Neutron gamma discrimination method and system based on pulse shape self-calibration

Abstract

The application discloses a neutron-gamma discrimination method and system based on pulse shape self-calibration, and relates to the technical field of nuclear radiation detection. The method comprises the following steps: collecting and pre-processing digital pulse waveforms output by a detector, extracting multi-dimensional pulse shape features and combining the features with effective digital pulse waveforms in time sequence; screening high-confidence neutron pulse samples and high-confidence gamma pulse samples based on current environmental state data, updating reference pulse templates and feature normalization parameters; dynamically correcting a neutron-gamma discrimination boundary by using the update result and outputting a discrimination result; and calculating discrimination performance quality according to the discrimination result, and triggering the next round of update when a preset decline condition is met. The application can make the discrimination boundary self-calibrate with the change of environmental state and detector response, and reduce the influence of fixed parameter drift on the discrimination result.

Description

Technical Field

[0001] This invention relates to the field of nuclear radiation detection technology, and in particular to a neutron gamma discrimination method and system based on pulse shape self-calibration. Background Technology

[0002] In the field of nuclear radiation detection, organic scintillator detectors are widely used for neutron measurements in mixed radiation fields. Since neutrons are usually accompanied by gamma rays, distinguishing between the two is a prerequisite for obtaining reliable data. Currently, pulse shape discrimination technology is a common method for differentiation, primarily utilizing the differences in fluorescence decay times excited by different particles in the scintillator for classification.

[0003] In practical applications, traditional discrimination methods typically rely on a fixed integration time window or discrimination threshold. However, during long-term operation, detectors are susceptible to interference from factors such as ambient temperature fluctuations, photomultiplier tube (PMT) aging, and voltage drift, leading to pulse signal baseline drift and gain fluctuations. Under these disturbances, fixed discrimination parameters cannot be adaptively adjusted, causing overlap in the characteristic distributions of neutrons and gamma rays, thus reducing discrimination accuracy.

[0004] To address this issue, existing technologies have introduced some model-based classification schemes. However, these methods largely rely on offline-collected static datasets for training, resulting in fixed classification boundaries. When the actual operating environment of the detector changes, the distribution of pulse features acquired in real-time deviates from the original model, leading to a decline in classification performance. To maintain discrimination quality, system operation needs to be interrupted and samples manually reacquired to calibrate parameters. This approach is labor-intensive, inefficient, and fails to meet the requirements for real-time, continuous monitoring of radiation fields in complex environments. Therefore, a technical solution that can adapt to environmental changes and autonomously manage discrimination boundaries is urgently needed. Summary of the Invention

[0005] This invention provides a neutron gamma discrimination method and system based on pulse shape self-calibration, which solves the problems of existing technologies where fixed parameters cannot adapt to environmental drift and rely on manual calibration.

[0006] The first aspect of this invention proposes a neutron gamma discrimination method based on pulse shape self-calibration, comprising: The digital pulse waveform output by the detector is acquired and preprocessed to obtain an effective digital pulse waveform. Feature extraction is performed on the effective digital pulse waveform to obtain multi-dimensional pulse shape features. The effective digital pulse waveform and its corresponding multi-dimensional pulse shape features are then combined in a time sequence to form continuously acquired data. The current environmental state data of the detector is acquired. Based on the current environmental state data, high-confidence neutron pulse samples and high-confidence gamma pulse samples are selected from the continuously collected data. The reference pulse template and feature normalization parameters are updated in real time using the high-confidence neutron pulse samples and high-confidence gamma pulse samples. The neutron gamma discrimination boundary is dynamically corrected using the updated reference pulse template and feature normalization parameters to obtain the corrected neutron gamma discrimination boundary. The multi-dimensional pulse shape features are input into the corrected neutron gamma discrimination boundary, and the corresponding neutron gamma discrimination results are output. The discrimination performance quality is evaluated based on the neutron gamma discrimination results. When the discrimination performance quality meets the preset descent condition, the next round of updating of the reference pulse template and feature normalization parameters is triggered.

[0007] In one optional implementation, the digital pulse waveform is preprocessed to obtain an effective digital pulse waveform, including: Baseline estimation and baseline subtraction are performed on the digital pulse waveform to determine the zero-level reference. Peak location results are obtained by performing a peak search on the digital pulse waveform after baseline subtraction based on a zero-level reference. The peak amplitude of the pulse is obtained based on the peak positioning result. The peak amplitude of the pulse is then used to normalize the amplitude of the digital pulse waveform after the baseline is removed, resulting in a normalized digital pulse waveform. Based on the preset noise threshold and pulse rise time validity criteria, the normalized digital pulse waveform is screened for pulse validity to obtain a valid digital pulse waveform.

[0008] In one optional implementation, feature extraction is performed on the effective digital pulse waveform to obtain multi-dimensional pulse shape features, including: The rise time, decay time, tail integration ratio, peak amplitude, and total integral of the effective digital pulse waveform are extracted to form time-domain statistical features. The rising edge, peak value and decay tail of the effective digital pulse waveform are decomposed into multiple scale components using the Haar wavelet basis function, and the wavelet approximation coefficients and detail coefficients of each scale component are extracted to form wavelet multi-scale features. The effective digital pulse waveform in one-dimensional form is converted into a two-dimensional pulse image through a time-series matrix mapping. The two-dimensional pulse image is then input into a pre-trained two-dimensional convolutional neural network for convolution extraction, and the output vector of the hidden layer is obtained as the depth feature of the pulse image. By combining time-domain statistical features, wavelet multi-scale features, and pulse image depth features, multi-dimensional pulse shape features are constructed.

[0009] In one optional implementation, the current environmental state data includes temperature data of the detector's environment, detector gain data, baseline noise data, and pulse amplitude range data. The current environmental state data of the detector is acquired, and based on this data, high-confidence neutron pulse samples and high-confidence gamma pulse samples are selected from continuously acquired data. The reference pulse template and feature normalization parameters are then updated in real time using these high-confidence neutron pulse samples and high-confidence gamma pulse samples, including: Real-time monitoring of current environmental status data; when the change in current environmental status data compared to preset historical baseline data exceeds a preset environmental drift threshold, a self-calibration update process is initiated. After the self-calibration update process is started, the initial neutron cluster center and the initial gamma cluster center are determined by clustering algorithm in the continuously collected data, and the data whose distance from the initial neutron cluster center and the initial gamma cluster center is within the preset safety radius are extracted and used as high-confidence neutron pulse samples and high-confidence gamma pulse samples, respectively. The mean waveform and globally optimized decay shape parameters of high-confidence neutron pulse samples and high-confidence gamma pulse samples are calculated respectively, and an updated reference pulse template is generated based on the mean waveform and globally optimized decay shape parameters. The feature distribution centers of high-confidence neutron pulse samples and high-confidence gamma pulse samples are calculated respectively. The feature compensation scale factor is calculated based on the feature distribution centers and used as the updated feature normalization parameter.

[0010] In one alternative implementation, the method further includes, during the process of generating the updated reference pulse template based on the mean waveform and globally optimized attenuation shape parameters: Determine whether the pulse amplitude range data in the current environmental state data falls within the nonlinear response region of the photomultiplier tube contained in the detector; If the sample falls into the nonlinear response region, the preset artificial neural network nonlinear response calibration model is invoked to perform high-dimensional spatial mapping distortion correction on the high-confidence neutron pulse sample and high-confidence gamma pulse sample in the nonlinear response region, and obtain nonlinear response correction sample data. By replacing the corresponding high-confidence neutron pulse samples and high-confidence gamma pulse samples within the nonlinear response region with nonlinear response correction sample data, the mean waveform and globally optimized attenuation shape parameters are calculated to correct the boundary bending error caused by the nonlinear response of the photomultiplier tube.

[0011] In one alternative implementation, the neutron gamma discrimination boundary is dynamically corrected using the updated reference pulse template and feature normalization parameters to obtain the corrected neutron gamma discrimination boundary, including: The extracted temporal statistical features and wavelet multi-scale features are compensated for feature drift using feature normalization parameters to obtain a normalized feature set. Using the normalized feature set as the input sample set and the updated reference pulse template, the support vector machine algorithm is used to recalculate the high-dimensional hyperplane weight matrix used to distinguish between neutron and gamma categories. The updated high-dimensional hyperplane generated based on the high-dimensional hyperplane weight matrix is ​​then determined as the corrected neutron-gamma discrimination boundary.

[0012] In one optional implementation, the multi-dimensional pulse shape features are input into the corrected neutron gamma discrimination boundary, and the corresponding neutron gamma discrimination results are output, including: The corresponding time-domain statistical features and wavelet multi-scale features in the normalized feature set are input into the multi-dimensional spatial projection decision classifier for classification and determination, and the confidence probability of the first category is obtained. The pulse image depth features are input into the terminal fully connected layer of a two-dimensional convolutional neural network for activation judgment to obtain the second category confidence probability; A decision fusion layer with a modified neutron-gamma discrimination boundary is adopted. The joint decision probability is obtained by weighted cross-entropy fusion of the confidence probability of the first category and the confidence probability of the second category. When the joint decision probability is greater than the preset boundary threshold, the neutron category label is output; otherwise, the gamma ray category label is output, and the neutron-gamma discrimination result composed of the corresponding labels is obtained.

[0013] In one optional implementation, a decision fusion layer with a modified neutron-gamma discrimination boundary is used. The first-class confidence probability and the second-class confidence probability are weighted and cross-entropy fused to obtain a joint decision probability. When the joint decision probability is greater than a preset boundary threshold, a neutron class label is output; otherwise, a gamma-ray class label is output. The neutron-gamma discrimination result, composed of the corresponding labels, is obtained, including: Obtain the real-time event count rate of the detector; When the real-time event count rate exceeds the preset congestion threshold, the preprocessed effective digital pulse waveform is converted into a signal vector form. The inner product of the effective digital pulse waveform in signal vector form and the updated reference pulse template is calculated to obtain the waveform similarity scalar used to characterize the waveform similarity. The waveform similarity scalar is used to replace the first category confidence probability obtained by classification and determination of time-domain statistical features and wavelet multi-scale features. The second category confidence probability is combined and input into the decision fusion layer for downgraded cross-entropy fusion determination, and the neutron gamma discrimination result is output.

[0014] In one optional implementation, the discrimination performance quality is evaluated based on the neutron gamma discrimination results. When the discrimination performance quality meets a preset descent condition, the next round of updates to the reference pulse template and feature normalization parameters is triggered, including: The neutron-gamma discrimination results of all valid digital pulse waveforms within a preset time window are statistically analyzed, and the data are divided into neutron data clusters and gamma data clusters based on the neutron-gamma discrimination results. Calculate the absolute distance between the feature distribution centers of the neutron data cluster and the gamma data cluster, and the sum of the full width at half maximum (FWHM) of the neutron data cluster and the full width at half maximum (FWHM) of the gamma data cluster, respectively. The ratio of absolute distance to the sum of half-height and full width is determined as the discrimination quality factor, and the discrimination quality factor is used as the discrimination performance quality. The system obtains the quality degradation rate of the identification quality factor compared to the pre-stored historical reference quality factor. When the quality degradation rate is greater than or equal to the warning threshold, it determines that the preset degradation condition is met and automatically sends a feedback signal to trigger the next round of updates to the reference pulse template and feature normalization parameters.

[0015] A second aspect of this invention proposes a neutron gamma discrimination system based on pulse shape self-calibration, comprising: The pulse acquisition and preprocessing module is used to acquire the digital pulse waveform output by the detector and preprocess the digital pulse waveform to obtain an effective digital pulse waveform. The feature extraction and data construction module is used to extract features from the effective digital pulse waveform to obtain multi-dimensional pulse shape features, and to combine the effective digital pulse waveform and its corresponding multi-dimensional pulse shape features in a time sequence to form continuously acquired data. The self-calibration sample screening and parameter update module is used to acquire the current environmental state data of the detector, and based on the current environmental state data, it selects high-confidence neutron pulse samples and high-confidence gamma pulse samples from the continuously acquired data, and uses the high-confidence neutron pulse samples and high-confidence gamma pulse samples to update the reference pulse template and feature normalization parameters in real time. The dynamic correction module for the discrimination boundary is used to dynamically correct the neutron gamma discrimination boundary using the updated reference pulse template and feature normalization parameters, so as to obtain the corrected neutron gamma discrimination boundary. The comprehensive judgment and quality assessment feedback module is used to input multi-dimensional pulse shape features into the corrected neutron gamma discrimination boundary, output the corresponding neutron gamma discrimination result, evaluate the discrimination performance quality based on the neutron gamma discrimination result, and trigger the next round of update of reference pulse template and feature normalization parameters when the discrimination performance quality meets the preset decline condition.

[0016] A third aspect of this invention provides an electronic device, comprising: a memory and a processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the neutron gamma discrimination method based on pulse shape self-calibration as proposed in the foregoing embodiments.

[0017] The fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the neutron gamma discrimination method based on pulse shape self-calibration as proposed in the foregoing embodiments.

[0018] Beneficial Effects: This invention provides a neutron-gamma discrimination method and system based on pulse shape self-calibration. By combining the detector's current environmental state data with continuously acquired pulse data, it achieves autonomous screening of high-confidence samples and real-time updating of the reference template, overcoming the limitations of existing technologies such as reliance on fixed parameters, weak resistance to environmental interference, and the need for manual calibration. By extracting multi-dimensional pulse shape features and combining them temporally, this method can dynamically correct the neutron-gamma discrimination boundary according to the real-time environmental state during operation. This allows the system to adapt to baseline and gain drift caused by temperature fluctuations or detector aging, maintaining the timeliness and reliability of the discrimination model. Furthermore, this invention introduces a performance quality evaluation and closed-loop feedback mechanism based on classification results. When the discrimination quality meets the set decline conditions, the system autonomously triggers the next round of parameter updates. This process reduces reliance on offline static training datasets, lowers the daily maintenance threshold of the system, and ensures that the detection system can maintain a long-term, stable discrimination state in dynamic working environments. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 A flowchart illustrating the steps of a neutron gamma discrimination method based on pulse shape self-calibration provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the functional modules of a neutron gamma discrimination system based on pulse shape self-calibration provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] The present invention will be further described below with reference to the accompanying drawings.

[0022] Reference Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0023] like Figure 1 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (Wi-Fi). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0024] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0025] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and computer programs.

[0026] exist Figure 1 In the illustrated electronic device, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the neutron gamma discrimination system based on pulse shape self-calibration stored in the memory 1005 through the processor 1001 and executes the neutron gamma discrimination method based on pulse shape self-calibration provided in the embodiment of the present invention.

[0027] Combined with appendix Figure 2As shown, the first aspect of this invention proposes a neutron-gamma discrimination method based on pulse shape self-calibration. This method is mainly applied in mixed radiation fields (such as around nuclear reactors, nuclide identification instruments, or deep-earth exploration environments), utilizing organic scintillator detectors (such as liquid scintillator BC501A, NE213, or solid scintillator Ce:LiCaAlF6, etc.) to accurately distinguish incident neutrons and gamma rays. A complete execution flow of this method is provided, specifically including the following steps S101 to S105: S101 acquires the digital pulse waveform output by the detector and preprocesses the digital pulse waveform to obtain the effective digital pulse waveform.

[0028] During nuclear radiation detection, when neutrons or gamma rays emitted from a radioactive source enter an organic scintillator detector (PMT), they interact with the target nucleus within the scintillator (neutrons typically undergo elastic scattering to produce recoil protons, while gamma rays generate recoil electrons through Compton scattering). These recoil particles deposit energy in the medium, exciting fluorescent molecules, which then de-excite and emit photons. The cathode of the PMT receives and converts these photons into photoelectrons, which are then multiplied by a dynamo and output as analog electrical pulse signals at the anode. These analog signals are then sampled at high speed using a high-frequency analog-to-digital converter (ADC) (typically at sampling rates of 500 MS / s to 2 GS / s) to obtain the digital pulse waveform output by the detector.

[0029] In one alternative implementation, the step of preprocessing the digital pulse waveform to obtain an effective digital pulse waveform in order to eliminate the interference of electronic noise, dark current, and baseline drift on subsequent pulse shape analysis further includes sub-steps S1011 to S1014: S1011. Perform baseline estimation and baseline subtraction on the digital pulse waveform to determine the zero-level reference.

[0030] Due to the AC coupling characteristics of the detector's front-end electronics, temperature drift, and the effects of pulse accumulation tails at high count rates, the baseline of the digital pulse waveform is not constant at absolute zero but rather experiences dynamic drift. The actuator first extracts background data from a pulse-free region before the trigger signal arrives (i.e., within the pre-trigger time window, such as the sequence of 50 sampling points before the pulse rising edge). The expected value of this background data is calculated or a polynomial fit is used to obtain the dynamic baseline value. Subsequently, this dynamic baseline value is subtracted from the value of each sampling point in the entire digital pulse waveform, thus completing baseline subtraction and shifting the signal to a true zero-level reference, ensuring the accuracy of the integrated charge calculation.

[0031] S1012. Based on the zero-level reference, perform peak search on the digital pulse waveform after baseline subtraction to obtain peak location results.

[0032] After establishing the zero-level reference, the execution unit uses a first-order differential or sliding window extremum determination algorithm to traverse and search the pulse data sequence. To prevent false peak misjudgment caused by high-frequency quantization noise, a slight five-point moving average smoothing can be performed before peak search. When the derivative of the signal is detected to change from positive to negative, and the corresponding signal amplitude is greater than the preset hardware trigger threshold, the time index where it is located is locked, thereby obtaining an accurate peak location result (i.e., the time point when the pulse reaches its maximum amplitude).

[0033] S1013. Obtain the pulse peak amplitude based on the peak positioning result, and use the pulse peak amplitude to perform amplitude normalization processing on the digital pulse waveform after deducting the baseline, so as to obtain the normalized digital pulse waveform.

[0034] Incident particles of different energies deposit different energies, resulting in significant differences in pulse amplitude. The core of PSD technology lies in the waveform shape, not the amplitude. Therefore, amplitude normalization is necessary, and its mathematical expression is shown in the formula:

[0035] In the formula, Indicates the time after normalization. The digital pulse sample value; This represents the raw digital pulse sample value acquired by the ADC; This represents the dynamic baseline value estimated in step S1011; This represents the absolute value of the pulse peak amplitude obtained based on the peak localization result. Through this operation, the peak values ​​of all digital pulse waveforms are uniformly scaled to a constant of 1, eliminating the absolute interference of particle energy differences on shape feature extraction.

[0036] S1014. Based on the preset noise threshold and pulse rise time validity discrimination conditions, the normalized digital pulse waveform is subjected to pulse validity screening to obtain a valid digital pulse waveform.

[0037] In actual detection, the environment contains overload pulses generated by cosmic rays, single-photon dark noise from PMT thermionic emission, and severe pile-up pulses caused by simultaneous incidence of multiple particles. The actuator calculates the time taken for the normalized waveform to rise from 10% amplitude to 90% amplitude. If this rise time is lower than the detector's inherent physical emission limit (considered as high-frequency glitches), or higher than the upper threshold (considered as cosmic rays or multiple overlapping pulses), the pulse is marked as invalid and discarded. Simultaneously, multi-peak detection is performed on the pulse tail; if a secondary spike is present, it is determined to be a pile-up abnormality pulse and discarded. The data retained after the above screening is considered the valid digital pulse waveform.

[0038] S102, extract features from the effective digital pulse waveform to obtain multi-dimensional pulse shape features, and combine the effective digital pulse waveform and its corresponding multi-dimensional pulse shape features in a time sequence to form continuously acquired data.

[0039] Traditional charge integration methods often rely on only a single one-dimensional feature, which leads to a sharp decline in discrimination performance in the low-energy region due to increased photon statistical fluctuations. This embodiment, however, integrates multi-dimensional features from the time domain, frequency domain, and spatial depth to comprehensively characterize the pulse shape.

[0040] In one optional implementation, feature extraction is performed on the effective digital pulse waveform to obtain multi-dimensional pulse shape features, including sub-steps S1021 to S1024: S1021. Extract the rise time, decay time, tail integration ratio, peak amplitude, and total integral of the effective digital pulse waveform to form time-domain statistical features.

[0041] According to the luminescence dynamics of scintillators, the recoil protons generated by neutron excitation have a high ionization density, leading to a large amount of "delayed fluorescence (slow component)" due to the de-excitation of scintillator molecules; while the recoil electrons generated by gamma-ray excitation have a low ionization density, mainly producing "instantaneous fluorescence (fast component)". The execution entity sets a short integration time window (covering the fast component region) and a long integration time window (covering the entire slow component region at the pulse tail), calculating the ratio of the integrated charge at the tail to the total integrated charge (i.e., the tail integration ratio). Simultaneously, the time constant of the pulse's exponential decay segment is recorded as the decay time. These physical parameters together constitute the temporal statistical characteristics characterizing the differences in the luminescence mechanisms.

[0042] S1022. Using the Haar wavelet basis function, the rising edge, peak value and decay tail of the effective digital pulse waveform are decomposed into multiple scale components, and the wavelet approximation coefficients and detail coefficients of each scale component are extracted to form wavelet multi-scale features.

[0043] Digital methods offer superior feature resolution compared to traditional analog circuit methods. By introducing the Haar wavelet transform, the actuator performs discrete wavelet decomposition on a one-dimensional time-series signal. The Haar wavelet, as the simplest orthogonal wavelet, possesses low computational cost and sensitivity to signal abrupt changes, making it ideal for processing flicker pulses. Through multi-scale decomposition, the waveform is broken down into "wavelet approximation coefficients," which contain the overall low-frequency attenuation envelope distribution of the signal, and "wavelet detail coefficients," which contain high-frequency statistical fluctuations and white noise. Extracting the approximation coefficients at specific scales can effectively filter out photoelectron statistical fluctuation noise in the PMT, while the energy distribution patterns of the detail coefficients also contain microscopic statistical differences between different particles; together, they constitute the multi-scale wavelet features.

[0044] S1023. The effective digital pulse waveform in one-dimensional form is converted into a two-dimensional pulse image through time-series matrix mapping, and the two-dimensional pulse image is input into a pre-trained two-dimensional convolutional neural network for convolution extraction to obtain the hidden layer output vector as the pulse image depth feature.

[0045] To fully exploit the hidden nonlinear dependencies in unprocessed digitized pulse snapshots, the execution agent employs a Gramian Angular Field or a one-dimensional folding method to fold the data into segments of length [missing information]. One-dimensional time series mapping is A two-dimensional pixel matrix is ​​generated to form a two-dimensional pulse image. This image is then fed into a two-dimensional convolutional neural network (CNN) containing several convolutional layers, non-linear activation layers (such as ReLU), and max pooling layers. The network was initially trained using backpropagation on a large-scale, offline-collected calibration dataset. During real-time inference, the high-dimensional tensor (i.e., the deep feature embedding vector) output from the hidden layer before the last pooling layer or fully connected layer is extracted as the deep features of the pulse image.

[0046] S1024. Combine time-domain statistical features, wavelet multi-scale features, and pulse image depth features to form multi-dimensional pulse shape features.

[0047] The feature vectors extracted from the three dimensions of physical statistical mechanism, time-frequency transformation space, and deep learning latent space are concatenated and fused to form a high-dimensional joint feature vector, namely, multi-dimensional pulse shape features, which enhances the ability of features to represent different pulse types and environmental changes. Subsequently, for subsequent model self-updating, the system timestamps each original valid digital pulse waveform data with this set of high-dimensional features and stores them in a time-series binding, forming a complete mapping record. As the detection experiment progresses, these records are aggregated to form a continuously acquired dataset.

[0048] S103: Acquire the current environmental state data of the detector, and based on the current environmental state data, filter out high-confidence neutron pulse samples and high-confidence gamma pulse samples from the continuously acquired data, and use the high-confidence neutron pulse samples and high-confidence gamma pulse samples to update the reference pulse template and feature normalization parameters in real time.

[0049] When detectors operate for extended periods in complex conditions such as outdoors or in nuclear reactors, changes in ambient temperature can cause significant shifts in the luminescence yield and the time decay constants of the fast and slow components of the scintillator. Simultaneously, temperature and strong radiation can also cause drift in the PMT gain. This leads to a decrease in the applicability of the original static discrimination boundary. The core of this step lies in endowing the method with the ability to automatically sense the environment and spontaneously calibrate the model.

[0050] In one optional implementation, the current environmental state data includes temperature data of the detector's environment, detector gain data, baseline noise data, and pulse amplitude range data. The current environmental state data of the detector is acquired, and based on this data, high-confidence neutron pulse samples and high-confidence gamma pulse samples are selected from continuously acquired data. The reference pulse template and feature normalization parameters are then updated in real time using these high-confidence neutron pulse samples and high-confidence gamma pulse samples, including sub-steps S1031 to S1034. S1031. Monitor the current environmental status data in real time. When the change in the current environmental status data compared with the preset historical benchmark data exceeds the preset environmental drift threshold, start the self-calibration update process.

[0051] The system periodically reads ambient temperature data, baseline noise root mean square (RMS), and detector gain data (using constant LED illumination to monitor the relative change in PMT output amplitude) through a temperature sensor, dark current monitoring circuit, and LED constant light source calibration circuit integrated into the front-end hardware. When a preset environmental drift threshold is detected, such as the current temperature fluctuating by more than 3 degrees Celsius compared to the reference temperature during the last calibration, or the gain change exceeding 5%, an interrupt request is triggered to start the self-calibration update process to prevent discrimination errors caused by parameter rigidity.

[0052] S1032. After starting the self-calibration update process, the initial neutron cluster center and the initial gamma cluster center are determined by clustering algorithm in the continuously collected data, and the data whose distance from the initial neutron cluster center and the initial gamma cluster center is within the preset safety radius are extracted and used as high-confidence neutron pulse samples and high-confidence gamma pulse samples, respectively.

[0053] Since manually labeled supervised data is unavailable in dynamic environments, the execution entity employs unsupervised clustering algorithms (such as K-Means or Gaussian Mixture Models, GMM) to find cluster centers in continuously acquired data within the latest sliding time window. Two-class clustering is performed using multi-dimensional pulse shape features, yielding two feature space centroids representing neutron and gamma characteristics upon convergence: the initial neutron cluster center and the initial gamma cluster center. Because the edge data of the clusters is highly likely to be misclassified data due to feature overlap, the execution entity sets a strict Euclidean or Mahalanobis distance as a preset safety radius, using these two centroids as the center, and extracts only the core region data falling within this safety radius from the continuously acquired data. These data exhibit strong class typicality and are designated as high-confidence neutron pulse samples and high-confidence gamma pulse samples.

[0054] S1033. Calculate the mean waveform and global optimized attenuation shape parameter of the high-confidence neutron pulse sample and the high-confidence gamma pulse sample respectively, and generate the updated reference pulse template based on the mean waveform and the global optimized attenuation shape parameter.

[0055] After obtaining high-confidence samples, besides simply averaging the time-domain waveforms to obtain the mean waveform, the more crucial step is to accurately fit the decay tail of the pulse (this tail is the core region for physical discrimination of neutrons and gamma rays). In actual calculations, if the traditional least squares method (L2 norm) is used for fitting, outliers caused by gamma rays or electronic glitches can drastically skew the fitted curve. Therefore, the execution unit uses a robust estimation method, employing L1 norm minimization (Least Absolute Deviations) combined with a global optimization algorithm to perform multi-exponential fitting on the decay segment of the high-confidence samples, extracting globally optimized decay shape parameters (such as the slow component decay time constant) with strong anti-interference capabilities. By fusing and reconstructing this noise-resistant decay parameter with the mean waveform, an updated reference pulse template representing the latest physical state of the detector can be generated.

[0056] To address the inherent nonlinear distortion problem of the detector under large pulse amplitudes, in one optional implementation, during the process of generating the updated reference pulse template based on the mean waveform and globally optimized attenuation shape parameters, i.e., during the execution of step S1033 above, the method also includes nested sub-steps S10331 to S10333: S10331. Determine whether the pulse amplitude range data in the current environmental state data falls within the nonlinear response region of the photomultiplier tube contained in the detector.

[0057] When the incident neutron or gamma energy is extremely high, the massive number of photons emitted by the scintillator causes a severe space charge effect at the final stage dynamo of the PMT. The mutual repulsion between photoelectrons leads to a decrease in collection efficiency, and the pulse amplitude, which should have increased linearly, undergoes saturation compression, forcing the pulse shape to broaden. The actuator reads the pulse amplitude range data to determine whether the peak value exceeds the factory-calibrated upper limit threshold current of the PMT linearity.

[0058] S10332. If the sample falls into the nonlinear response region, the preset artificial neural network nonlinear response calibration model is invoked to perform high-dimensional spatial mapping distortion correction on the high-confidence neutron pulse sample and high-confidence gamma pulse sample in the nonlinear response region, so as to obtain nonlinear response correction sample data.

[0059] Once a nonlinear response is confirmed, the traditional PSD parameters will be severely distorted, causing the discrimination boundary to bend to one side. At this point, the actuator activates a lightweight Artificial Neural Network (ANN). This ANN model is trained offline using different attenuators to forcibly acquire aligned mapping data of linear and nonlinear waveforms. The distorted waveform is input into the ANN, and the network uses a multilayer perceptron nonlinear transformation to map it back to a virtual linear high-dimensional space, recovering the theoretical waveform before compression by the space charge effect, thus obtaining nonlinear response correction sample data.

[0060] S10333: Replace the corresponding high-confidence neutron pulse samples and high-confidence gamma pulse samples in the nonlinear response region with the nonlinear response correction sample data to calculate the mean waveform and global optimized attenuation shape parameters, so as to correct the boundary bending error caused by the nonlinear response of the photomultiplier tube.

[0061] The original distorted, inferior samples were replaced with sample data that restored the true shape, and then participated in the above L1 norm robust fitting calculation. This ensures that even in the extremely high energy range (PMT nonlinear region), the generated reference pulse template can accurately characterize the intrinsic physical properties of the particles, reducing the boundary offset caused by the "Banana Shape" in the high energy range.

[0062] S1034. Calculate the feature distribution centers of the high-confidence neutron pulse sample and the high-confidence gamma pulse sample respectively. Calculate the feature compensation scale factor based on the feature distribution centers and use the feature compensation scale factor as the updated feature normalization parameter.

[0063] For the newly extracted multi-dimensional pulse shape feature set, its mean and variance in the multi-dimensional space are calculated to obtain the latest feature distribution center. By comparing the center offset between the current calibration cycle and the previous calibration cycle, the scaling matrix and translation vector (i.e., feature compensation scaling factor) of each feature dimension are calculated, which are used as the updated feature normalization parameters. This parameter is used to subsequently force the drifted feature data back to the standard distribution range.

[0064] S104. Using the updated reference pulse template and feature normalization parameters, the neutron gamma discrimination boundary is dynamically corrected to obtain the corrected neutron gamma discrimination boundary.

[0065] In one alternative implementation, the neutron-gamma discrimination boundary is dynamically corrected using the updated reference pulse template and feature normalization parameters to obtain the corrected neutron-gamma discrimination boundary, including sub-steps S1041 and S1042: S1041. Use the feature normalization parameter to perform feature drift compensation on the extracted time-domain statistical features and wavelet multi-scale features to obtain a normalized feature set.

[0066] As subsequent real-time data continues to flow in, the detector's operating environment continues to change slightly. The execution unit applies the translation vector and scaling matrix of the feature compensation scale factor to the newly extracted features in real time. This mathematical compensation operation forcibly cancels the baseline and luminescence intensity drift caused by gain and temperature, ensuring that all normalized feature sets after compensation are at the same calibration scale, providing a stable input data source for the classifier.

[0067] S1042. Using the normalized feature set as the input sample set, combined with the updated reference pulse template, the support vector machine algorithm is used to recalculate the high-dimensional hyperplane weight matrix used to distinguish between neutron and gamma categories, and the updated high-dimensional hyperplane generated based on the high-dimensional hyperplane weight matrix is ​​determined as the corrected neutron-gamma discrimination boundary.

[0068] At this point, the executing entity uses the high-confidence normalized feature set with precise labels (obtained through high-confidence clustering and self-annotated in S1032) as training samples. Employing the Support Vector Machine (SVM) algorithm, and selecting an appropriate kernel function (such as the Radial Basis Function (RBF), it solves the optimization problem of maximizing the classification margin through quadratic programming, recalculating the support vectors of the support boundary and the high-dimensional hyperplane weight matrix. The optimal classification decision surface defined by this weight matrix is ​​the latest, corrected neutron gamma discrimination boundary that adapts to the current environmental state.

[0069] S105 inputs the multi-dimensional pulse shape features into the corrected neutron gamma discrimination boundary, outputs the corresponding neutron gamma discrimination result, evaluates the discrimination performance quality based on the neutron gamma discrimination result, and triggers the next round of update of the reference pulse template and feature normalization parameters when the discrimination performance quality meets the preset descent condition.

[0070] In order to make full use of the advantages of various feature extraction methods and take into account the reliability of extreme scenarios, this embodiment constructs a decision fusion mechanism and forms a closed-loop feedback at the output end.

[0071] In one optional implementation, the multi-dimensional pulse shape features are input into the corrected neutron gamma discrimination boundary, and the corresponding neutron gamma discrimination result is output, including sub-steps S1051 to S1053: S1051. Input the corresponding time-domain statistical features and wavelet multi-scale features in the normalized feature set into the multi-dimensional spatial projection decision classifier for classification and determination, and obtain the confidence probability of the first category.

[0072] This step utilizes traditional, baseline-compensated statistical and time-frequency features, and through the hyperplane distance mapping mechanism of the SVM classifier (such as Platt Scaling), converts the geometric distance from the input sample to the corrected neutron-gamma discrimination boundary into a posterior probability between 0 and 1, outputting the first-class confidence probability that the waveform belongs to either neutron or gamma. This probability has excellent physical interpretability.

[0073] S1052. Input the pulse image depth features into the fully connected layer at the end of the two-dimensional convolutional neural network for activation judgment to obtain the second category confidence probability.

[0074] In parallel, the feature vectors from the deep learning branch are fed into the fully connected layer network at the tail of the CNN, and normalized exponentially after passing through the Softmax activation function, outputting the second-class confidence probability that represents the nonlinear feature extraction result of the neural network.

[0075] S1053. A decision fusion layer with a modified neutron-gamma discrimination boundary is adopted. The first category confidence probability and the second category confidence probability are weighted and cross-entropy fused to obtain the joint decision probability. When the joint decision probability is greater than the preset boundary threshold, the neutron category label is output; otherwise, the gamma ray category label is output, and the neutron-gamma discrimination result composed of the corresponding labels is obtained.

[0076] The decision fusion layer assigns weights to probabilities based on the performance of different features during historical calibration. The joint decision probability is then output using a weighted probability and cross-entropy cost function. When the joint decision probability exceeds a set threshold (e.g., 0.5), the particle is classified as a neutron; otherwise, it is classified as a gamma ray.

[0077] In particular, in applications such as nuclide leakage or accelerator high-pulse target bombardment, the radiation field can lead to extremely high count rates (e.g., greater than 1 Mcps). At this point, the severe overlap of numerous pulses at their beginnings and ends causes both complex CNN features and wavelet decomposition to suffer from reduced recognition accuracy and increased computational resource consumption due to waveform overlap. To address this extreme problem, in one optional implementation, step S1053—which employs a decision fusion layer with a corrected neutron gamma discrimination boundary to perform weighted cross-entropy fusion of the first and second category confidence probabilities to obtain a joint decision probability and output the result—further includes sub-steps S10531 to S10533 as a degradation response strategy: S10531. Obtain the real-time event count rate of the detector; when the real-time event count rate exceeds the preset congestion threshold, convert the preprocessed effective digital pulse waveform into a signal vector form.

[0078] Trigger an interrupt and stop time-consuming deep / multi-scale operations such as convolutional neural networks and wavelet transforms. Treat the raw waveform sequence acquired by the current ADC directly as a vector in multi-dimensional space, that is, construct a signal vector containing the intensity of time sampling points.

[0079] S10532. Calculate the inner product of the effective digital pulse waveform in signal vector form and the updated reference pulse template to obtain a waveform similarity scalar used to characterize the waveform similarity.

[0080] By utilizing the dot product operation in pattern recognition technology, the inner product (or cosine similarity) between the current signal vector and the updated reference pulse template vector generated in the previous self-calibration step is directly calculated. The computational complexity of this inner product is lower than the complete feature extraction and classification process, and can be pipelined within a single clock cycle using the multiply-accumulate unit (MAC) of the underlying hardware DSP. The inner product result serves as a high-speed waveform similarity scalar, measuring the proximity of the real-time pulse to the neutron or gamma standard template.

[0081] S10533: Replace the first category confidence probability obtained by classification and determination of time-domain statistical features and wavelet multi-scale features with waveform similarity scalar, and input the second category confidence probability into the decision fusion layer for downgraded cross-entropy fusion determination, and output the neutron gamma discrimination result.

[0082] Under system overload and congestion conditions, the confidence probability of the first category is discarded, and the waveform similarity scalar obtained by the inner product operation is injected into the fusion layer to maintain the basic discrimination output at high count rates.

[0083] Finally, to ensure the long-term stability of the entire self-calibration method, this invention establishes a result-oriented closed-loop quality evaluation system. In one optional implementation, the discrimination performance quality is evaluated based on the neutron-gamma discrimination results. When the discrimination performance quality meets a preset decline condition, the next round of updating the reference pulse template and feature normalization parameters is triggered, including sub-steps S1054 to S1056: S1054. Statistically analyze the neutron-gamma discrimination results of all valid digital pulse waveforms within the preset time window, and divide the data into neutron data clusters and gamma data clusters based on the neutron-gamma discrimination results.

[0084] The executing entity opens a macroscopic statistical window lasting several minutes to collect all the identified data during this period. Based on the classification labels they carry, the data is aggregated into independent neutron data clusters and gamma data clusters.

[0085] S1055. Calculate the absolute distance between the feature distribution centers of the neutron data cluster and the gamma data cluster, and the sum of the full width at half maximum (FWHM) of the neutron data cluster and the full width at half maximum (FWHM) of the gamma data cluster. Determine the ratio of the absolute distance to the sum of the full width at half maximum (FWHM) as the discrimination quality factor, and use the discrimination quality factor as the discrimination performance quality.

[0086] The effectiveness of the discrimination performance depends on whether the two classes of data are "sufficiently separated" (large inter-class distance) and "sufficiently concentrated" (small intra-class variance) in the feature space. The implementing entity uses the Figure of Merit (FOM), a common evaluation standard in the field of nuclear radiation measurement, for quantitative assessment. Its mathematical formula is shown below:

[0087] In the formula, This represents the characteristic distribution center (mean) of the neutron data cluster along the one-dimensional projection direction. This represents the characteristic distribution center (mean) of the gamma data cluster along the one-dimensional projection direction. This represents the absolute distance between the two types of centers; a larger distance indicates a higher degree of separation. The full width at half maximum (FWHM) represents the distribution of neutron data clusters on the map, characterizing the degree of dispersion of the neutron clusters. This represents the full width at half maximum (FWHM) of the gamma data cluster, characterizing the degree of dispersion of the gamma cluster. The calculated... The value reflects the current discrimination performance quality. Typically, in organic scintillators, A score greater than 1.2 is considered to achieve good differentiation, while a score greater than 1.5 indicates a high degree of class separation.

[0088] S1056. Obtain the quality degradation rate of the identification quality factor compared to the pre-stored historical reference quality factor. When the quality degradation rate is greater than or equal to the warning threshold, determine that the preset degradation condition is met, and automatically send a feedback signal to trigger the next round of updating of the reference pulse template and feature normalization parameters.

[0089] The real-time identification quality factor calculated in the current time window The historical reference quality factor recorded by the detector under optimal operating conditions The comparison is performed, and the percentage decrease is calculated. If it is found that the quality degradation rate has reached the set tolerance limit (e.g., the FOM value drops by more than 15% warning threshold) due to slow environmental drift, cumulative device aging, or unforeseen external noise interference, it indicates that the existing classification boundary is no longer suitable for the current detection physical state. At this time, the execution entity automatically generates an internal system-level interrupt feedback signal. This signal directly returns to execute the aforementioned steps S103 and S104, forcing the algorithm to recapture the current background noise and waveform features, start a new unsupervised clustering process, and iteratively refresh the reference pulse template and reconstruct the decision high-dimensional hyperplane based on the latest real high-confidence samples.

[0090] This invention also provides a neutron gamma discrimination system based on pulse shape self-calibration, referring to... Figure 3 The diagram shows a functional block diagram of a neutron gamma discrimination system 300 based on pulse shape self-calibration according to the present invention. The system may include the following modules: The pulse acquisition and preprocessing module 301 is used to acquire the digital pulse waveform output by the detector and preprocess the digital pulse waveform to obtain an effective digital pulse waveform. The feature extraction and data construction module 302 is used to extract features from the effective digital pulse waveform to obtain multi-dimensional pulse shape features, and to combine the effective digital pulse waveform and its corresponding multi-dimensional pulse shape features in a time sequence to form continuously acquired data. The self-calibration sample screening and parameter update module 303 is used to acquire the current environmental state data of the detector, screen out high-confidence neutron pulse samples and high-confidence gamma pulse samples from continuously acquired data based on the current environmental state data, and use the high-confidence neutron pulse samples and high-confidence gamma pulse samples to update the reference pulse template and feature normalization parameters in real time. The dynamic correction module 304 for the discrimination boundary is used to dynamically correct the neutron gamma discrimination boundary using the updated reference pulse template and feature normalization parameters, so as to obtain the corrected neutron gamma discrimination boundary. The comprehensive judgment and quality assessment feedback module 305 is used to input multi-dimensional pulse shape features into the corrected neutron gamma discrimination boundary, output the corresponding neutron gamma discrimination result, evaluate the discrimination performance quality based on the neutron gamma discrimination result, and trigger the next round of update of reference pulse template and feature normalization parameters when the discrimination performance quality meets the preset decline condition.

[0091] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the neutron gamma discrimination method based on pulse shape self-calibration of the present invention.

[0092] The communication bus mentioned in the aforementioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the figure, but this does not mean that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include RAM, or it can include NVM, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0093] The processors mentioned above can be general-purpose processors, including CPUs, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0094] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the neutron gamma discrimination method based on pulse shape self-calibration of the embodiments of the present invention.

[0095] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0096] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, electronic devices, apparatuses, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A neutron gamma discrimination method based on pulse shape self-calibration, characterized in that, include: The digital pulse waveform output by the detector is acquired, and the digital pulse waveform is preprocessed to obtain an effective digital pulse waveform; Feature extraction is performed on the effective digital pulse waveform to obtain multi-dimensional pulse shape features, and the effective digital pulse waveform and its corresponding multi-dimensional pulse shape features are combined in a time sequence to form continuously acquired data. The current environmental state data of the detector is obtained, and high-confidence neutron pulse samples and high-confidence gamma pulse samples are selected from the continuously collected data based on the current environmental state data. The reference pulse template and feature normalization parameters are updated in real time using the high-confidence neutron pulse samples and the high-confidence gamma pulse samples. Using the updated reference pulse template and the feature normalization parameters, the neutron gamma discrimination boundary is dynamically corrected to obtain the corrected neutron gamma discrimination boundary. The multi-dimensional pulse shape features are input into the corrected neutron gamma discrimination boundary, and the corresponding neutron gamma discrimination result is output. The discrimination performance quality is evaluated based on the neutron gamma discrimination result. When the discrimination performance quality meets the preset decline condition, the next round of updating of the reference pulse template and the feature normalization parameters is triggered.

2. The neutron gamma discrimination method based on pulse shape self-calibration according to claim 1, characterized in that, The preprocessing of the digital pulse waveform to obtain an effective digital pulse waveform includes: The digital pulse waveform is subjected to baseline estimation and baseline subtraction to determine a zero-level reference. Based on the zero-level reference, peak search is performed on the digital pulse waveform after baseline subtraction to obtain peak location results; The peak amplitude of the pulse is obtained based on the peak positioning result. The peak amplitude of the pulse is then used to normalize the amplitude of the digital pulse waveform after deducting the baseline, resulting in a normalized digital pulse waveform. Based on a preset noise threshold and pulse rise time validity criteria, the normalized digital pulse waveform is subjected to pulse validity screening to obtain the valid digital pulse waveform.

3. The neutron gamma discrimination method based on pulse shape self-calibration according to claim 1, characterized in that, The step of extracting features from the effective digital pulse waveform to obtain multi-dimensional pulse shape features includes: The rise time, decay time, tail integration ratio, peak amplitude, and total integral of the effective digital pulse waveform are extracted to form time-domain statistical features; The rising edge, peak value and decay tail of the effective digital pulse waveform are decomposed into multiple scale components using Haar wavelet basis functions, and the wavelet approximation coefficients and detail coefficients of each scale component are extracted to form wavelet multi-scale features. The effective digital pulse waveform in one-dimensional form is converted into a two-dimensional pulse image through a time-series matrix mapping, and the two-dimensional pulse image is input into a pre-trained two-dimensional convolutional neural network for convolution extraction to obtain the hidden layer output vector as the pulse image depth feature; The time-domain statistical features, the wavelet multi-scale features, and the pulse image depth features are combined to form the multi-dimensional pulse shape features.

4. The neutron gamma discrimination method based on pulse shape self-calibration according to claim 1, characterized in that, The current environmental state data includes temperature data of the detector's environment, detector gain data, baseline noise data, and pulse amplitude range data. The process of acquiring the detector's current environmental state data, filtering high-confidence neutron pulse samples and high-confidence gamma pulse samples from the continuously acquired data based on the current environmental state data, and using the high-confidence neutron pulse samples and high-confidence gamma pulse samples to update the reference pulse template and feature normalization parameters in real time includes: The system monitors the current environmental status data in real time. When the change in the current environmental status data compared to the preset historical baseline data exceeds the preset environmental drift threshold, a self-calibration update process is initiated. After the self-calibration update process is started, the initial neutron cluster center and the initial gamma cluster center are determined by clustering algorithm in the continuously collected data, and the data whose distance from the initial neutron cluster center and the initial gamma cluster center is within the preset safety radius are extracted and used as the high-confidence neutron pulse sample and the high-confidence gamma pulse sample, respectively. The mean waveform and globally optimized decay shape parameter of the high-confidence neutron pulse sample and the high-confidence gamma pulse sample are calculated respectively, and the updated reference pulse template is generated based on the mean waveform and the globally optimized decay shape parameter. The feature distribution centers of the high-confidence neutron pulse sample and the high-confidence gamma pulse sample are calculated respectively. The feature compensation scale factor is calculated based on the feature distribution centers, and the feature compensation scale factor is used as the updated feature normalization parameter.

5. The neutron gamma discrimination method based on pulse shape self-calibration according to claim 4, characterized in that, In the process of generating the updated reference pulse template based on the mean waveform and the globally optimized attenuation shape parameters, the method further includes: Determine whether the pulse amplitude range data in the current environmental state data falls within the nonlinear response region of the photomultiplier tube contained in the detector; If the sample falls within the nonlinear response region, a preset artificial neural network nonlinear response calibration model is invoked to perform high-dimensional spatial mapping distortion correction on the high-confidence neutron pulse sample and the high-confidence gamma pulse sample within the nonlinear response region, thereby obtaining nonlinear response correction sample data. The mean waveform and the globally optimized attenuation shape parameters are calculated by replacing the corresponding high-confidence neutron pulse samples and high-confidence gamma pulse samples within the nonlinear response region using the nonlinear response correction sample data, in order to correct the boundary bending error caused by the nonlinear response of the photomultiplier tube.

6. The neutron gamma discrimination method based on pulse shape self-calibration according to claim 3, characterized in that, The step of dynamically correcting the neutron-gamma discrimination boundary using the updated reference pulse template and the feature normalization parameters to obtain the corrected neutron-gamma discrimination boundary includes: The extracted time-domain statistical features and wavelet multi-scale features are compensated for feature drift using the feature normalization parameters to obtain a normalized feature set. Using the normalized feature set as the input sample set, and combining it with the updated reference pulse template, the support vector machine algorithm is used to recalculate the high-dimensional hyperplane weight matrix used to distinguish between neutron and gamma categories, and the updated high-dimensional hyperplane generated based on the high-dimensional hyperplane weight matrix is ​​determined as the corrected neutron-gamma discrimination boundary.

7. The neutron gamma discrimination method based on pulse shape self-calibration according to claim 6, characterized in that, The step of inputting the multi-dimensional pulse shape features into the corrected neutron gamma discrimination boundary and outputting the corresponding neutron gamma discrimination result includes: The time-domain statistical features and wavelet multi-scale features corresponding to the normalized feature set are input into a multi-dimensional spatial projection decision classifier for classification and determination, and the confidence probability of the first category is obtained. The pulse image depth features are input into the fully connected layer at the end of the two-dimensional convolutional neural network for activation judgment to obtain the second category confidence probability; A decision fusion layer with the corrected neutron-gamma discrimination boundary is used to perform weighted cross-entropy fusion of the confidence probabilities of the first category and the confidence probabilities of the second category to obtain a joint decision probability. When the joint decision probability is greater than a preset boundary threshold, the neutron category label is output; otherwise, the gamma ray category label is output, and the neutron-gamma discrimination result composed of the corresponding labels is obtained.

8. The neutron gamma discrimination method based on pulse shape self-calibration according to claim 7, characterized in that, The decision fusion layer employing the modified neutron-gamma discrimination boundary performs weighted cross-entropy fusion of the first category confidence probability and the second category confidence probability to obtain a joint decision probability. When the joint decision probability is greater than a preset boundary threshold, a neutron category label is output; otherwise, a gamma ray category label is output. This process obtains the neutron-gamma discrimination result composed of the corresponding labels, including: Obtain the real-time event count rate of the detector; When the real-time event count rate exceeds the preset congestion threshold, the preprocessed effective digital pulse waveform is converted into a signal vector form; Calculate the inner product of the effective digital pulse waveform in the form of the signal vector and the updated reference pulse template to obtain a waveform similarity scalar used to characterize the waveform similarity; The waveform similarity scalar is used to replace the first category confidence probability obtained by the classification and determination of the time domain statistical features and the wavelet multi-scale features. The second category confidence probability is then input into the decision fusion layer for downgraded cross-entropy fusion determination, and the neutron gamma discrimination result is output.

9. The neutron gamma discrimination method based on pulse shape self-calibration according to claim 1, characterized in that, The step of evaluating the discrimination performance quality based on the neutron gamma discrimination results, and triggering the next round of updating the reference pulse template and the feature normalization parameters when the discrimination performance quality meets a preset decline condition, includes: The neutron-gamma discrimination results of all valid digital pulse waveforms within a preset time window are statistically analyzed, and the data are divided into neutron data clusters and gamma data clusters based on the neutron-gamma discrimination results. Calculate the absolute distance between the feature distribution center of the neutron data cluster and the feature distribution center of the gamma data cluster, and the sum of the full width at half maximum (FWHM) of the neutron data cluster and the full width at half maximum (FWHM) of the gamma data cluster. The ratio of the absolute distance to the sum of the half-height and full width is determined as the discrimination quality factor, and the discrimination quality factor is used as the discrimination performance quality. The quality degradation rate of the discrimination quality factor compared to the pre-stored historical reference quality factor is obtained. When the quality degradation rate is greater than or equal to the warning threshold, it is determined that the preset degradation condition is met, and a feedback signal is automatically issued to trigger the next round of updating of the reference pulse template and the feature normalization parameter.

10. A neutron gamma discrimination system based on pulse shape self-calibration, characterized in that, include: The pulse acquisition and preprocessing module is used to acquire the digital pulse waveform output by the detector and preprocess the digital pulse waveform to obtain an effective digital pulse waveform. The feature extraction and data construction module is used to extract features from the effective digital pulse waveform to obtain multi-dimensional pulse shape features, and to combine the effective digital pulse waveform and its corresponding multi-dimensional pulse shape features in a time sequence to form continuously acquired data. The self-calibration sample screening and parameter update module is used to acquire the current environmental state data of the detector, screen high-confidence neutron pulse samples and high-confidence gamma pulse samples from the continuously acquired data based on the current environmental state data, and use the high-confidence neutron pulse samples and high-confidence gamma pulse samples to update the reference pulse template and feature normalization parameters in real time. The dynamic correction module for the discrimination boundary is used to dynamically correct the neutron gamma discrimination boundary using the updated reference pulse template and the feature normalization parameters, so as to obtain the corrected neutron gamma discrimination boundary. The comprehensive judgment and quality assessment feedback module is used to input the multi-dimensional pulse shape features into the corrected neutron gamma discrimination boundary, output the corresponding neutron gamma discrimination result, evaluate the discrimination performance quality based on the neutron gamma discrimination result, and trigger the next round of updating of the reference pulse template and the feature normalization parameters when the discrimination performance quality meets the preset decline condition.

Images