Radiation particle identification method based on improved ResNet-18 network and CIS transient response technology

Abstract

The invention discloses a radiation particle identification method based on an improved ResNet-18 network and a CIS transient response technology, and the method comprises the steps: collecting a CIS dark field image in a radiation environment, and generating a multi-modal data set according to neutron, proton and heavy ion radiation experiment sample images; a transient response image data set is obtained through theoretical calculation based on radiation analog simulation software; preprocessing the transient response image data set, and marking particle type, energy and angle information of each sample image as a label of multi-task learning; a pre-trained ResNet-18 model is used, an input layer is modified to adapt to the image size, and a channel attention module is added; and carrying out compression quantification on the model by adopting an optimizer in combination with gradient cutting through a self-adaptive category weight adjustment strategy. By constructing the intelligent feature extraction network, the system realizes automatic identification of transient response geometric features, and breaks through the bottleneck that a traditional method depends on artificially defined features.

Description

Technical Field

[0001] This invention belongs to the field of radiation detection technology, and particularly relates to a radiation particle identification method based on an improved ResNet-18 network and CIS transient response technology. It is applicable to the identification of complex radiation source parameters in nuclear reactor monitoring, space radiation environment detection, and high-energy physics experimental devices. Background Technology

[0002] Traditional radiation detection methods primarily rely on devices such as scintillator detectors, semiconductor detectors (e.g., HPGe), or gas ionization chambers. While these devices offer high accuracy, they also have limitations. For instance, a complete detection system using a scintillator detector and photomultiplier tube is expensive; HPGe detectors require liquid nitrogen cooling, making maintenance complex. Furthermore, these devices are functionally limited, typically used to detect radiation intensity or particle energy, and cannot perform particle type classification or incident angle inversion, making them unsuitable for detecting complex integrated radiation fields. Additionally, traditional detector data processing relies mainly on offline analysis, hindering real-time location and dynamic tracking of radiation sources.

[0003] CMOS image sensors, with their advantages of low power consumption, high integration, fast response, and strong environmental adaptability, have been widely used in aerospace, medical imaging, and nuclear industry monitoring. Of particular note is their sensitivity to radiated particles, which gives them unique application value in the field of radiation detection. Research shows that the transient response induced by high-energy particles in CMOS image sensors exhibits significant geometric characteristics, and their morphological parameters are strongly correlated with the type, energy, and incident angle of the radiated particles.

[0004] Traditional radiation detection technologies primarily rely on devices such as scintillator detectors, semiconductor detectors (e.g., high-purity germanium HPGe detectors), or gas ionization chambers. While these devices offer high detection accuracy, they suffer from several limitations: First, scintillator detection systems require photomultiplier tubes, leading to high overall costs; second, HPGe detectors require liquid nitrogen cooling, significantly increasing maintenance complexity; and more importantly, traditional devices are functionally limited, only capable of detecting radiation intensity or particle energy, unable to perform particle type classification or incident angle inversion, thus failing to meet the multi-parameter detection requirements of complex radiation fields. Furthermore, the offline data analysis mode of traditional systems also restricts the real-time location and dynamic tracking capabilities of radiation sources. Summary of the Invention

[0005] Objective: This invention aims to provide a radiation particle identification method based on an improved ResNet-18 network and CIS transient response technology. By constructing an intelligent feature extraction network, the system achieves automatic identification of transient response geometric features, overcoming the bottleneck of traditional methods that rely on manually defined features. This solution significantly reduces equipment complexity and maintenance costs while ensuring detection accuracy, providing a cost-effective technical path for multi-dimensional parameter monitoring of radiation fields.

[0006] Technical solution: The present invention provides a radiation particle identification method based on an improved ResNet-18 network and CIS transient response technology, comprising the following steps:

[0007] Step 1: Acquire dark-field images of CIS under radiation environment, generate multimodal datasets based on neutron, proton and heavy ion irradiation experimental sample images; and obtain transient response image datasets through theoretical calculations based on radiation simulation software.

[0008] Step 2: Preprocess the transient response image dataset and label each sample image with particle type, energy, and angle information as labels for multi-task learning;

[0009] Step 3: Using a pre-trained ResNet-18 model, modify the input layer to fit the image size and add a channel attention module;

[0010] Step 4: Adaptive class weight adjustment strategy, AdamW optimizer combined with gradient clipping, and weighted average SWA introduced at the end of training to improve generalization performance. Layered unfreezing and channel attention regularization are used to adapt to the training requirements of small sample ROI subgraphs.

[0011] Step 5: Compress and quantize the ResNet-18 model, deploy it to the FPGA platform, and realize real-time inference prediction of radiation transient response.

[0012] Furthermore, step 1 specifically involves: the sensitive region of a CIS pixel unit is made of silicon material. The energy required to generate an electron-hole pair in silicon is 3.6 eV. Therefore, the charge number inside each pixel unit during radiation simulation is calculated. According to the definition of the noise model in EMVA1288, the noise intensity DN of each pixel is calculated. DN is a dimensionless coefficient used to describe the gray value of noise in the image. Its calculation formula is as follows:

[0013] In the formula, K represents the gain coefficient of CIS;

[0014] By sampling all pixel units in the calculated pixel array, the distribution of noise DN can be obtained, and the transient response image caused by radiation can be plotted.

[0015] Furthermore, in step 2, the preprocessing includes cutting, rotating, and translating.

[0016] Furthermore, in step 3, a channel attention module (SE Block) is added to the end of the path of each residual block in the ResNet-18 model.

[0017] Furthermore, step 4 specifically involves: using an improved Focal Loss function to dynamically adjust the weights for positive-to-negative sample ratios greater than 1:10, defined by the formula: ; in, For category weight coefficients, By focusing parameters on difficult samples and increasing the loss contribution of hard-to-classify samples, the overfitting tendency of simple samples is suppressed. This represents the model's predicted probability for each sample class; it is dynamically adjusted based on the class of each batch during training. When the proportion of positive samples is below a threshold, their weight coefficient is automatically increased; The AdamW optimizer is used to decouple weight decay from gradient update. When processing high dynamic range images, a gradient norm threshold is set. At the end of training, random weight averaging (SWA) is enabled to move the model parameters by average to improve generalization.

[0018] Further, step 5 specifically involves: first, converting the FP32 floating-point model to INT8 fixed-point format to reduce model size and computational resource consumption; then, using Matlab Coder to convert the quantized model into fixed-point C++ code; finally, using the High-Level Synthesis tool to convert the C++ code into Verilog / VHDL. Finally, the code is programmed onto the FPGA platform for deployment.

[0019] Furthermore, the High-Level Synthesis tool used is Xilinx Vitis HLS; the FPGA platform includes Xilinx Zynq.

[0020] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0021] The present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0022] The present invention also discloses a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method of the present invention.

[0023] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0024] This invention innovatively proposes a radiation particle identification method based on an improved ResNet-18 network and CIS transient response technology. Its core innovation lies in combining the physical mechanism of radiation with an efficient deep learning model, which enables real-time detection of radiation particles such as neutrons, protons, and heavy ions.

[0025] This invention embeds a channel attention (SE) module into an improved ResNet-18 model and combines it with a sample dataset theoretically calculated using Geant4 3D Monte Carlo simulation modeling software. This enables the model to identify the morphological characteristics of the transient response of radiated particles in a CIS (Continuous Induction Sensor). For example, in nuclear reactor monitoring scenarios, it can distinguish the transient response characteristics of neutrons in different energy ranges in real time. Furthermore, compared to the high cost and complex maintenance of traditional radiation detection equipment, this invention utilizes a commercially available CIS sensor and an FPGA embedded solution, combined with a lightweight model to achieve radiation particle detection, resulting in significant cost savings.

[0026] This invention provides innovative solutions for fields such as nuclear industry, aerospace, and medicine through a technical closed loop of "physical mechanism + deep learning + embedded deployment", and has broad market prospects.

[0027] This invention utilizes low-cost hardware alternatives: it captures transient responses using lower-cost commercial CMOS image sensors, replacing dedicated detectors. By combining convolutional neural networks and deep learning models, it achieves accurate detection and classification of radiation particle sources. Furthermore, by lightweighting the model, the trained model is deployed to a low-power embedded platform (such as an FPGA), enabling real-time analysis of CIS output images and providing feedback on the type and energy range of radiation particles. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0029] Figure 2 A simplified pixel unit structure for simulation purposes.

[0030] Figure 3 This is a CIS area array model used for simulation.

[0031] Figure 4 This is the unmodified ResNet-18 model framework.

[0032] Figure 5 This is the improved SE-ResNet-18 model framework.

[0033] Figure 6 This is a transient radiation response image with morphological characteristics.

[0034] Figure 7 These are transient response data samples preprocessed after theoretical calculations. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0036] Figure 1 The following is a flowchart illustrating the implementation of the radiation particle identification method based on the improved ResNet-18 network and CIS transient response technology of this invention, including the following steps:

[0037] S1. Data acquisition involves a mixture of experimental data and simulation data. The simulation data is generated using the three-dimensional Monte Carlo simulation software Geant4. In Geant4, a three-dimensional physical model of CIS is constructed, and irradiation experiments of different radiating particles are simulated. The tracks of secondary charged particles and the energy deposition distribution are tracked. Based on the definition of the noise model of CIS in the EMVA1288 standard, the transient response imaging generated by radiation is calculated.

[0038] S2. Experimental data acquisition in data acquisition mainly involves acquiring dark-field imaging of CIS under proton accelerators, neutron sources, reactors, and heavy ion devices. The exposure time of CIS is set to less than 1 second to record transient response imaging under different radiation environments.

[0039] S3. Merge simulation images and experimental data and preprocess them, crop the dataset for deep learning model training, and label the type of radiating particles, energy range, and recoil angle.

[0040] S4. Construct a deep learning model. By using an improved ResNet-18 model, add a channel attention module (SE Block) after the residual block to enhance the sensitivity to the spatial features of the transient response.

[0041] S5. The image classification model is trained using the Focal Loss function, AdamW optimizer, gradient clipping, and random weight SWA strategy introduced at the end of training.

[0042] S6. Using Matlab to calibrate the dataset, a fixed-point model is generated. The quantized model is then converted into fixed-point C++ code using Matlab Coder and integrated into the Xilinx Zynq platform. This enables lightweight deployment of the model and ultimately achieves real-time particle detection.

[0043] Specifically, in step S1, the simulation data is generated. The steps for generating the simulation data are as follows:

[0044] 1) Figure 2 A simplified model of the CIS pixel unit, using a stacked material structure. A simplified 3D pixel unit model is constructed in Geant4 based on the dimensions, material parameters, doping concentration, and other information from the pixel unit model. A 3D area array model of the CIS is then constructed according to the pixel arrangement.

[0045] 2) Set the parameters of particle sources such as neutrons, protons, and heavy ions with different energies, conduct irradiation simulation experiments, and track the tracks of secondary particles and the energy deposition distribution along the way.

[0046] 3) The sensitive region of a CIS pixel unit is mainly made of silicon. The energy required to generate an electron-hole pair in silicon is 3.6 eV. Therefore, the charge number inside each pixel unit during radiation simulation can be calculated. According to the EMVA1288 definition of the noise model, the noise intensity DN of each pixel can be calculated. DN is a dimensionless coefficient used to describe the grayscale value of noise in the image. Its calculation formula is:

[0047]

[0048] In the formula, K represents the gain coefficient of CIS.

[0049] 4) By sampling all pixel units in the calculated pixel array, the distribution of noise DN can be obtained, and the transient response image caused by radiation can be drawn.

[0050] Specifically, in step S3, the simulation data is generated. The steps for generating the simulation data are as follows:

[0051] 1) The transient response generated by radiation is denoised to a certain extent, because DN needs to reach a certain threshold to be displayed as noise in the image. In this invention, the noise threshold is set to three times the average dark signal DN, and this threshold is related to the type of CIS.

[0052] 2) Based on the dynamic threshold segmentation in the previous step, locate the transient response region, such as... Figure 6 As shown, the cropped sub-images are organized into a dataset for subsequent model training, such as... Figure 7 As shown.

[0053] Specifically, in step S4, the simulation data is generated. The steps for generating the simulation data are as follows:

[0054] Using a pre-trained ResNet-18 model, the input layer was modified to fit the image size, and a channel attention module (Squeeze-and-Excitation, SE Block) was added. Figure 4 For the original deep learning model, Figure 5 This is the improved deep learning model. The main improved parameters are shown in Table 1 below:

[0055] Table 1

[0056]

[0057] Specifically, in step S5, the simulation data is generated. The steps for generating the simulation data are as follows:

[0058] An improved Focal Loss function is used to dynamically adjust the weights for positive-to-negative sample ratios greater than 1:10. The formula is defined as follows:

[0059]

[0060] in For category weight coefficients, The parameters are focused on difficult samples, and the overfitting tendency of easy samples is suppressed by increasing the loss contribution of hard-to-classify samples. The parameters are dynamically adjusted according to the class of each batch during training. When the proportion of positive samples is low, their weight coefficient is automatically increased.

[0061] The AdamW optimizer is employed to decouple weight decay from gradient updates, avoiding parameter update bias caused by decoupling in traditional Adam optimizers. When processing high dynamic range images (e.g., nuclear reactor monitoring scenarios), a gradient norm threshold is set to prevent gradient explosion. Stochastic weight averaging (SWA) is enabled at the end of training to move the model parameters and improve generalization.

[0062] Specifically, in step S6, the simulation data is generated. The steps for generating the simulation data are as follows:

[0063] First, the FP32 floating-point model is converted to INT8 fixed-point format to reduce model size and computational resource consumption. Then, Matlab Coder is used to convert the quantized model into fixed-point C++ code. Finally, a High-Level Synthesis tool (such as Xilinx Vitis HLS) is used to convert the C++ code into Verilog / VHDL. Finally, the code is programmed onto an FPGA platform for deployment (such as Xilinx Zynq).

Claims

1. A radiation particle identification method based on an improved ResNet-18 network and CIS transient response technology, characterized in that, Comprise the following steps: Step 1, collect CIS dark field images in a radiation environment, generate a multi-modal data set according to neutron, proton and heavy ion irradiation sample images; and obtain a transient response image data set through theoretical calculation based on radiation simulation software; Step 2, pre-process the transient response image data set, and label the particle type, energy and angle information of each sample image as the label of multi-task learning; Step 3, use a pre-trained ResNet-18 model, modify the input layer to adapt to the image size, and add a channel attention module; Step 4, through an adaptive class weight adjustment strategy, use an AdamW optimizer combined with gradient clipping, and introduce weight averaging SWA at the end of training to improve the generalization performance, through hierarchical unfreezing and channel attention regularization, adapt to the training requirements of small sample ROI subgraph; Step 5, compress and quantize the ResNet-18 model, deploy it to the FPGA platform, and realize real-time inference prediction of radiation transient response.

2. The method of claim 1, wherein the method is based on an improved ResNet-18 network and a CIS transient response technique. The step 1 is specifically: the sensitive area of the CIS pixel unit is silicon material, and the energy required for generating a pair of electron-hole pairs in the silicon material is 3.6eV, so that the number of charges in each pixel unit inside the CIS during radiation simulation is calculated According to the definition of the noise model of EMVA1288, the noise intensity DN of each pixel is calculated, DN is a dimensionless coefficient, and is used to describe the gray value of the noise on the image, and the calculation formula is: ； In the formula, K represents the gain coefficient of CIS; By sampling all pixel units in the calculated pixel array, the distribution of noise DN can be obtained, and the transient response image caused by radiation can be drawn.

3. The method of claim 1, wherein the method is characterized by, In step 2, the pre-processing includes cropping, rotating and translating.

4. The method of claim 1, wherein the method is characterized by, In step 3, a channel attention module SE Block is added at the end of the path of each residual block of the ResNet-18 model.

5. The method of claim 1, wherein the method is based on an improved ResNet-18 network and a CIS transient response technique. Step 4 is: using an improved Focal Loss function, dynamically adjusting the weight when the positive and negative sample ratio is greater than 1:10, and the formula is defined as: ； wherein, is a class weight coefficient, is a hard sample focus parameter, by increasing the loss contribution of difficult classification samples, inhibiting the overfitting tendency of simple samples, is the predicted probability of the model for the class of the sample; according to the class of each batch in the training process , the positive sample is automatically raised when the positive sample is lower than the threshold. Using the AdamW optimizer, decoupling weight decay and gradient update, setting the gradient norm threshold when processing high dynamic range images; enable random weight averaging SWA at the end of training to slide the model parameters to improve generalization.

6. The method of claim 1, wherein the method is based on an improved ResNet-18 network and a CIS transient response technique. Step 5 is: first, convert the FP32 floating point model to INT8 fixed point format; use Matlab Coder to convert the quantized model to fixed point C++ code; use High-Level Synthesis tools to convert C++ code to Verilog / VHDL; finally, burn to FPGA platform deployment.

7. The method of claim 6, wherein the method is characterized by, The High-Level Synthesis tool selects Xilinx Vitis HLS; the FPGA platform includes Xilinx Zynq.

8. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, causes the processor to perform the method of any one of claims 1 to 7. The processor executes the computer program to implement the steps of the method of claim 1.

9. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method of claim 1.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method of claim 1.

Images