neural network; The noisy kernel pulse digital signal is input into the neural network autoencoder model to obtain the denoised kernel pulse digital signal; Determine the loss function value according to the denoised core pulse digital signal and the noise-free core pulse digital signal; According to the loss function value, the parameters of the neural network autoencoder model are adjusted and the step of "inputting the noisy core pulse digital signal into the neural network autoencoder model to obtain a denoised core pulse digital signal" is returned until the loss function values ​​of the preset number of consecutive iterations are less than the loss function value threshold, thereby obtaining the trained neural network autoencoder model.

9. The method for processing nuclear pulse signals based on a neural network autoencoder according to claim 8, characterized in that: The neural network autoencoder model includes an input layer, a multi-encoding layer, a multi-decoding layer and an output layer connected in sequence; The multiple coding layers include a plurality of coding layers connected in sequence; The multiple decoding layers include a plurality of decoding layers connected in sequence; The number of the encoding layers is equal to the number of the decoding layers.

Images