A prompt gamma-ray spectroscopy nuclide information inversion method using an artificial neural network

Abstract

The application discloses a prompt gamma-ray spectrometry nuclide information inversion method using an artificial neural network and belongs to the technical field of nuclear analysis. The method comprises the following steps: acquiring prompt gamma-ray spectrometry of a sample to be measured; and adaptively selecting a window method or a full-spectrum analysis method path for processing according to the requirement for analysis speed or accuracy. The window method path extracts characteristic energy window data to form a characteristic vector, which is input into a pre-trained lightweight machine learning model of an embedded AI processor; and the full-spectrum analysis method path inputs full-spectrum data into a pre-trained deep learning model after standardizing and preprocessing the energy spectrum. Both the two models are obtained by using a training set containing simulation and measurement data. The application adaptively learns the energy spectrum characteristics through an artificial intelligence model, effectively overcomes the problem that a traditional algorithm has insufficient resolving capacity for overlapping peaks and weak peaks, and realizes rapid, online and high-precision inversion of nuclide information in combination with embedded deployment.

Description

Technical Field

[0001] This invention relates to the field of nuclear analysis technology, and specifically to a method for inverting nuclide information in transient gamma spectroscopy using an artificial neural network. Background Technology

[0002] Prompt Gamma Neutron Activation Analysis (PGNAA) is an important nuclide analysis technique. Its basic principle is to irradiate the sample with a neutron beam. The nuclei of different nuclides in the sample capture neutrons, forming excited-state composite nuclei. During de-excitation, these composite nuclei release characteristic gamma rays with specific energies. By detecting and analyzing the energy spectrum of these characteristic gamma rays, qualitative and quantitative analysis of the types and amounts of nuclides in the sample can be achieved. This technique has significant advantages such as high analysis speed, simultaneous analysis of multiple elements, and no need for complex chemical pretreatment. It has been widely applied in fields such as coal quality testing, cement raw material proportioning, mineral exploration, and environmental monitoring.

[0003] In the specific implementation and analysis of PGNAA technology, the analysis method has long relied mainly on a combination of traditional digital processing algorithms and manual interpretation. These traditional methods typically include automatic or manual peak finding, peak area fitting, and net peak area calculation of the acquired gamma spectrum, followed by comparison and identification based on a pre-set nuclide characteristic energy database.

[0004] However, traditional PGNAA nuclide analysis methods face a series of inherent limitations in practical applications: First, when the sample composition is complex, the characteristic gamma-ray peaks emitted by different nuclides often overlap significantly in the energy spectrum. Traditional algorithms struggle to effectively resolve these overlapping peaks and cannot accurately distinguish and quantify the independent contributions of each nuclide. Second, matrix effects (such as density and composition variations) can cause distortions in the characteristic peak shapes. Simultaneously, statistical fluctuations exist in the background radiation during measurement, and the fixed algorithm parameters of traditional methods are difficult to adapt to these disturbances, leading to decreased analytical accuracy. Furthermore, for trace nuclides with extremely low concentrations, the characteristic peak signals are weak and the signal-to-noise ratio is low, resulting in significantly insufficient identification sensitivity of traditional methods and a high risk of missed detections. These problems collectively limit the reliability, accuracy, and practicality of PGNAA technology under complex and variable operating conditions.

[0005] Therefore, developing a novel nuclide information inversion method that can intelligently analyze complex energy spectra, overcome the difficulties in identifying overlapping and weak peaks, and possess adaptive capabilities has become a key requirement for enhancing the core competitiveness of PGNAA technology. Summary of the Invention

[0006] The purpose of this invention is to provide a method for inverting nuclide information in the transient gamma spectrum using an artificial neural network, so as to overcome the above-mentioned problems in the prior art, and to intelligently analyze complex energy spectra, overcome the difficulty of identifying overlapping peaks and weak peaks, and improve the adaptive ability to matrix effects and noise.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for inverting transient gamma spectral nuclide information using artificial neural networks includes the following steps: S1. Energy spectrum acquisition: The sample to be tested is irradiated with neutrons, and the transient gamma rays emitted by the sample are collected by a gamma detector to form the transient gamma energy spectrum to be tested. S2. Data Processing and Model Inference: Based on the preset analysis requirements, select any of the following methods to process the measured transient gamma spectrum, and input the processing results into the corresponding deployed inversion model for inference to obtain the inference results: Method 1: Extract window data corresponding to the preset nuclide characteristic energy from the transient gamma energy spectrum to be tested, form a feature vector, and input the feature vector into the first inversion model for inference; the first inversion model is a pre-trained lightweight machine learning model deployed in an embedded AI processor; Method 2: Perform energy scaling and count rate normalization preprocessing on the transient gamma spectrum to be tested to obtain standardized full-spectrum data, and input the standardized full-spectrum data into the second inversion model for inference; the second inversion model is a pre-trained deep learning model deployed in an embedded AI processor; S3. Result Output: Based on the reasoning results, output the information on the types and contents of nuclides in the sample to be tested in real time.

[0008] Furthermore, in the first method, the extraction of window data corresponding to the preset nuclide feature energy includes: taking the peak position corresponding to the preset nuclide feature energy as the center, determining the window width according to the energy resolution of the gamma detector, and calculating the net peak count after subtracting the background within the window, and the feature vector is composed of multiple net peak counts.

[0009] Furthermore, in the second method, the preprocessing includes: The transient gamma spectrum to be measured is calibrated with energy, and its channel address is converted into the corresponding energy value. The count rate is normalized by calibrating the energy spectrum, and the counts of each channel are divided by the measurement time to obtain the standardized full spectrum data.

[0010] Furthermore, the lightweight machine learning model is one of artificial neural networks, random forests, support vector machines, or XGBoost models.

[0011] Furthermore, the deep learning model is a one-dimensional convolutional neural network, a long short-term memory network, a Transformer neural network, or a hybrid neural network model that combines one-dimensional convolutional neural networks and long short-term memory networks.

[0012] Furthermore, at least one of the first inversion model and the second inversion model undergoes operator fusion and weight quantization processing before deployment.

[0013] Furthermore, the embedded AI processor is a neural network processor, a field-programmable gate array (FPGA), or an embedded graphics processor.

[0014] Furthermore, the preset analysis requirements include: When focusing on speed analysis, choose method one. When focusing on analysis accuracy, choose method two.

[0015] Furthermore, the first inversion model and the second inversion model are deployed as independent models in the embedded AI processor, and the corresponding model is called according to the preset analysis requirements.

[0016] Furthermore, the second inversion model is a multi-task neural network model, which can simultaneously output the nuclide classification results and the nuclide content regression results.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Significantly improves the analytical accuracy and model robustness of complex energy spectra: Addressing the challenges of resolving overlapping peaks, weak peaks, and peak distortion caused by matrix effects that traditional algorithms struggle with, the deep learning models employed in this invention (such as one-dimensional convolutional neural networks and long short-term memory networks) can autonomously learn deep features and complex nonlinear relationships within the full spectrum data end-to-end. For example, one-dimensional convolutional neural networks can effectively extract local peak shape features, and long short-term memory networks can capture long-range dependencies in energy spectrum sequences (such as the relative intensities of different characteristic peaks), thereby achieving high-precision and high-reliability identification and quantification of nuclides against complex interference backgrounds.

[0018] 2. A dual-path intelligent analysis solution is provided to flexibly adapt to business needs: The innovative dual-path architecture of the window method and the full-spectrum analysis method allows the system to make the optimal choice based on the different emphases on speed or accuracy in actual applications. The window method extracts limited feature energy window data and combines it with lightweight machine learning models (such as random forests and XGBoost), resulting in low computational cost and fast inference speed, making it particularly suitable for online embedded analysis scenarios with extremely high real-time requirements; the full-spectrum method utilizes standardized full energy spectrum information and analyzes it through a more complex deep learning model, resulting in high information utilization and superior analysis accuracy, suitable for offline precision analysis scenarios with extremely strict requirements for result accuracy.

[0019] 3. Achieved online real-time analysis and on-site deployment capabilities: By compressing the trained neural network model through optimization techniques such as operator fusion and weight quantization, and deploying it in a dedicated embedded AI processor (such as an NPU or FPGA), the computationally complex intelligent analysis algorithm was successfully migrated from the server to the analysis terminal. This deployment strategy effectively overcomes the bottleneck of traditional neural network analysis, which relies on high-performance computers and cannot be performed online in real time. This enables PGNAA technology to perform rapid and continuous on-site online nuclide analysis, greatly expanding its practical application in industrial process control (such as coal and cement production) and on-site detection. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the technical process of the method described in this invention. Figure 2 This is a schematic diagram of the fully connected neural network used in the present invention to retrieve nuclide information; Figure 3 This invention provides the nuclide information obtained through random forest inversion. Figure 4 This is a schematic diagram of the support vector machine inversion of nuclide information according to the present invention; Figure 5 This is a schematic diagram of the XGBoost model for inverting nuclide information according to the present invention; Figure 6 This is a schematic diagram of the deep convolutional neural network used to retrieve nuclide information according to the present invention; Figure 7 This is a schematic diagram of the residual neural network used to retrieve nuclide information according to the present invention; Figure 8 This is a schematic diagram of the long short-term memory neural network used in this invention to retrieve nuclide information; Figure 9 This is a schematic diagram of the Transformer neural network used to invert nuclide information according to the present invention; Figure 10This is a schematic diagram of the online inversion of nuclide information by deploying a neural network model in the embedded processor of the present invention. (a) is a schematic diagram of the multi-model deployment of a single-task neural network in the embedded processor, and (b) is a schematic diagram of the single-model deployment of a multi-task neural network in the embedded processor. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] This embodiment provides a method for inverting instantaneous gamma-ray spectrum nuclide information using an artificial neural network, comprising the following steps: S1. Energy spectrum acquisition: The sample to be tested is irradiated with neutrons, and the transient gamma rays emitted by the sample are collected by a gamma detector to form the transient gamma energy spectrum to be tested. S2. Data Processing and Model Inference: Based on the preset analysis requirements, select any of the following methods to process the measured transient gamma spectrum, and input the processing results into the corresponding deployed inversion model for inference to obtain the inference results: Method 1: Extract window data corresponding to the preset nuclide characteristic energy from the transient gamma energy spectrum to be tested, form a feature vector, and input the feature vector into the first inversion model for inference; the first inversion model is a pre-trained lightweight machine learning model deployed in an embedded AI processor; Method 2: Perform energy scaling and count rate normalization preprocessing on the transient gamma spectrum to be tested to obtain standardized full-spectrum data, and input the standardized full-spectrum data into the second inversion model for inference; the second inversion model is a pre-trained deep learning model deployed in an embedded AI processor; S3. Result Output: Based on the reasoning results, output the information on the types and contents of nuclides in the sample to be tested in real time.

[0023] Specifically, the method is implemented based on a hardware system, mainly including a neutron source, a gamma detector array, a multichannel pulse amplitude analyzer, an embedded AI processor, and a host computer. The neutron source generates pulsed neutrons to irradiate the sample; the gamma detector array (such as a high-purity germanium detector or a scintillator detector) collects the transient gamma rays emitted by the sample; the multichannel pulse amplitude analyzer converts the pulse signal output by the detector into a digital energy spectrum; the embedded AI processor (such as an NPU, FPGA, or GPU) deploys a pre-trained inversion model, responsible for intelligently analyzing the energy spectrum; and the host computer displays and stores the inversion results. The complete technical flow of the method is as follows: Figure 1 As shown, the overall path from physical signal acquisition and dual-path intelligent processing to embedded online inversion is illustrated.

[0024] The implementation process of the method described in this embodiment will be explained in detail below.

[0025] Step S1: Energy Spectrum Acquisition Fast neutrons are generated using a neutron source (such as a neutron generator). These fast neutrons first lose energy through elastic collisions with the nuclei of moderator atoms, eventually being moderated into thermal neutrons. The moderated thermal neutron beam then uniformly irradiates the sample to be tested (such as coal or ore). The nuclei of various nuclides in the sample capture the thermal neutrons, forming excited-state recombination nuclei. During the de-excitation process, these excited-state recombination nuclei emit transient gamma rays with characteristic energies of the nuclides. These gamma rays are collected by a high-resolution gamma detector array. Commonly used detectors include high-purity germanium detectors with excellent energy resolution or scintillator detectors with high detection efficiency. When the gamma rays interact with the detector medium, their energy is converted into a weak electrical pulse signal, the amplitude of which is proportional to the gamma ray energy. This signal is then amplified and shaped in two stages by a preamplifier and a main amplifier, and finally, the amplitude of the analog pulse is precisely converted into the corresponding digital channel by an analog-to-digital converter. Finally, a multichannel pulse amplitude analyzer classifies pulses of different amplitudes according to their channel addresses and accumulates their counts, forming an energy spectrum distribution map with channel address as the x-axis and cumulative count as the y-axis, which is the measured transient gamma energy spectrum. This complete energy spectrum acquisition process provides the most basic source of raw data for subsequent intelligent qualitative and quantitative analysis.

[0026] Step S2: Data Processing and Model Inference The system adaptively selects any of the following methods to process the transient gamma spectrum under test based on preset analysis requirements (e.g., prioritizing analysis speed or prioritizing analysis accuracy), and inputs the processing result into the corresponding deployed inversion model for inference to obtain the inference result. This dual-path adaptive architecture enables the system to flexibly respond to different application scenarios and achieve an optimal balance between real-time performance and accuracy.

[0027] 1. Method 1: Window method combined with lightweight machine learning model (focusing on high-speed analysis) Window data corresponding to the preset nuclide characteristic energy is extracted from the transient gamma spectrum to be measured, forming a feature vector. The feature vector is then input into a first inversion model for inference. The first inversion model is a pre-trained lightweight machine learning model deployed in an embedded AI processor, characterized by high computational efficiency, thereby enabling rapid inversion of nuclide information. The extraction of window data corresponding to the preset nuclide characteristic energy includes: centering on the peak position corresponding to the preset nuclide characteristic energy, determining the window width according to the energy resolution of the gamma detector, subtracting the background within the window, and calculating the net peak count. The feature vector is composed of multiple net peak counts.

[0028] 1) Feature extraction (nuclear fingerprint construction) First, based on the library of nuclides to be analyzed, one or more preset characteristic energies for each target nuclide are determined (e.g., the characteristic energy of hydrogen (H) is approximately 2223 keV, and the characteristic energy of iron (Fe) is approximately 7631 keV). The system employs a multi-scale peak detection algorithm to automatically identify the spectral peak positions (peak positions) corresponding to these characteristic energies on the transient gamma spectrum to be measured.

[0029] Subsequently, a dynamic energy window is adaptively determined based on the energy resolution of the gamma detector used (usually characterized by the full width at half maximum (FWHM)) and centered on each identified characteristic peak. For example, for high-purity germanium detectors with extremely high energy resolution, the window width can be set to 0.5 to 1 times the FWHM; for scintillator detectors with lower resolution, the window width needs to be appropriately increased to cover a wider peak shape.

[0030] Within each defined energy window, background subtraction is performed to obtain the pure signal. Specifically, a linear or polynomial fitting method is used to estimate the background radiation level in the window region. Then, the total count integral within the window is calculated, and the estimated background integral is subtracted to obtain the net peak count of the characteristic peak. For a nuclide containing multiple characteristic peaks, the net peak counts of all its characteristic peaks are combined sequentially to form a multidimensional feature vector characterizing the nuclide information of the sample. This vector is essentially a quantitative expression of the unique fingerprint of the nuclide.

[0031] 2) Lightweight model inference The extracted feature vectors are input into the first inversion model deployed in the embedded AI processor for inference. This model is a pre-trained lightweight machine learning model, and the selectable model is one of artificial neural networks, random forests, support vector machines, or XGBoost (extreme gradient boosting model). This embodiment uses an artificial neural network. This network adopts a shallow structure, typically containing 1 to 3 hidden layers. The number of input layer nodes is equal to the dimension of the feature vector, and the number of output layer nodes is equal to the number of nuclides to be analyzed. The number of hidden layer nodes can be flexibly configured according to the task complexity (usually 50%-150% of the number of input layer nodes). For the specific process of using the window method combined with the lightweight model for nuclide information inversion, please refer to [link to documentation]. Figures 2-5 They respectively demonstrated fully connected neural networks ( Figure 2 Random Forest Figure 3 Support Vector Machine (SVM) Figure 4 ) and XGBoost Figure 5 How does a typical lightweight model receive feature vectors and infer nuclide content?

[0032] Because this path processes only a limited amount of data within the feature peak window, the input dimensionality is significantly reduced, allowing the model structure to remain lightweight. This results in extremely fast inference speeds and low computational resource consumption, making it particularly suitable for online and embedded analysis scenarios with extremely high real-time requirements. It enables the rapid extraction of key information from complex energy spectra and intelligent interpretation.

[0033] 2. Method Two: Full-spectrum analysis combined with deep learning models (emphasizing high-precision analysis) This method abandons the traditional manual feature extraction steps, preprocessing the transient gamma spectrum to be tested with energy scaling and count rate normalization to obtain standardized full-spectrum data. This standardized full-spectrum data is then input into a second inversion model for inference. The second inversion model is a pre-trained deep learning model deployed in an embedded AI processor. By autonomously learning the deep features and complex patterns contained in the energy spectrum end-to-end, the model achieves higher-precision nuclide identification and quantitative analysis.

[0034] 1) Full-spectrum data preprocessing and enhancement First, the raw transient gamma spectrum to be measured undergoes standardized preprocessing to eliminate the influence of instrument conditions and measurement parameters, ensuring data consistency. This preprocessing includes: energy calibration of the transient gamma spectrum to be measured, converting its channel addresses to corresponding energy values; and count rate normalization of the energy-calibrated spectrum, dividing the counts of each channel by the measurement time to obtain the standardized full-spectrum data. Specifically: Energy calibration: The detection system is calibrated using a standard radioactive source with known energy to establish the correspondence between the energy spectrum address and the gamma ray energy (for example, by establishing a conversion relationship through quadratic polynomial fitting), thereby converting the abscissa of the energy spectrum from the address to the standard energy value (unit: keV).

[0035] Count rate normalization: The original count for each energy channel (or channel address) in the energy spectrum after energy calibration is divided by the total acquisition time of that energy spectrum to obtain the count rate per unit time. This is intended to eliminate count differences caused by different measurement times.

[0036] After the above preprocessing, a standardized full-energy spectrum data vector is obtained, whose length (number of channels) is usually between 1024 and 16384, which can be directly used as input for deep learning models.

[0037] 2) Deep learning model inference The standardized full-spectrum data vector is input into a second inversion model deployed in an embedded AI processor for inference. This model is a pre-trained deep learning model, and its architecture can be a one-dimensional convolutional neural network, a long short-term memory network, a Transformer neural network, or a hybrid neural network model that combines one-dimensional convolutional neural networks and long short-term memory networks.

[0038] Specifically, the convolutional kernels of a one-dimensional convolutional neural network (1D-CNN) slide along the energy axis, automatically and effectively extracting local feature patterns in the energy spectrum, such as the peak shape of characteristic peaks and Compton edges. Long Short-Term Memory (LSTM) networks excel at capturing long-range dependencies in energy spectrum data sequences, such as the relative intensity ratios between different characteristic peaks, which is crucial for nuclide identification. Transformer neural networks, relying on their core self-attention mechanism, can process the entire energy spectrum sequence in parallel and dynamically calculate the global dependencies between any energy channels in the spectral data, showing potential in modeling complex spectral line structures and overlapping peaks. A hybrid neural network model combining the 1D-CNN and LSTM networks can, for example, employ an architecture where local features are extracted by 1D convolutional layers, followed by LSTM layers to capture sequence dependencies, and finally output through fully connected layers, thus combining the advantages of CNN local feature extraction and LSTM long-range dependency modeling. Furthermore, the model can introduce an attention mechanism, enabling the network to adaptively focus on the most information-rich regions in the energy spectrum (such as regions with strong characteristic peaks).

[0039] The output layer of the deep learning model can be designed as a multi-task learning architecture. For example, a softmax function can be used for nuclide classification, while a linear activation function can be used for nuclide content regression; or two independent neural networks can be used to implement the classification and regression tasks respectively. For content prediction, the model not only outputs point estimates, but also provides an assessment of prediction uncertainty through quantile regression or Bayesian neural network techniques, providing a credibility index for industrial decision-making.

[0040] For the specific process of nuclide information inversion using the full-spectrum analysis method combined with a deep learning model, please refer to [link / reference needed]. Figures 6-9 The document showcases deep convolutional neural networks (an implementation of one-dimensional convolutional neural networks). Figure 6 Residual neural networks (another implementation of one-dimensional convolutional neural networks) Figure 7 Long Short-Term Memory Network (LSTM) Figure 8 ) and Transformer Neural Network ( Figure 9 How to process full-spectrum data end-to-end and output inversion results.

[0041] This approach utilizes all information from the energy spectrum, including weak characteristic peaks, overlapping peak regions, and the Compton continuum, enabling the model to learn more comprehensive and in-depth physical characteristics. Therefore, it significantly improves the resolution of complex samples, overlapping peaks, and weak peaks, as well as its robustness against matrix effects and noise, achieving higher inversion accuracy. It is suitable for precision analysis scenarios where extremely high accuracy of analytical results is required.

[0042] 3. Model training, optimization, and deployment To ensure that the above-mentioned intelligent inversion method can operate stably and efficiently in actual industrial settings, this embodiment includes a complete model training, optimization, and deployment process.

[0043] 1) Model Training In Method 1, the training of the first inversion model (lightweight machine learning model) requires a large amount of standard sample energy spectrum data with known nuclide composition. First, the transient gamma spectra of these standard samples are acquired using a PGNAA measurement system, and feature vectors are extracted using the method described above. These feature vectors and the corresponding true nuclide contents (obtained through laboratory chemical analysis) constitute training sample pairs. During training, model parameters are optimized using backpropagation algorithms (for ANNs) or greedy algorithms (for tree models) to minimize the error between predicted and true values. In the inference phase, the same feature extraction process is performed on the energy spectra of unknown samples, and the resulting feature vectors are input into the trained model to obtain the predicted content values ​​for each nuclide.

[0044] In Method Two, the training of the second inversion model (deep learning model) employs a multi-stage strategy: first, the model is pre-trained using a large amount of synthetic energy spectrum data generated by Monte Carlo simulation; then, it is fine-tuned using energy spectrum data obtained from actual measurements to reduce the difference between simulation and measured data. The loss function used for training is a combination of weighted mean square error and cross-entropy, where the weights of different nuclides can be adjusted according to their importance or analytical accuracy requirements.

[0045] 2) Model Optimization Before being deployed to resource-constrained embedded AI processors, the model must undergo rigorous optimization and compression, including: Operator fusion: Combines continuous linear operations (such as convolution, batch normalization, activation functions) in neural networks into a single composite operation, significantly reducing memory access overhead and intermediate data caching.

[0046] Weight quantization: Converts model parameters from 32-bit floating-point numbers to 8-bit or even lower-bit integer representations, significantly reducing model storage size and memory usage, and accelerating the computation process. Quantization-aware training can be used to compensate for accuracy loss.

[0047] Model optimization and compression also include the following optional techniques: Model pruning: For neural network models (such as artificial neural networks or deep learning models), sparse models are generated by removing redundant connections or neurons in the network, thereby reducing computational complexity without significantly affecting accuracy.

[0048] Model distillation: Especially for lightweight tree models (such as random forests and XGBoost), distillation techniques can be used to transfer knowledge from complex models (or model ensembles) to models with a more streamlined structure, thereby achieving model miniaturization.

[0049] 3) Embedded deployment and real-time inference The optimized model is deployed in an embedded AI processor. Depending on the application scenario, a neural network processor (NPU), a field-programmable gate array (FPGA), or an embedded graphics processor (GPU) can be selected. To achieve efficient operation of the PGNAA system in practical applications, this invention proposes differentiated deployment and acceleration schemes for these heterogeneous computing platforms.

[0050] For the FPGA platform, deployment is targeted at edge computing and low-power scenarios, leveraging its high customizability and parallel processing capabilities. The deployment process includes: first, converting the trained neural network model into a hardware description language using High-Level Synthesis (HLS); then, utilizing the FPGA's parallel architecture to implement dataflow-driven operations for core operators such as convolution and matrix multiplication; and finally, achieving real-time processing of spectral data through deep pipeline technology and memory access optimization. In particular, for feature vectors extracted using the windowing method, fully parallel computation of multilayer perceptron (MLP) networks can be implemented on the FPGA, with single inference latency controllable in the microsecond range.

[0051] For GPU platforms, deployment is suitable for server-side analysis scenarios requiring high throughput, fully leveraging their massively parallel computing advantages. Through CUDA or OpenCL programming models, operations such as convolution and matrix transformations in full-spectrum analysis are mapped to thousands of computing cores for parallel execution. Data locality is optimized using GPU's shared memory and register resources, reducing global memory access latency. For self-attention mechanisms in potential Transformer model architectures, warp-shuffle instructions and Tensor Cores can be used to accelerate matrix multiplication operations, significantly speeding up model inference.

[0052] For the NPU platform, a dedicated acceleration is deployed for edge devices, leveraging its customized architecture for neural network computation. Through dedicated instruction sets and computing units, high-efficiency energy-saving computations are achieved for network layers such as CNN and LSTM. This invention optimizes the operator fusion and memory layout of the neural network computation graph within the fixed-function architecture of the NPU to maximize hardware performance.

[0053] The first and second inversion models are deployed as independent model files on the processor. The above deployment strategy can be summarized into two typical architectures, such as... Figure 10 As shown. Figure 10 (a) demonstrates the embedded processor multi-model deployment strategy for single-task neural networks, which involves deploying and scheduling multiple independent models (such as window method models and full spectrum method models) separately; Figure 10 (b) illustrates the embedded processor single-model deployment strategy for multi-task neural networks, which employs a single multi-task model capable of simultaneously outputting nuclide types and abundances. Both architectures aim to achieve real-time online analysis.

[0054] Regardless of the hardware solution used, a unified software inference interface was developed to receive energy spectrum data from the multichannel pulse amplitude analyzer, call the corresponding optimization model, and output the nuclide analysis results.

[0055] The real-time online inference engine is the core software component of the system, responsible for loading the optimized model, managing computing resources, and executing efficient inference. The engine employs a multi-threaded pipeline design: the main thread receives energy spectrum data from the gamma detector; the preprocessing thread performs data format conversion and normalization (or feature extraction); the inference thread calls the AI ​​processor to perform forward computation of the model; and the post-processing thread parses the model output to generate results on the types and abundance of nuclides. Furthermore, the system possesses self-monitoring and fault-tolerance capabilities, monitoring the detector status, neutron source intensity, and AI processor load in real time. Upon detecting an anomaly, it can automatically trigger an alarm or switch to a backup model. This design ensures the long-term stable operation of the system in complex industrial environments.

[0056] The deployment strategy is as follows: the first and second inversion models can be deployed as independent model files on the processor. The system intelligently schedules and calls the models according to preset analysis requirements (if speed is emphasized, the first inversion model of method one is called; if accuracy is emphasized, the second inversion model of method two is called).

[0057] Step S3, Result Output After completing model inference, the embedded AI processor generates structured inference results. These results include information on the types of nuclides identified in the sample and their predicted concentrations (usually expressed as percentages, ppm, etc.). The results are then output in real-time to host computer monitoring software, databases, or industrial control systems via communication interfaces (such as Ethernet, RS485, CAN bus, etc.), thus completing a fully automated process from physical signal acquisition and intelligent processing to information output. This process represents a significant leap for PGNAA technology from traditional offline, intermittent analysis to online, continuous, and intelligent inversion.

[0058] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), 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 apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A prompt gamma-ray spectroscopy nuclide information inversion method using an artificial neural network, characterized by, Includes the following steps: S1. Energy spectrum acquisition: The sample to be tested is irradiated with neutrons, and the transient gamma rays emitted by the sample are collected by a gamma detector to form the transient gamma energy spectrum to be tested. S2. Data Processing and Model Inference: Based on the preset analysis requirements, select any of the following methods to process the measured transient gamma spectrum, and input the processing results into the corresponding deployed inversion model for inference to obtain the inference results: Method 1: Extract window data corresponding to the preset nuclide characteristic energy from the transient gamma energy spectrum to be tested, form a feature vector, and input the feature vector into the first inversion model for inference; the first inversion model is a pre-trained lightweight machine learning model deployed in an embedded AI processor; Method 2: Perform energy scaling and count rate normalization preprocessing on the transient gamma spectrum to be tested to obtain standardized full-spectrum data, and input the standardized full-spectrum data into the second inversion model for inference; the second inversion model is a pre-trained deep learning model deployed in an embedded AI processor; S3. Result Output: Based on the reasoning results, output the information on the types and contents of nuclides in the sample to be tested in real time.

2. The method for inverting transient gamma-ray spectral nuclide information using an artificial neural network according to claim 1, characterized in that, In the first method, the extraction of window data corresponding to the preset nuclide feature energy includes: taking the peak position corresponding to the preset nuclide feature energy as the center, determining the window width according to the energy resolution of the gamma detector, and calculating the net peak count after subtracting the background within the window, and the feature vector is composed of multiple net peak counts.

3. The method for inverting transient gamma-ray spectral nuclide information using an artificial neural network according to claim 1, characterized in that, In the second method, the preprocessing includes: The transient gamma spectrum to be measured is calibrated with energy, and its channel address is converted into the corresponding energy value. The count rate is normalized by calibrating the energy spectrum, and the counts of each channel are divided by the measurement time to obtain the standardized full spectrum data.

4. The method for inverting transient gamma-ray spectral nuclide information using an artificial neural network according to claim 1, characterized in that, The lightweight machine learning model is one of the following: artificial neural network, random forest, support vector machine, or XGBoost model.

5. The method for inverting transient gamma-ray spectral nuclide information using an artificial neural network according to claim 1, characterized in that, The deep learning model is a one-dimensional convolutional neural network, a long short-term memory network, a Transformer neural network, or a hybrid neural network model that combines one-dimensional convolutional neural networks and long short-term memory networks.

6. A method for inverting transient gamma-ray spectral nuclide information using an artificial neural network according to any one of claims 1 to 5, characterized in that, At least one of the first inversion model and the second inversion model undergoes operator fusion and weight quantization processing before deployment.

7. A method for inverting transient gamma-ray spectral nuclide information using an artificial neural network according to any one of claims 1 to 5, characterized in that, The embedded AI processor is a neural network processor, a field-programmable gate array (FPGA), or an embedded graphics processor.

8. The method for inverting transient gamma-ray spectral nuclide information using an artificial neural network according to claim 1, characterized in that, The preset analysis requirements include: When focusing on speed analysis, choose method one. When focusing on analysis accuracy, choose method two.

9. The method for inverting transient gamma-ray spectral nuclide information using an artificial neural network according to claim 1, characterized in that, The first inversion model and the second inversion model are deployed as independent models in the embedded AI processor, and the corresponding model is called according to the preset analysis requirements.

10. The method for inverting transient gamma-ray spectral nuclide information using an artificial neural network according to claim 1, characterized in that, The second inversion model is a multi-task neural network model, which can simultaneously output the nuclide classification results and the nuclide content regression results.

Images