Decommissioned nuclear graphite potential prediction method and system based on multi-modal vision

Abstract

The invention belongs to the technical field of nuclear graphite decommissioning treatment, and particularly relates to a decommissioning nuclear graphite potential prediction method and system based on multi-modal vision, and the method comprises the steps: collecting the multi-modal image data of the surface morphology, temperature distribution, radiation intensity and internal structure of decommissioning nuclear graphite, carrying out the preprocessing, and carrying out the prediction of the potential of the decommissioning nuclear graphite. A deep learning model is utilized to extract key features, a feature fusion technology is adopted to generate a multi-dimensional feature map, a key region is focused in combination with an attention mechanism, the attention of the model on the features is enhanced, then a feature extraction network is constructed through a U-Net image segmentation algorithm for training, and the optimized model can display the state and prediction result of nuclear graphite in real time. And an appropriate waste treatment scheme is selected in an assisted manner. According to the method, the limitation of a traditional single detection mode is overcome, the potential of the decommissioning nuclear graphite is accurately predicted through multi-dimensional data fusion and intelligent analysis, and the safety of the nuclear reactor decommissioning process is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of nuclear graphite decommissioning treatment, and in particular, relates to a method and system for predicting the potential of retired nuclear graphite based on multimodal vision. Background Art

[0002] Graphite for Nuclear Applications is a commonly used material in nuclear reactors, particularly as a neutron moderator. The primary function of a neutron moderator is to slow down the fast neutrons produced by nuclear fission into thermal neutrons, allowing them to undergo fission reactions with nuclear fuel (such as uranium-235 or plutonium-239), thereby generating more energy. In certain types of nuclear reactors, such as heavy water reactors and graphite reactors, graphite is widely used to moderate neutrons and improve reactor efficiency.

[0003] However, during normal nuclear reactor operation, nuclear graphite absorbs a significant amount of neutrons, generating radioactive substances in the process. These substances may have environmental and human health impacts after reactor decommissioning. Therefore, accurately predicting the potential energy (including radioactive and heat release) in decommissioned nuclear graphite has become a critical issue in nuclear waste disposal and environmental safety.

[0004] The potential of decommissioned nuclear graphite is often triggered by changes in its environment, such as elevated temperatures, mechanical impact, or chemical corrosion. These changes can cause the graphite to release radioactive substances and generate heat, all of which increase the potential risk of nuclear hazards. Traditional methods for analyzing the potential of decommissioned nuclear graphite, such as sample collection and analysis, rely primarily on detecting radioactive elements in decommissioned nuclear graphite, typically using a gamma spectrometer to measure the energy and intensity of characteristic gamma rays from different radionuclides. This method can provide some nuclear radiation information, but it has the following problems:

[0005] Limitations of a single testing method: Traditional methods for predicting the potential of nuclear graphite primarily rely on gamma-ray analysis to analyze radioactive materials. However, this approach focuses solely on radioactive elements, ignoring other factors that may affect the state of decommissioned nuclear graphite. For example, changes in chemical composition, physical state, and even microstructural defects can affect graphite stability and safety. Gamma-ray testing alone cannot fully capture the multidimensional information about nuclear graphite.

[0006] Lack of intelligent data analysis and fusion: Existing radiation prediction methods rely heavily on traditional manual analysis and single-data testing, lacking intelligent data processing and fusion mechanisms. This prevents comprehensive consideration of various data types, such as chemical composition, structural damage, and radiation levels. This analysis often fails to accurately predict the potential of decommissioned nuclear graphite, potentially leading to errors in the assessment of the radioactive content within, which in turn impacts the comprehensiveness and safety of nuclear waste disposal.

[0007] Therefore, in view of the limitations of traditional nuclear graphite decommissioning potential prediction methods, it is urgent to develop a new prediction technology based on multi-dimensional data fusion and intelligent analysis. Summary of the Invention

[0008] In response to the above problems, the present invention provides a method and system for predicting the potential of retired nuclear graphite based on multimodal vision, which solves the limitations of the traditional method's single detection method and the lack of multi-dimensional data fusion and intelligent analysis. By combining high-resolution appearance images, gamma-ray imaging, infrared visual sensing technology and nuclear graphite internal ultrasonic detection technology, it comprehensively evaluates the physical, chemical and radiation properties of nuclear graphite, accurately predicts its potential risks, and provides a scientific basis for nuclear waste treatment and environmental monitoring.

[0009] In one aspect, the present invention provides a method for predicting the potential of retired nuclear graphite based on multimodal vision, the method comprising:

[0010] Step 1: Collect multimodal image data of the surface morphology, temperature distribution, radiation intensity, and internal structure of retired nuclear graphite to obtain comprehensive nuclear graphite status information;

[0011] Step 2: Preprocess the collected multimodal image data;

[0012] Step 3: Use the deep learning model to extract key features from the multimodal image data after preprocessing, and integrate the extracted key features into a multi-dimensional feature map through feature fusion technology;

[0013] Step 4: Select the corresponding loss function based on the decommissioned nuclear graphite potential prediction task and the collected data characteristics;

[0014] Step 5: Introduce the attention mechanism to enhance the model's attention to the key features of nuclear graphite by focusing on the key areas in the image;

[0015] Step 6: Build a feature extraction network and train the U-Net model by combining the multi-dimensional feature map fused in step 3, the loss function defined in step 4, and the attention mechanism in step 5.

[0016] Step 7: Evaluate the trained model and optimize the model based on the evaluation results;

[0017] Step 8: The status and prediction results of nuclear graphite are displayed in real time through the human-computer interaction interface, and a waste treatment plan is selected based on the prediction information.

[0018] Furthermore, in step 4, the retired nuclear graphite potential prediction task includes at least a regression task, a classification task, and an image segmentation task, wherein the regression task is used to predict the potential risks of retired nuclear graphite, the classification task is used to classify the potential risks of retired nuclear graphite, and the image segmentation task is used to segment cracks, hot spots, and high-risk radiation areas in retired nuclear graphite images.

[0019] Furthermore, the loss function selected for the regression task is the mean square error or Huber loss function, the loss function selected for the classification task is the cross entropy loss function, and the loss function selected for the image segmentation task is the Dice loss function.

[0020] Furthermore, in step 5, the key areas include but are not limited to surface cracks and defect areas, temperature hotspot areas, high radiation areas and internal structural defect areas; the attention mechanism includes a channel attention mechanism, a spatial attention mechanism, a combined channel attention mechanism and a spatial attention mechanism.

[0021] Furthermore, step 6 includes:

[0022] Step 6.1: Select U-Net as the training model and combine it with ResNet as the feature extraction network;

[0023] Step 6.2: Use ResNet to extract features from the multi-dimensional feature map in step 3 and fuse them to form a multimodal fused feature map. Divide each fused feature map into multiple patches, and input each patch into U-Net for processing. Apply the channel attention mechanism after each layer of the U-Net network module. Apply the combined channel attention mechanism and spatial attention mechanism after the last layer of the U-Net to further enhance the expression of spatial and channel features.

[0024] Step 6.3: Select a combined loss function, combining Dice loss and cross entropy loss as the loss function of the U-Net model;

[0025] Step 6.4: Use the Adam optimizer to adjust the network parameters and gradually update the weights of the U-Net model according to the changes in the loss function during training. Repeat the training and monitor the changes in the loss function until the preset number of training rounds is reached or the loss function changes tend to be stable and the model converges.

[0026] Furthermore, in step 7, the trained model is evaluated by analyzing the model's accuracy and intersection-over-union (IoU) metrics; the model is optimized based on the evaluation results, including but not limited to hyperparameter adjustment, data augmentation, regularization, and model fine-tuning.

[0027] Furthermore, step 3 includes:

[0028] Step 3.1: Use a convolutional neural network to extract surface defect features, temperature gradient features, radiation distribution features, and internal structure features from the preprocessed multimodal image data;

[0029] Step 3.2: Normalize the feature maps extracted from different modality images;

[0030] Step 3.4: Set the corresponding fusion weight according to the influence of different feature maps on the nuclear graphite potential prediction results;

[0031] Step 3.5: Perform weighted fusion of the feature maps of each modality through convolutional neural network to obtain a multi-dimensional feature map.

[0032] Furthermore, step 1 collects multimodal image data through a high-resolution camera, an infrared thermal imager, a gamma-ray camera, and an ultrasonic flaw detector; the multimodal image data includes surface images of retired nuclear graphite, infrared thermal images, radionuclide distribution images and radiation intensity distribution images, and ultrasonic flaw detection images.

[0033] Furthermore, step 2 preprocesses the collected image data including denoising, enhancement, calibration, stitching and normalization.

[0034] On the other hand, the present invention also provides a retired nuclear graphite potential prediction system based on multimodal vision, the system comprising:

[0035] Multimodal image acquisition module, used to collect multimodal image data of the surface morphology, temperature distribution, radiation intensity and internal structure of retired nuclear graphite to obtain comprehensive nuclear graphite status information;

[0036] A data preprocessing module, used for preprocessing the collected multimodal image data;

[0037] A feature extraction and fusion module is used to extract key features from the preprocessed image using a deep learning model and integrate the key features into a multi-dimensional feature map through feature fusion technology;

[0038] A loss function selection module is used to select a corresponding loss function based on the decommissioned nuclear graphite potential prediction task and the characteristics of the collected data;

[0039] The attention mechanism module is used to introduce the attention mechanism to enhance the model's attention to the key features of nuclear graphite by focusing on the key areas in the image;

[0040] A feature extraction network module is used to train a U-Net model by combining the multidimensional feature map fused in the feature extraction and fusion module, the loss function defined in the loss function selection module, and the attention mechanism in the attention mechanism module;

[0041] A model evaluation and optimization module is used to evaluate the trained model and optimize the model based on the evaluation results;

[0042] The human-computer interaction interface module is used to display the status and prediction results of nuclear graphite in real time, and provide options for waste treatment solutions based on the prediction information.

[0043] The beneficial effects of the present invention are:

[0044] First, the present invention's multimodal vision-based method for predicting the energy potential of retired nuclear graphite enables comprehensive assessment and accurate prediction of retired nuclear graphite through multi-sensor data acquisition and the application of deep learning models. The method first acquires comprehensive information on the nuclear graphite's state by collecting multimodal image data, including surface morphology, temperature distribution, radiation intensity, and internal structure, ensuring a full understanding of its potential problems and risks. Subsequently, the collected image data undergoes preprocessing and feature fusion techniques to remove noise and improve data quality. Simultaneously, multiple sensory information is fused into a multidimensional feature map to enhance feature extraction accuracy. This method utilizes a deep learning model to automatically extract key features from the image and, combined with an attention mechanism, enhances focus on important areas, enabling the model to more accurately identify potential defects or abnormal areas in the nuclear graphite. Furthermore, through the training and optimization of a U-Net model, the reliability and accuracy of the prediction results are ensured, thereby providing an accurate prediction of the potential distribution of retired nuclear graphite. Finally, the system displays the nuclear graphite's state and prediction results in real time through a human-computer interface, helping operators quickly understand its internal conditions and select appropriate waste disposal solutions based on the predicted information. This method not only improves the safety assessment capabilities of nuclear graphite, but also reduces manual intervention, improves the accuracy and efficiency of waste disposal, and provides effective technical support for the safe disposal of retired nuclear graphite. It overcomes the problem of the relatively limited potential prediction methods of nuclear graphite in traditional methods.

[0045] Second, in a preferred implementation, step 4 of the present invention ensures accurate and targeted prediction of the potential risk of decommissioned nuclear graphite by selecting appropriate loss functions for regression, classification, and image segmentation. The regression task uses a mean square error or Huber loss function, which helps accurately predict the potential risk of nuclear graphite; the classification task uses a cross-entropy loss function, which effectively classifies potential risks; and the image segmentation task uses a Dice loss function, which accurately segments cracks, hotspots, and high-risk radiation areas, improving the overall prediction accuracy and processing power.

[0046] Third, in a preferred implementation, step 5 of the present invention effectively focuses on key areas in the nuclear graphite image, such as surface cracks and defects, temperature hotspots, high-radiation areas, and internal structural defects, by introducing an attention mechanism. This ensures that the model can more accurately identify potential risk areas. Channel attention mechanisms, spatial attention mechanisms, and their combination further enhance the model's focus on these key features, improving the accuracy and reliability of prediction results. This is particularly true when processing complex images, enabling the model to highlight and precisely locate important information, thereby optimizing the overall prediction effect.

[0047] Fourth, in the preferred implementation, step 6 of the present invention combines U-Net and ResNet to give full play to the advantages of both and realize efficient image segmentation and feature extraction. In step 6, ResNet, as a feature extraction network, can efficiently extract key features in the multi-dimensional feature map, and form a multimodal fusion feature map through fusion, and then divide each feature map into multiple patches, which are input into U-Net for fine processing. The image segmentation capability of U-Net is further enhanced by the channel attention mechanism and the combined channel-space attention mechanism, which effectively improves the expression of spatial and channel features and optimizes the segmentation effect of cracks, hot spots and high radiation areas. By selecting a combined loss function and optimizing the network parameters using the Adam optimizer, the model weights are gradually updated to ensure the convergence of model training and the accuracy of prediction. The combination of this method improves the image segmentation accuracy in the prediction of the potential of retired nuclear graphite, and provides strong technical support for processing complex nuclear graphite images.

[0048] Fifth, in a preferred implementation, step 7 of the present invention comprehensively evaluates the trained model by analyzing its accuracy and intersection-over-union (IoU) metrics to ensure its efficiency and accuracy in predicting the potential of decommissioned nuclear graphite. Optimization based on the evaluation results, including hyperparameter adjustment, data augmentation, regularization, and model fine-tuning, further improves the model's robustness and adaptability, effectively preventing overfitting and enhancing predictive performance. This optimization process not only enhances the model's generalization capabilities but also ensures the reliability and accuracy of predictions in practical applications.

[0049] Sixth, in a preferred implementation, step 3 of the present invention effectively extracts a variety of key features from the pre-processed multimodal image data through a convolutional neural network (CNN), including information such as surface defects, temperature gradients, radiation distribution, and internal structure. These features can comprehensively reflect the state of nuclear graphite and provide a rich data basis for potential prediction. By standardizing the feature maps extracted from different modal images, the differences between various types of data are eliminated, ensuring the consistency and accuracy of feature fusion. In addition, weighted fusion is performed based on the degree of influence of different feature maps on the potential prediction results, so that the model can dynamically adjust the information weight according to the importance of each feature, and finally obtain a multi-dimensional feature map through weighted fusion, thereby improving the accuracy and reliability of nuclear graphite potential prediction.

[0050] Seventh, in a preferred implementation, step 1 of the present invention uses multimodal image data collected by a high-resolution camera, infrared thermal imager, gamma-ray camera, and ultrasonic flaw detector to comprehensively capture information such as the surface morphology, thermal distribution, radiation intensity, and internal structure of decommissioned nuclear graphite, providing rich data support for subsequent potential prediction. Step 2, through pre-processing such as denoising, enhancement, calibration, splicing, and normalization of these image data, effectively improves data quality, reduces noise interference, unifies data formats, and ensures data accuracy and consistency.

[0051] Eighth, the prediction system of the present invention realizes a comprehensive assessment and accurate prediction of retired nuclear graphite through multi-sensor data acquisition and deep learning technology. The system provides comprehensive information on the status of nuclear graphite by collecting multimodal image data such as surface morphology, temperature distribution, radiation intensity and internal structure, ensuring early identification of potential problems. After the data is pre-processed and feature fused, the system can extract and integrate key features, and use deep learning models combined with attention mechanisms to focus on important areas, thereby improving the accuracy and robustness of the model. Through U-Net model training, the system can efficiently predict the potential distribution of nuclear graphite, and continuously improve the prediction accuracy through evaluation and optimization modules. The human-computer interaction interface provides real-time monitoring and prediction results to help operators select appropriate waste treatment solutions based on prediction information, improve the accuracy and safety of nuclear graphite treatment, reduce human intervention and errors, and improve overall work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of a method for predicting the potential of retired nuclear graphite based on multimodal vision according to an embodiment of the present invention;

[0053] Figure 2 4 is a schematic diagram of a system for predicting the potential of retired nuclear graphite based on multimodal vision according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the technical solution of the present application, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0055] The terms "up", "down", "left", "right", "front", and "back" in this application are based on the positional relationships shown in the accompanying drawings. The corresponding positional relationships may vary depending on the drawings, and should not be construed as limiting the scope of protection.

[0056] In this application, the terms "installed," "connected," "connected," "connected," "fixed," etc. should be understood in a broad sense. For example, they can refer to fixed connection, detachable connection, integral connection, mechanical connection, electrical connection, or mutual communication. They can also be directly connected or indirectly connected through an intermediate medium. They can also refer to internal communication between two components or interaction between two components. For those skilled in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0057] The present invention aims to provide a method and system for predicting the potential of retired nuclear graphite based on multimodal vision. The system combines high-resolution appearance images, gamma-ray imaging, infrared visual sensing technology, and nuclear graphite internal ultrasonic detection technology to construct a multi-level visual detection platform. By fusing multiple feature maps, it obtains deeper information, thereby improving the accuracy of the prediction results. In terms of prediction methods, the U-Net image segmentation algorithm is adopted, and the attention mechanism module is introduced. With the help of the encoder-decoder structure, key information in the multimodal feature map is efficiently extracted. This technical architecture can more accurately capture the physical, chemical, and radiation properties of nuclear graphite, providing strong technical support for assessing the potential risks of retired nuclear graphite. Through intelligent prediction, the system can comprehensively analyze the internal potential of nuclear graphite, providing a more scientific and reliable decision-making basis for waste treatment, environmental monitoring, and decommissioning processes, further improving the safety and efficiency of the nuclear reactor decommissioning process.

[0058] Refer to the instruction manual Figure 1 A multimodal vision-based method for predicting the decommissioning potential of nuclear graphite is proposed. This method combines multi-dimensional information fusion with intelligent algorithms to achieve accurate prediction of the decommissioning potential of nuclear graphite. The specific steps are as follows:

[0059] Step 1: Collect multimodal image data of the surface morphology, temperature distribution, radiation intensity, and internal structure of retired nuclear graphite to obtain comprehensive nuclear graphite status information.

[0060] The purpose of Step 1 is to comprehensively capture the surface morphology, thermal conductivity, radiation intensity distribution, and internal structure of retired nuclear graphite through multimodal data acquisition, providing comprehensive raw data support for subsequent potential prediction and analysis. This data will provide foundational information for predictive models, thereby helping to assess the potential risks of retired nuclear graphite and providing a scientific basis for subsequent waste treatment and environmental monitoring.

[0061] Specifically, step 1 collects multimodal data through a high-resolution camera, an infrared thermal imager, a gamma-ray camera, and an ultrasonic flaw detector. The multimodal data includes surface images of retired nuclear graphite, infrared thermal images, radionuclide distribution images and radiation intensity distribution images, and ultrasonic flaw detection images.

[0062] Use a high-resolution camera (e.g., 20MP or higher) to capture the surface of decommissioned nuclear graphite, capturing detailed topography, including microcracks, surface defects, and discoloration. These images provide essential data for subsequent defect analysis. The data includes information on surface topography, cracks, microscopic defects, and wear.

[0063] Infrared thermal imaging cameras are used to monitor the surface and internal temperature distribution of decommissioned nuclear graphite in real time. These cameras can accurately capture temperature variations in different areas, reflecting their heat transfer characteristics. The data includes information such as surface temperature, internal heat distribution, and hotspots.

[0064] Gamma-ray cameras are used for nuclide identification, measuring the gamma-ray signature of radionuclides. Gamma-ray cameras can capture the intensity, distribution, and type of different radiation sources. By analyzing the energy spectrum of different nuclides, they can accurately map the nuclide distribution. The data includes nuclide distribution maps, radiation intensity distribution, and the type of radioactive material.

[0065] Ultrasonic flaw detectors are used to inspect the interior of decommissioned nuclear graphite for structural defects such as cracks and pores. Ultrasonic signals are reflected and refracted within the graphite to determine its structural integrity, generating an image of the internal structure. The data includes internal defect maps, cracks, pores, and structural damage information.

[0066] Step 2: Preprocess the collected multimodal image data.

[0067] The purpose of step 2 is to improve data quality, reduce measurement errors, and provide clear and accurate data for subsequent feature extraction and analysis.

[0068] Specifically, the preprocessing of the collected image data includes denoising, enhancement, calibration, stitching and normalization.

[0069] By removing and enhancing images, we can effectively reduce noise caused by environmental interference or equipment defects, improving image quality and usability. Common denoising methods include median filtering and mean filtering, which help remove unnecessary interference from images. To enhance the expressiveness and diversity of images, data augmentation techniques are also used. Common enhancement methods include geometric transformations (such as rotation and scaling), color transformations (such as brightness and contrast adjustment), and affine transformations. These operations help enrich training data and improve model robustness.

[0070] Temperature calibration is performed on images captured by infrared thermal imagers to ensure accurate temperature data and mitigate the effects of instrumental errors. First, the infrared thermal imager is calibrated to ensure that its temperature measurement range, sensitivity, and accuracy meet requirements. Temperature calibration is performed using a standard object with a known temperature (such as a blackbody radiation source or a heat source of known temperature). During calibration, the temperature measured by the thermal imager is compared with the standard temperature data to ensure the accuracy of the captured image temperature data. Linear or nonlinear calibration methods are used to adjust the temperature value of each pixel in the thermal image to eliminate the influence of instrumental errors and environmental factors, thereby ensuring that the collected temperature data is consistent with the actual temperature. For large areas of decommissioned nuclear graphite surfaces and interiors, multiple thermal images must be stitched together to form a complete thermal image distribution map. Image registration technology is used during stitching to ensure precise alignment of the images by detecting key points and feature points within the images or using edge detection algorithms. Overlapping areas are used to match the stitched images to avoid temperature gaps or discontinuities at the seams. A gradient blending algorithm is used to smooth the seams to ensure visual consistency in the stitched thermal image. The assembled thermal image will show the temperature changes on the surface and inside the entire decommissioned nuclear graphite. To visually display the temperature distribution, heat map technology is used to map temperature values to different color gradients. The temperature color scale clearly identifies abnormal areas and hot spots, allowing intuitive identification of temperature anomalies, hot spots, and possible heat sources. High-resolution temperature data is used to generate thermal imaging distribution maps, which are then post-processed (such as color mapping and temperature contour drawing) to produce easy-to-understand and analyze graphical results, facilitating subsequent evaluation of the thermal conductivity and safety of nuclear graphite.

[0071] Normalization is performed on gamma-ray camera and ultrasound images to eliminate the effects of differences in instruments, measurement conditions, and equipment performance. This allows for uniform values of radiation intensity, defect information, and other information in the images, facilitating subsequent data fusion and analysis. Logarithmic normalization is used for normalization, which is suitable for images with a wide pixel distribution range. This operation scales the pixel values of the image to the range of [0,1], allowing pixel values between different images to be compared on the same scale. The logarithmic normalization formula is:

[0072]

[0073] Where: x represents the pixel value in the original image; x max Indicates the maximum pixel value in the image.

[0074] Gaussian filtering is performed on ultrasound images to reduce noise generated during signal transmission and improve image quality. This helps improve image clarity and reduce noise caused by environmental factors or equipment errors, especially for detecting internal structural defects (such as cracks and pores) in the image. The image is smoothed using a Gaussian function to reduce the impact of high-frequency noise in the image and highlight low-frequency information. The expression of the Gaussian function is:

[0075]

[0076] Where G(x,y) represents the pixel values after Gaussian filtering; σ represents the standard deviation of the Gaussian distribution. A larger standard deviation results in a more pronounced smoothing effect. By adjusting σ, you can optimize the image based on its noise intensity to achieve the best possible noise reduction.

[0077] Step 3: Use the deep learning model to extract key features from the multimodal image data after preprocessing, and integrate the extracted key features into a multi-dimensional feature map through feature fusion technology.

[0078] The purpose of step 3 is to extract surface defects, temperature gradients, radiation distribution, and internal structural features from multimodal images and fuse these features to form a multidimensional feature map, aiming to comprehensively display the potential information of decommissioned nuclear graphite. This approach can more accurately assess the potential risks of nuclear graphite and provide data support for subsequent waste treatment and environmental monitoring.

[0079] Step 3 specifically includes:

[0080] Step 3.1: Use convolutional neural network (CNN) to extract surface defect features, temperature gradient features, radiation distribution features, and internal structure features from the preprocessed multimodal image data.

[0081] The feature extraction in step 3 is obtained using a convolutional network (CNN). The multidimensional feature map is a fusion image of one or more of the following: surface image, thermal map, radiation distribution map, and ultrasound image. Temperature gradient features are extracted from infrared thermal imaging data to capture thermal conductivity information on the surface and interior of nuclear graphite. The temperature distribution in the thermal map is analyzed using a CNN model to obtain temperature variation areas and hotspot locations. Radioactive nuclide distribution features and radiation intensity distribution features are extracted from gamma-ray images. CNN is used to accurately analyze the radiation distribution of different nuclides to help identify potential high-radiation areas. Structural features within the nuclear graphite, such as cracks, pores, and other defects, are extracted from ultrasound images. Through feature extraction from ultrasound images, potential structural damage is identified and located.

[0082] Step 3.2: Normalize the feature maps extracted from images of different modalities.

[0083] In order to ensure that the features of different modal images can be effectively fused, the extracted features need to be normalized. Step 3.2 can ensure that the data range of each feature map is consistent, so that features from different sources can be effectively fused at the same scale. The normalization steps for each feature are as follows:

[0084] Temperature gradient features: Normalize the temperature data in infrared thermal images to ensure that the temperature values are within the same dimensional range. Radiation distribution features: Normalize the radiation intensity in gamma-ray images to eliminate dimensional differences in radiation intensity. Surface defect features: Standardize the range of surface image features through statistical processing. Ultrasonic image features: Normalize internal structural features (such as cracks and pores) in ultrasound images to ensure consistency with other modalities during feature fusion.

[0085] Step 3.3: Resize the feature maps of different modalities to the same size.

[0086] For effective feature fusion, a unified size is used to ensure that the subsequent fusion process can consistently process each pixel. For example, if the size of one modality image is 256×256 and the size of the other modality image is 512×512, the larger image needs to be resized to the same size by downsampling or the smaller image needs to be expanded by interpolation.

[0087] Step 3.4: Set the corresponding fusion weights according to the degree of influence of different feature maps on the nuclear graphite potential prediction results.

[0088] Infrared thermal map (temperature gradient characteristics): Temperature change areas and hot spots are very important for the potential risk assessment of nuclear graphite, so they are assigned a higher weight. Gamma ray map (radiation intensity distribution characteristics): The distribution of radioactive areas directly affects safety, so they are assigned a higher weight. Surface defect map (surface defect characteristics): Surface defects have a smaller impact when assessing potential, but their contribution to the overall risk assessment still needs to be considered and a medium weight is assigned. Ultrasound image (internal structure characteristics): Ultrasound images provide key information on the internal structure of nuclear graphite, but compared with radiation and temperature information, their contribution to potential risks is relatively small, so they are given a lower weight.

[0089] Step 3.5: Perform weighted fusion of the feature maps of each modality through convolutional neural network to obtain a multi-dimensional feature map.

[0090] The feature map of each modality image is scaled according to the weights set in step 3.4. That is, each pixel value of each feature map is multiplied by the corresponding weight coefficient. The weighted calculation formula is:

[0091] F fused =w1·F1+w2·F2+……w n ·F n

[0092] Where: F fused Represents a multi-dimensional feature map; F1, F2...F n Represents feature maps from different modalities (such as surface images, temperature images, etc.); w1, w2...w n Indicates the corresponding weight value.

[0093] The weighted fusion feature map integrates information such as surface condition, thermal conductivity, radiation distribution, and internal structural defects, comprehensively reflecting the potential of nuclear graphite. This multi-dimensional feature map provides rich input data for subsequent potential prediction. The surface condition of the weighted fusion feature map includes information such as surface defects, cracks, and wear. Thermal conductivity includes information such as temperature gradients, hot spots, and heat conduction paths. Radiation distribution includes areas of varying radiation intensity and nuclide distribution. Internal structural defects such as internal cracks and pores are identified through ultrasound images.

[0094] Step 4: Select the corresponding loss function based on the decommissioned nuclear graphite potential prediction task and the collected data characteristics.

[0095] The purpose of step 4 is to improve the model's prediction accuracy and generalization ability on specific tasks. The choice of loss function directly affects model optimization during training and can help improve model performance and convergence speed.

[0096] Specifically, the decommissioned nuclear graphite potential prediction task includes at least regression task, classification task and image segmentation task.

[0097] The regression task used to predict the potential risks of decommissioned nuclear graphite (such as radiation risk, thermal damage, etc.) is a continuous value prediction problem. The loss functions selected are the mean squared error (MSE) and the Huber loss function. The mean squared error (MSE) is suitable for regression tasks that predict continuous values. If the result of the potential prediction is a specific value (such as radiation intensity, temperature change, etc.), the MSE is used to measure the difference between the predicted value and the actual value. The Huber loss function is used when the data may contain outliers. It can effectively reduce the impact of outliers on model training and improve the robustness of the prediction.

[0098] Take the Huber loss function as an example: The Huber loss function is a loss function for regression problems. It behaves like a squared loss function when the data point is close to the predicted value, and behaves like an absolute loss function when the data point is far from the predicted value. Therefore, it combines the advantages of both functions to some extent. Its mathematical expression is:

[0099]

[0100] Where y represents the target output of the model (i.e., the value we want to predict). For example, in a regression task, it represents the actual numerical label. f(x) represents the predicted value of the model for the input data x, which is a prediction made on the input data based on the mapping relationship learned by the model. δ represents an adjustable threshold. |yf(x)| represents the difference between the actual value y and the predicted value f(x). represents the square loss part; Represents the linear loss part.

[0101] When the prediction error |yf(x)| is less than δ, square loss is used For calculation, square loss is the standard regression loss, which is suitable for cases with small errors and can better smooth the model output. When the error |yf(x)| is greater than δ, linear loss is used. When performing calculations, linear loss penalizes larger errors less severely, preventing excessive weights from impacting model training. It also handles abnormal values or outliers more robustly. The value of δ is typically determined through cross-validation to balance the penalties for small and large errors.

[0102] The advantage of the Huber loss function is that it combines the smoothness of the squared loss function and the robustness of the absolute loss function when approaching the predicted value. Therefore, it is relatively less affected by outliers, making the model more robust. In practical applications, the Huber loss function is often used to fit data with many outliers or regression tasks that are sensitive to outliers.

[0103] The classification task involves classifying the potential risks of decommissioned nuclear graphite into categories such as high risk, medium risk, and low risk. This is a discrete category prediction task. The cross-entropy loss function is chosen as the loss function for this task. The cross-entropy loss function is used in binary or multi-classification tasks to measure the difference between the predicted class probability and the true class label, making it suitable for classification tasks that determine different risk levels.

[0104] Cross-Entropy Loss: This function measures the difference between the probability distribution of the model output and the true probability distribution. It is primarily applicable to classification tasks and works well in both binary and multi-classification tasks. It is often used in conjunction with activation functions in classification tasks to effectively optimize the model's classification performance. Its mathematical formula is:

[0105]

[0106] Where: y i The probability value representing the actual label, usually the true probability of the target class or 0 or 1 (representing the actual class) in a binary classification problem; Represents the probability value predicted by the model, that is, the probability that the predicted value is the i-th category; Represents the logarithm of the model's predicted probability.

[0107] For each sample's i-th class label y i Is a true value, which represents the actual probability that the sample belongs to the class; for classification problems, usually y i =1 means the sample belongs to category i, y i =0 means it does not belong to this category. The predicted probability of each class output by the model is For each sample of class i, is the probability that the model believes that the sample belongs to this category. It is the key part of the cross entropy loss function. By performing logarithmic operations on the predicted probability, the prediction error can be better quantified, especially giving greater penalties to errors with lower probabilities.

[0108] The image segmentation task is used to segment high-risk areas (such as cracks, hot spots, and radiation zones) in decommissioned nuclear graphite images. The Dice loss function is chosen as the loss function for this task. The Dice loss function is particularly effective in image segmentation tasks, especially when dealing with an imbalance of positive and negative samples. It effectively optimizes the model and ensures maximum overlap between the predicted and true regions, making it suitable for segmenting risky areas in nuclear graphite.

[0109] Dice loss function: The Dice loss function is calculated based on the intersection and union of the predicted area and the true area. This allows the model to focus more on how to achieve better overlap between the predicted area and the true area during training. In image tasks, it is suitable for dealing with imbalanced positive and negative samples. Its mathematical formula is:

[0110]

[0111] Where: p i represents the predicted value of the model, that is, the predicted probability that the i-th pixel (or area) belongs to the target category; g i Indicates the true label value, that is, whether the i-th pixel (or area) is the target category; Represents the sum of the intersection of the predicted area and the true area; represents the sum of squares of the model prediction area; Represents the sum of squares of the true area; N represents the number of pixels (or regions) in the image or dataset, which is used to calculate the loss for each pixel (or region) and accumulate it.

[0112] For image segmentation tasks, p i represents the probability value of the model predicting that the i-th pixel in the image belongs to the target category (such as foreground). For each pixel or region, g i =1 means the pixel belongs to the target category, g i =0 means that the pixel does not belong to the target category. This term measures the overlap between the predicted area and the true area. The larger it is, the closer the model's prediction is to the true label. Measures the "total amount" of the model prediction area, that is, the sum of all pixels predicted by the model as the target category. A larger value indicates a larger area predicted by the model. It measures the “total amount” of the true label area, that is, the sum of pixels that are actually the target category.

[0113] In summary, if the primary goal of predicting the potential of decommissioned nuclear graphite is a regression problem (e.g., predicting radiation intensity or temperature changes), choose the MSE or Huber loss function as the loss function. If the task is to classify the potential risk of nuclear graphite, choose the cross-entropy loss function for optimization. If the task involves image segmentation (e.g., segmenting high-risk areas), choose the Dice loss function to optimize the segmentation effect.

[0114] Step 5: Introduce the attention mechanism to enhance the model's attention to the key features of nuclear graphite by focusing on the key areas in the image.

[0115] The purpose of step 5 is to introduce an attention mechanism, enabling the model to focus on the most important areas of the image for the classification task and suppress interference from irrelevant areas. This effectively improves the ability to extract key features, avoids information loss, and thus improves classification accuracy and model robustness. Especially when processing complex image data, the attention mechanism can help the model focus on areas with high discriminative power, improving task performance.

[0116] Critical areas in the image include, but are not limited to, surface cracks and defects, temperature hotspots, high radiation areas, and internal structural defects. Surface cracks and defects, including defects such as cracks, wear, and corrosion, are important indicators of potential risks in nuclear graphite, impacting its structural strength, thermal conductivity, and radiation tolerance. Temperature hotspots are areas of high temperature or uneven heat conduction shown in the temperature distribution map. These areas indicate potential thermal damage or localized overheating, which in turn affects the lifespan and safety of nuclear graphite. High radiation areas are areas of high radiation intensity in gamma-ray images. These areas are where radioactive material accumulates and are often associated with potential radiation risks in nuclear graphite, requiring special attention. Internal structural defects, including internal structural defects such as cracks and pores identified in ultrasound images, are key factors leading to nuclear graphite cracking, structural failure, or uneven thermal conductivity. By focusing on and enhancing these critical areas, the model can more accurately assess the potential risks of decommissioned nuclear graphite, providing stronger data support for subsequent waste disposal and environmental monitoring.

[0117] The attention mechanisms in step 5 include channel attention mechanism (SENet), spatial attention mechanism, and scSE attention mechanism (combining channel and spatial attention).

[0118] Specifically, the channel-wise attention mechanism (SENet) introduces an attention mechanism at the channel level, dynamically adjusting the importance of each channel. In predicting the energy potential of decommissioned nuclear graphite, the model prioritizes channel features with strong discriminative power, such as radiation distribution and temperature variation. SENet utilizes max pooling and normalization to prioritize important channels, suppress interference from irrelevant channels, and enhance classification capabilities.

[0119] The implementation of the SENet mechanism involves: first, performing a max-pooling operation on the input feature map to capture global information. The pooled features are then normalized to obtain per-channel weights. Finally, the weights are multiplied by each channel of the original input, thereby increasing attention to important channels and suppressing the influence of irrelevant channels. The SENet mechanism enhances the model's focus on key channels, reduces interference from irrelevant channels, and improves classification capabilities.

[0120] The spatial attention mechanism uses max and average pooling operations to help the model identify important spatial locations in an image. In the task of predicting the energy potential of decommissioned nuclear graphite, the spatial attention mechanism helps the model focus on cracks, hot spots, and areas of high radiation intensity, which are crucial for energy potential prediction and effectively improve prediction accuracy.

[0121] The implementation process of the spatial attention mechanism involves: first, performing max pooling and average pooling operations on each channel of the input data, respectively, to obtain two feature maps (representing the maximum and average values of the key regions). These two feature maps are then concatenated along the channel dimension to form a new feature map. A convolution operation is then used to reduce the dimensionality, resulting in a spatial attention map with one channel. This is then normalized using a sigmoid function to generate spatial weights. Finally, the spatial attention weights are multiplied by the original data to obtain a spatially weighted feature map. Spatial weighting enables the model to focus on the most discriminative areas of the image, thereby improving classification performance.

[0122] The scSE attention mechanism combines channel attention and spatial attention, making it suitable for tasks that require attention to both spatial and channel features. In the prediction of the energy potential of decommissioned nuclear graphite, both spatial features (such as cracks, hot spots, and radiation areas) and channel features (such as radiation intensity and temperature variations) are crucial. Therefore, the scSE attention mechanism is used for joint optimization, further improving the model's ability to extract key region and channel features from images.

[0123] The implementation process of the scSE attention mechanism includes the following steps: First, a spatial attention map is generated using a convolution operation (Conv2d) and normalized using sigmoid to obtain spatial attention weights. This is then multiplied with the input feature map to obtain a weighted spatial feature map. The input is pooled using adaptive pooling to obtain the weight of each channel. Channel weights are generated using a sigmoid function and multiplied with the original data to obtain a weighted channel feature map. Finally, the feature maps weighted by spatial and channel attention are fused to obtain the final weighted feature map. By combining the spatial and channel attention mechanisms, the model can simultaneously focus on important spatial regions and important channel features in the image, thereby further improving the accuracy of image classification.

[0124] Step 5: Select an appropriate attention mechanism based on the different feature requirements of different tasks, image complexity, and computational efficiency. For example, for tasks that require in-depth analysis of spatial features, a spatial attention mechanism may be more appropriate; for tasks that emphasize the importance of channel features, a channel attention mechanism may be more effective. If the information distribution in the image is complex and both spatial and channel features are crucial to the results, consider combining channel and spatial attention mechanisms (such as the scSE mechanism). Some attention mechanisms (such as SENet and spatial attention) may be more computationally efficient than the scSE mechanism. Choosing the appropriate attention mechanism based on computational resources is also an important consideration.

[0125] Step 6: Build a feature extraction network and train the U-Net model by combining the multi-dimensional feature map fused in step 3, the loss function defined in step 4, and the attention mechanism in step 5.

[0126] The purpose of step 6 is to train a U-Net model, as ultrasound images and radiation distribution images share similarities with medical imaging in processing methods, particularly in image segmentation and feature extraction. The U-Net model, through its encoder-decoder structure, efficiently extracts deep image features and accurately restores spatial information, making it suitable for multimodal image processing in the nuclear graphite potential prediction task. This structure enables the U-Net model to learn richer image information from the fused feature maps, thereby providing accurate predictions for nuclear graphite potential assessment.

[0127] It should be noted that U-Net is a classic image segmentation network that utilizes an encoder-decoder architecture to efficiently extract image features and accurately restore the image's spatial resolution. The encoder extracts high-level features from the input image, while the decoder restores these features to the original image's spatial resolution. It is commonly used in image segmentation tasks. For ultrasound images and radiation distribution images used in nuclear graphite potential prediction, U-Net is able to efficiently extract multi-dimensional features and generate accurate potential predictions.

[0128] The encoder progressively extracts image features through a series of convolutional and pooling layers, increasing the number of channels at each layer. As the network progresses, the spatial dimensions of the image decrease, and the features become increasingly abstract. In the decoder, upsampling restores the spatial dimensions of the image, gradually reducing the size of the feature map to the size of the input image. This allows the decoder to recover image detail and output a classification or prediction for each pixel. The skip connections in the U-Net connect shallow features in the encoder with deep features in the decoder, helping the model retain more detailed information and improving image restoration accuracy.

[0129] Step 6 specifically includes:

[0130] Step 6.1: Select U-Net as the training model and combine it with ResNet as the feature extraction network.

[0131] U-Net was chosen as the primary network architecture. Its strengths in image segmentation make it well-suited for multimodal image processing of decommissioned nuclear graphite. ResNet (residual network) was used as the feature extraction network, utilizing the first half (encoder portion) of a pre-trained ResNet18 or ResNet50 model. ResNet, through its residual connections, effectively mitigates the vanishing gradient problem in deep networks and possesses powerful feature extraction capabilities, making it suitable for extracting high-level features from complex images.

[0132] The process of combining U-Net and ResNet is as follows: ResNet extracts deep features from the image, which are gradually reduced in spatial size after being processed by each residual block. The U-Net decoder restores the spatial resolution of the image. It selects the first half (encoder) of a pre-trained model such as ResNet18 or ResNet50 and combines the shallow features (encoder) of the ResNet with the deep features of the U-Net decoder through skip connections, thereby preserving more detailed information.

[0133] Step 6.2: Use ResNet to extract features from the multi-dimensional feature map in step 3 and fuse them to form a multimodal fusion feature map. Divide each fused feature map into multiple small blocks (patches), and input each patch into U-Net for processing. Apply the channel attention mechanism after each layer of the U-Net network module, and apply the scSE attention mechanism after the last layer of the U-Net to further enhance the expression of spatial and channel features.

[0134] Step 6.3: Select a combined loss function and combine Dice loss with cross entropy loss as the loss function of the U-Net model.

[0135] Step 6.4: Use the Adam optimizer to adjust the network parameters and gradually update the weights of the U-Net model according to the changes in the loss function during training. Repeat the training and monitor the changes in the loss function until the preset number of training rounds is reached or the loss function changes tend to be stable and the model converges.

[0136] It's important to note that Adam (Adaptive Moment Estimation) is a common optimization algorithm based on gradient descent, combining the principles of momentum and RMSProp. Momentum primarily considers the accumulation of previous gradients to accelerate parameter updates in relevant directions, while RMSProp adaptively adjusts the learning rate of each parameter, minimizing the learning rate of frequently updated parameters.

[0137] The training process of step 6.4 specifically includes:

[0138] Step 6.4.1: Initialize parameters: Randomly initialize the parameters of the convolutional network, set the learning rate, decay rate and small constant for numerical stability.

[0139] Specifically, the first-order moment estimation vector m and the second-order moment estimation vector v are initialized to 0, and the learning rate α, decay rates β1 and β2, and a small constant ε for numerical stability are set.

[0140] Step 6.4.2: Forward Propagation: Pass the input data through the convolutional network to calculate the probability value predicted by the model And through the predicted probability value Calculate the loss function with the true label y

[0141] Step 6.4.3: Calculate the gradient: Calculate the gradient of the loss function with respect to the convolution kernel weight w and bias b through the backpropagation algorithm and

[0142] Step 6.4.4: Update moment estimates: Let the current time step be t, and the first-order moment estimate be and The second-order moment estimate is and

[0143] Step 6.4.5: Bias correction: Because m and v are 0 at the time of initialization, there will be bias in the early estimation, so correction is required. The first-order moment estimate of the parameter is corrected to The second-order moment estimate of the parameter is corrected to

[0144] Step 6.4.6: Update parameters: Update the convolutional network parameters based on the corrected moment estimates. The updated weight parameters are Update the bias parameters to

[0145] Step 6.4.7: Repeat iterations: Repeat steps 6.4.2-6.4.6 until the preset number of training rounds is reached or until the loss function changes tend to be stable and the model converges.

[0146] Step 6.6: Input the new image data into the trained U-Net model and output the potential prediction value of each pixel. Based on the prediction results, the potential risk of nuclear graphite is assessed and a reference is provided for waste disposal and environmental monitoring.

[0147] Step 7: Evaluate the trained model and optimize the model based on the evaluation results.

[0148] The purpose of step 7 is to evaluate the trained model and analyze key metrics such as accuracy and Intersection over Union (IOU) to determine its performance in the nuclear graphite potential prediction task. If the model's accuracy is insufficient or its IOU is low, the model is optimized based on the evaluation results to improve prediction accuracy and robustness.

[0149] Specifically, the trained model is evaluated, and indicators such as model accuracy and intersection over union (IOU) are analyzed. Accuracy is a measure of the consistency between the model prediction results and the true labels. In the nuclear graphite potential prediction task, accuracy can help us determine whether the model can correctly predict the potential risk areas of nuclear graphite (such as cracks, hot spots, radiation intensity areas, etc.). In the nuclear graphite potential prediction, accuracy mainly reflects the model's ability to correctly identify potential risk areas within the entire prediction area. Intersection over union (IOU) is a commonly used evaluation indicator in image segmentation tasks, which is used to measure the degree of overlap between the model prediction area and the actual real area. For nuclear graphite potential prediction, the IOU value can help determine the spatial accuracy of the model in predicting risk areas (such as high radiation areas, crack areas, etc.).

[0150] The IOU formula is:

[0151]

[0152] A high IOU value (close to 1) indicates that the model accurately predicts the risk areas of nuclear graphite, indicating that the model has strong spatial localization capabilities in risk areas and can identify key areas (such as cracks, hot spots, and radiation zones). A low IOU value indicates that the overlap between the model's predicted areas and the actual areas is low, indicating that the model may have missed detections or misjudged some areas and requires further optimization.

[0153] The optimization of the model based on the evaluation results includes but is not limited to hyperparameter adjustment, data augmentation, regularization (to prevent overfitting) and model fine-tuning.

[0154] Hyperparameter tuning involves adjusting model hyperparameters to improve performance based on evaluation results (such as low accuracy or insufficient IOU). Common tuning parameters include the learning rate, batch size, and loss function. Adjusting the learning rate controls the speed of parameter updates during training, preventing overfitting or underfitting. Adjusting the batch size can affect the training speed and stability of the model. Based on the characteristics of the task, an appropriate loss function (such as Dice loss or cross-entropy loss) should be selected to best suit the nuclear graphite potential prediction task.

[0155] Data augmentation increases the diversity of training data by performing transformations such as rotation, translation, and scaling, thereby improving the model's adaptability to different image types. For nuclear graphite potential prediction tasks, if the model performs poorly on certain image types (such as those showing unusual cracks or areas of high radiation), data augmentation can be used to supplement these image types and enhance the model's generalization capabilities.

[0156] Regularization refers to the problem of overfitting if a model performs well on training data but loses accuracy on test data. Methods such as adding a dropout layer or using L2 regularization can improve the model's generalization ability. L2 regularization (weight decay) prevents the model from over-relying on certain features, improving its generalization ability on unseen data. The dropout layer randomly discards some neurons in the neural network to prevent the model from becoming overly complex, thereby reducing the risk of overfitting.

[0157] Model fine-tuning refers to adjusting the model structure or training strategy based on the evaluation results of an existing pre-trained model. Specific methods include: using a deeper network (such as ResNet101 instead of ResNet50) to improve feature extraction capabilities; using more complex feature extraction methods (such as more advanced convolution methods or more detailed image segmentation techniques) to improve model performance.

[0158] Step 8: The status and prediction results of nuclear graphite are displayed in real time through the human-computer interaction interface, and a waste treatment plan is selected based on the prediction information.

[0159] The purpose of step 8 is to design and implement an intuitive user interface that allows operators to view the status and prediction results of nuclear graphite in real time, mark and classify the internal potential of nuclear graphite through the interface, and select appropriate waste treatment solutions based on the prediction results.

[0160] Specifically, the user interface displays the nuclear graphite state image, feature map, and prediction results predicted by the model in real time. The interface includes the following functions: real-time monitoring image, feature map display, and prediction result display.

[0161] The real-time monitoring image function displays images of the current state of nuclear graphite. These images are derived from data from different modalities, such as ultrasound images (used to inspect internal structural defects in nuclear graphite), infrared images (used to monitor thermal conductivity and hot spots in nuclear graphite), and gamma-ray images (used to identify radiation intensity and radioactivity distribution). These images provide real-time information on the nuclear graphite's status and risk assessment.

[0162] The feature map display function can display various feature maps extracted by the model, such as spatial feature maps and channel feature maps, to help operators understand the model's image analysis process. The spatial feature map displays the spatial feature map extracted by the model to help operators understand the spatial layout of nuclear graphite (for example, cracks, fissures, hot spots, etc.). This map can highlight potential risk areas and help operators quickly locate important monitoring areas. The channel feature map displays the channel features extracted by the convolutional neural network (CNN), helping users understand the model's understanding and focus on nuclear graphite at different levels. The risk heat map displays a heat map of the nuclear graphite potential prediction results, highlighting potential risk areas (such as high radiation areas, temperature anomaly areas, crack areas, etc.). This heat map can help operators quickly identify problem areas and respond.

[0163] The prediction results display function displays the model's predictions in a graphical or heat map format, clearly identifying areas with high potential and potential risk. This allows operators to clearly understand the risk level of each area and make appropriate decisions accordingly. The model output will classify nuclear graphite according to the predicted potential strength (e.g., high risk, medium risk, low risk), and the image uses different colors to identify areas of different risk levels. For example, red indicates high-risk areas (such as high radiation areas, hot spots, and deep cracks), orange indicates medium risk, and green indicates low risk areas.

[0164] Based on the potential prediction results of nuclear graphite, the waste treatment plan can be carried out as follows:

[0165] 1. High-risk areas for nuclear graphite include: areas with high radiation intensity, areas with abnormally high temperatures, and areas with large internal cracks or defects. For these areas, waste disposal plans should include:

[0166] Radiation protection: High-radiation areas are isolated using radiation-proof materials and barriers to reduce the impact of radiation on the environment and personnel. Temperature control management: For hot spots, rapid cooling measures (such as liquid nitrogen cooling) are implemented to prevent further heating of the nuclear graphite, which could lead to more severe structural damage. Special waste storage: For high-risk areas with significant internal damage, special sealed containers are used for waste storage and transportation to prevent leakage of internal contaminants.

[0167] 2. Medium-risk areas may include areas where cracks are not deep and local radiation levels are high. Waste disposal options for these areas include:

[0168] Localized repair: Reinforce and repair cracked areas to reduce the risk of further damage. Radiation monitoring: Regularly monitor areas with elevated radiation levels to ensure they remain within acceptable limits. Isolation management: Isolate these medium-risk areas from low-risk areas and develop a plan for regular inspections and treatment.

[0169] 3. Low-risk areas include areas without obvious cracks, low radiation intensity, and small temperature changes. For these areas, the waste disposal plan can be relatively simple, mainly including:

[0170] Conventional storage: Waste in these areas will be stored according to conventional nuclear waste storage standards, without the need for special radiation or temperature control measures. Regular monitoring: Although these are low-risk areas, they still require regular inspections to ensure that no new risks have emerged.

[0171] By designing an intuitive user interaction interface, real-time monitoring images, feature maps and nuclear graphite status are provided, allowing operators to quickly understand the internal conditions of nuclear graphite, mark and classify nuclear graphite with different internal potentials through prediction results, and select appropriate nuclear waste treatment processes according to the strength of the internal potential of nuclear graphite.

[0172] Example 1

[0173] Nuclear graphite is a key component in the decommissioning process of nuclear power plants. Due to prolonged radiation exposure and high temperatures, nuclear graphite can develop cracks, defects, and residual radiation, which directly impact its post-decommissioning potential and safety. Therefore, accurately assessing the potential of decommissioned nuclear graphite is crucial for nuclear waste management and environmental monitoring. This example demonstrates a multimodal vision-based method for predicting the potential of decommissioned nuclear graphite, combining real-time monitoring with waste disposal solutions to improve the safety and efficiency of nuclear graphite decommissioning.

[0174] Step 1: Acquire multimodal image data of retired nuclear graphite.

[0175] A high-resolution camera (e.g., a Canon EOS 5DS, 50.6MP) was used to photograph the surface of the nuclear graphite, achieving image resolution exceeding 20MP, ensuring the capture of microcracks, wear, and surface defects. Camera parameters were set to an exposure time of 1 / 500 second, an aperture of f / 8, and an ISO of 100 to reduce noise and enhance image quality. Subsequently, a FLIR T1K thermal imaging camera was used to monitor the temperature distribution of the nuclear graphite. This instrument features a resolution of 1000 × 1000 pixels, a temperature accuracy of ±2°C, and a scanning frequency of 20 frames per second, enabling the capture of surface and internal hotspots. A Mirion Technologies RadEye PRD gamma-ray camera was used to measure the radiation intensity of the nuclear graphite, with a measurement range of 0.01 μSv / h to 10 Sv / h and an accuracy of 1%. Finally, internal flaw detection was performed using a GE process-controlled USN60 ultrasonic flaw detector with a frequency range of 0.5 MHz to 10 MHz. The probe frequency was adjusted according to the crack depth to ensure comprehensive detection of defects such as cracks and pores.

[0176] Step 2: Preprocess the collected multimodal image data.

[0177] The collected image data was denoised, and median filtering and mean filtering (3x3 kernel) were used to remove environmental interference. Data enhancement included rotation of ±15°, scaling by 0.9-1.1 times, and color adjustment to ensure the robustness of the model. In the temperature calibration of infrared thermal images, a known standard temperature source was used to calibrate the temperature of the infrared thermal imager to ensure a temperature measurement accuracy of ±0.5°C. For large-area images, ORB or SIFT algorithms were used for stitching, and a gradient fusion algorithm was used to ensure the continuity and seamless transition of the stitched images. Gamma-ray images and ultrasound images were logarithmically normalized, and the image pixel values were unified to the range of [0,1]. The ultrasound images were smoothed by Gaussian filtering (standard deviation σ = 1.0) to reduce noise and highlight low-frequency information.

[0178] Step 3: Use deep learning models to extract key features from the preprocessed images.

[0179] In the feature extraction stage, a convolutional neural network (CNN) is used to extract key features from the preprocessed images. For infrared thermal images, temperature gradient features are extracted, and the CNN model is used to analyze the hot spots and temperature distribution in the image. Gamma-ray images use CNN to extract radiation distribution features to help identify high-radiation areas. Ultrasound images use CNN to extract internal structural features to accurately identify defects such as cracks and pores. To ensure that the features of images with different modalities can be effectively integrated, the features of all images are adjusted to the same size (256×256) after standardization, and weights are set based on the degree of influence of different modalities on the prediction results of nuclear graphite potential. The weights of temperature gradient and radiation distribution features are higher, while the weights of surface defects and ultrasound images are lower.

[0180] Step 4: Select the corresponding loss function based on the decommissioned nuclear graphite potential prediction task and the collected data characteristics.

[0181] For the potential prediction task of decommissioned nuclear graphite, the Huber loss function was chosen. This is suitable for regression tasks, especially when the data may contain outliers. Huber loss can effectively reduce the impact of outliers and improve the robustness of the model. The Huber loss function combines the advantages of squared loss and linear loss. When the prediction error is less than a set threshold (δ), the squared loss is used; when the error is greater than the threshold, the linear loss is used, thus avoiding excessive penalty for large errors.

[0182] Step 5: Introduce the attention mechanism.

[0183] During feature extraction, a channel-wise attention mechanism (SENet) is employed to enhance the model's focus on key feature channels. After each convolutional layer, a max pooling operation is used to capture global information. The pooled features are then normalized to obtain per-channel weights. By multiplying these weights by each channel of the original input, the model's focus on key channels, such as temperature variation and radiation distribution, is enhanced, suppressing the influence of irrelevant channels and improving classification capabilities.

[0184] Step 6: Build a feature extraction network and train it in conjunction with the U-Net model.

[0185] U-Net was selected as the model architecture for image segmentation, combined with the encoder portion of ResNet50 for feature extraction. ResNet50 effectively mitigates the vanishing gradient problem through residual connections, enabling high-level feature extraction from complex images. The decoder portion of U-Net restores the spatial resolution of the image, ensuring sufficient spatial accuracy in the prediction results. A combination of the Dice loss function and the cross-entropy loss function was used to optimize image segmentation and classification. During training, the Adam optimizer was used with a learning rate of 0.001 and a momentum decay rate of 0.9.

[0186] Step 7: Evaluate and optimize the model.

[0187] The trained model is evaluated using accuracy and IOU as metrics. In nuclear graphite potential prediction, accuracy measures the model's ability to correctly identify potential risk areas within the entire image, while IOU measures the overlap between the model's predicted and actual areas. Model performance is further optimized by adjusting hyperparameters such as the learning rate, batch size, and loss function. If the model's IOU falls below 0.7, data augmentation, regularization (such as L2 regularization), and model fine-tuning (such as increasing network depth) are used to improve model accuracy.

[0188] Step 8: The status and prediction results of nuclear graphite are displayed in real time through the human-computer interaction interface, and a waste treatment plan is selected based on the prediction information.

[0189] In the human-computer interaction interface, the nuclear graphite state image, feature map and prediction results obtained from the model prediction are displayed in real time. The interface includes real-time monitoring images, feature map display and risk heat map, which can display key information such as surface defects, hot spots, radiation intensity distribution of nuclear graphite. Operators can classify nuclear graphite into high, medium and low risk levels based on the prediction results and adopt corresponding waste treatment plans. For high-risk areas (such as high radiation, deep cracks, and hot spots), the system automatically recommends radiation protection, temperature control management and special storage plans; for medium-risk areas (such as local cracks and areas with increased temperature), local repair and radiation monitoring are recommended; for low-risk areas, routine storage and regular inspections are recommended.

[0190] This example achieves accurate prediction of the potential of decommissioned nuclear graphite through precise device selection and parameter settings, combined with deep learning models, attention mechanisms, and multimodal image data processing. The model's accuracy and robustness are enhanced through an optimized loss function (Huber loss), a channel attention mechanism (SENet), and a U-Net model. Furthermore, through an intuitive user interface and waste disposal solution, potential risks during nuclear graphite decommissioning are scientifically assessed and effectively managed, providing strong support for environmental monitoring and waste disposal.

[0191] Example 2

[0192] Traditional methods for predicting the potential of decommissioned nuclear graphite rely primarily on single-site detection of radioactive elements, typically using a gamma spectrometer to measure the energy and intensity of characteristic gamma rays from different radionuclides. The decommissioned nuclear graphite potential prediction method proposed in this application, based on multimodal vision and deep learning intelligent algorithms, can overcome the limitations of traditional methods and provide a more comprehensive, intelligent, and accurate assessment solution. The advantages of this application are illustrated below by comparing existing decommissioned nuclear graphite potential prediction methods with the decommissioned nuclear graphite potential prediction method proposed in this application.

[0193]

[0194]

[0195] In summary, this application solves the limitations of traditional nuclear graphite potential prediction methods by combining multimodal data acquisition, deep learning models and intelligent analysis. Through high-resolution cameras, infrared thermal imagers, gamma-ray cameras and ultrasonic flaw detectors and other equipment, comprehensive acquisition of multi-dimensional information such as the surface morphology, temperature distribution, radiation intensity and internal structure of nuclear graphite, the U-Net image segmentation algorithm and ResNet50 are used for feature extraction and fusion, combined with SENet and scSE attention mechanisms to accurately identify potential risk areas, providing efficient and accurate waste treatment solutions. Compared with traditional methods, this application can comprehensively consider multiple factors, perform intelligent data analysis and fusion, greatly improving the accuracy, robustness and real-time performance of potential prediction, and ensuring the safety and efficiency of waste treatment.

[0196] As the instruction manual Figure 2 The present invention also provides a decommissioned nuclear graphite potential prediction system based on multimodal vision, the system comprising:

[0197] The multimodal image acquisition module is used to collect multimodal image data of the surface morphology, temperature distribution, radiation intensity and internal structure of retired nuclear graphite to obtain comprehensive nuclear graphite status information.

[0198] The data preprocessing module is used to preprocess the collected multimodal image data.

[0199] The feature extraction and fusion module is used to extract key features from the preprocessed image using a deep learning model, and integrate the key features into a multi-dimensional feature map through feature fusion technology.

[0200] The loss function selection module is used to select the corresponding loss function according to the retired nuclear graphite potential prediction task and the collected data characteristics.

[0201] The attention mechanism module is used to introduce the attention mechanism to enhance the model's attention to the key features of the kernel by focusing on the key areas in the image.

[0202] The feature extraction network module is used to train the U-Net model by combining the multidimensional feature map fused in the feature extraction and fusion module, the loss function defined in the loss function selection module, and the attention mechanism in the attention mechanism module.

[0203] The model evaluation and optimization module is used to evaluate the trained model and optimize the model based on the evaluation results.

[0204] The human-computer interaction interface module is used to display the status and prediction results of nuclear graphite in real time, and provide options for waste treatment solutions based on the prediction information.

[0205] Furthermore, the multimodal image acquisition module includes four main sensors: a high-resolution camera, an infrared thermal imager, a gamma-ray camera, and an ultrasonic flaw detector. The high-resolution camera is used to capture surface images of decommissioned nuclear graphite, capturing its surface morphology, cracks, and other surface defects. The infrared thermal imager records the surface and internal temperature distribution of decommissioned nuclear graphite, reflecting its internal heat conduction characteristics and revealing possible thermal stress points. The gamma-ray camera, through its nuclide identification function, generates nuclide distribution images and radiation intensity distribution images to identify the radioactivity distribution of nuclear waste. The ultrasonic flaw detector is used to detect defects such as cracks and pores within the nuclear graphite, generating images of its internal structure for analysis of potential internal defects.

[0206] The data preprocessing module performs image enhancement (such as contrast enhancement and noise suppression) on the collected multimodal image data to improve image quality. It also standardizes the images, including resizing and brightness normalization, to facilitate subsequent feature extraction. De-noising algorithms are used to eliminate the effects of sensor errors and environmental interference.

[0207] The feature extraction and fusion module uses deep learning methods (such as convolutional neural networks (CNNs)) to extract key features from preprocessed images, such as crack width, radiation intensity distribution, and areas of temperature anomaly. Using multimodal data fusion techniques (such as weighted fusion and weighted summation), data from different sensors is fused into a multidimensional feature map that comprehensively reflects the internal state of decommissioned nuclear graphite.

[0208] The loss function selection module selects an appropriate loss function (such as mean squared error (MSE) or cross entropy loss function) based on the specific requirements of the decommissioned nuclear graphite potential prediction task. The choice of loss function should be consistent with the specific task objectives, such as minimizing prediction error or improving classification accuracy.

[0209] The attention mechanism module introduces an attention mechanism to focus on key areas in the image, such as cracks, radiation sources, or thermal stress areas, increasing the model's attention to these areas. Through self-attention or channel-attention mechanisms, the model can identify more critical latent features in multimodal data.

[0210] The feature extraction network module combines the multidimensional feature maps, loss functions, and attention mechanisms from the aforementioned modules to train a U-Net model to extract the potential characteristics of decommissioned nuclear graphite. The U-Net model, a deep learning network commonly used in medical image processing, has excellent feature extraction and image segmentation capabilities, making it suitable for predicting the internal potential distribution of nuclear graphite.

[0211] The model evaluation and optimization module evaluates the model after training, using metrics such as accuracy, recall, and F1 score to measure its performance. Based on the evaluation results, the model is optimized, including adjusting the network structure, hyperparameters (such as learning rate and batch size), and retraining the model.

[0212] The human-computer interaction interface module provides a user-friendly interface for real-time monitoring of nuclear graphite status images, characteristic graphs, and potential prediction results. Users can view nuclear graphite status information based on the prediction results, classify and label the waste, and select appropriate waste treatment options. The interface should support data visualization, image display, and selection of treatment processes for different types of nuclear waste.

[0213] The above is only an embodiment of the present invention, and common sense such as the specific structure and characteristics of the scheme are not described in detail here. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes that fall within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claim involved.

Claims

1. A method for predicting the potential of retired nuclear graphite based on multimodal vision, characterized in that: The method comprises: Step 1: Collect multimodal image data of the surface morphology, temperature distribution, radiation intensity, and internal structure of retired nuclear graphite to obtain comprehensive nuclear graphite status information; Step 2: Preprocess the collected multimodal image data; Step 3: Use the deep learning model to extract key features from the multimodal image data after preprocessing, and integrate the extracted key features into a multi-dimensional feature map through feature fusion technology; Step 4: Select the corresponding loss function based on the decommissioned nuclear graphite potential prediction task and the collected data characteristics; Step 5: Introduce the attention mechanism to enhance the model's attention to the key features of nuclear graphite by focusing on the key areas in the image; Step 6: Build a feature extraction network and train the U-Net model by combining the multi-dimensional feature map fused in step 3, the loss function defined in step 4, and the attention mechanism in step 5. Step 7: Evaluate the trained model and optimize the model based on the evaluation results; Step 8: The status and prediction results of nuclear graphite are displayed in real time through the human-computer interaction interface, and a waste treatment plan is selected based on the prediction information.

2. The method for predicting the potential of retired nuclear graphite based on multimodal vision according to claim 1, characterized in that: In step 4, the retired nuclear graphite potential prediction task includes at least a regression task, a classification task, and an image segmentation task. The regression task is used to predict the potential risks of retired nuclear graphite, the classification task is used to classify the potential risks of retired nuclear graphite, and the image segmentation task is used to segment cracks, hot spots, and high-risk radiation areas in retired nuclear graphite images.

3. The method for predicting the potential of retired nuclear graphite based on multimodal vision according to claim 2, characterized in that: The loss function selected for the regression task is the mean square error or Huber loss function, the loss function selected for the classification task is the cross entropy loss function, and the loss function selected for the image segmentation task is the Dice loss function.

4. The method for predicting the potential of retired nuclear graphite based on multimodal vision according to claim 3 is characterized in that: In step 5, the critical areas include but are not limited to surface cracks and defect areas, temperature hot spots, high radiation areas, and internal structural defect areas; The attention mechanism includes channel attention mechanism, spatial attention mechanism, combined channel attention mechanism and spatial attention mechanism.

5. The method for predicting the potential of retired nuclear graphite based on multimodal vision according to claim 4 is characterized in that: Step 6 includes: Step 6.1: Select U-Net as the training model and combine it with ResNet as the feature extraction network; Step 6.2: Use ResNet to extract features from the multi-dimensional feature map in step 3 and fuse them to form a multimodal fused feature map. Divide each fused feature map into multiple patches, and input each patch into U-Net for processing. Apply the channel attention mechanism after each layer of the U-Net network module. Apply the combined channel attention mechanism and spatial attention mechanism after the last layer of the U-Net to further enhance the expression of spatial and channel features. Step 6.3: Select a combined loss function, combining Dice loss and cross entropy loss as the loss function of the U-Net model; Step 6.4: Use the Adam optimizer to adjust the network parameters and gradually update the weights of the U-Net model according to the changes in the loss function during training. Repeat the training and monitor the changes in the loss function until the preset number of training rounds is reached or the loss function changes tend to be stable and the model converges.

6. The method for predicting the potential of retired nuclear graphite based on multimodal vision according to claim 1, characterized in that: In step 7, the trained model is evaluated by analyzing the model's accuracy and intersection-over-union (IoU) metrics. The model is optimized based on the evaluation results, including but not limited to hyperparameter adjustment, data augmentation, regularization, and model fine-tuning.

7. The method for predicting the potential of retired nuclear graphite based on multimodal vision according to claim 1, characterized in that: Step 3 includes: Step 3.1: Use a convolutional neural network to extract surface defect features, temperature gradient features, radiation distribution features, and internal structure features from the preprocessed multimodal image data; Step 3.2: Normalize the feature maps extracted from different modality images; Step 3.4: Set the corresponding fusion weight according to the influence of different feature maps on the nuclear graphite potential prediction results; Step 3.5: Perform weighted fusion of the feature maps of each modality through convolutional neural network to obtain a multi-dimensional feature map.

8. The method for predicting the potential of retired nuclear graphite based on multimodal vision according to claim 1, characterized in that: Step 1: Multimodal image data is collected using a high-resolution camera, an infrared thermal imager, a gamma-ray camera, and an ultrasonic flaw detector; the multimodal image data includes surface images of decommissioned nuclear graphite, infrared thermal images, radionuclide distribution images and radiation intensity distribution images, and ultrasonic flaw detection images.

9. The method for predicting the potential of retired nuclear graphite based on multimodal vision according to claim 1, characterized in that: Step 2 preprocesses the collected image data, including denoising, enhancement, calibration, stitching and normalization.

10. A decommissioned nuclear graphite potential prediction system based on multimodal vision, characterized in that: The system comprises: Multimodal image acquisition module, used to collect multimodal image data of the surface morphology, temperature distribution, radiation intensity and internal structure of retired nuclear graphite to obtain comprehensive nuclear graphite status information; A data preprocessing module, used for preprocessing the collected multimodal image data; A feature extraction and fusion module is used to extract key features from the preprocessed image using a deep learning model and integrate the key features into a multi-dimensional feature map through feature fusion technology; A loss function selection module is used to select a corresponding loss function based on the decommissioned nuclear graphite potential prediction task and the characteristics of the collected data; The attention mechanism module is used to introduce the attention mechanism to enhance the model's attention to the key features of nuclear graphite by focusing on the key areas in the image; A feature extraction network module is used to train a U-Net model by combining the multidimensional feature map fused in the feature extraction and fusion module, the loss function defined in the loss function selection module, and the attention mechanism in the attention mechanism module; A model evaluation and optimization module is used to evaluate the trained model and optimize the model based on the evaluation results; The human-computer interaction interface module is used to display the status and prediction results of nuclear graphite in real time, and provide options for waste treatment solutions based on the prediction information.

Images