Nuclear radiation online monitoring method based on convolutional neural network and multi-source data fusion

Abstract

The invention discloses a convolutional neural network and multi-source data fusion-based nuclear radiation on-line monitoring method, which comprises the following steps of: acquiring laser radar point cloud data, a radiation dose rate and a radiation image of a monitoring area by adopting an unmanned aerial vehicle, and de-noising the radiation image by adopting a double-branch convolutional neural network; determining the track position of the unmanned aerial vehicle according to the laser radar point cloud and the de-noised image; performing space-time registration on the radiation dose rate and the track position of the unmanned aerial vehicle to form a data set, and obtaining a continuous radiation field of the monitoring area by adopting Kriging interpolation; according to the continuous radiation field, an L-BFGS-B algorithm is adopted to carry out minimization solution on a loss function with the radiation source position as an optimization variable, and position coordinates of the radiation source are obtained; performing data cleaning on the radiation dose in the continuous radiation field, and then performing multi-dimensional feature extraction to obtain a feature vector; and inputting the feature vectors into a plurality of prediction models to obtain a plurality of predicted radiation doses, and carrying out weighted fusion to obtain a predicted dose rate of the radiation source.

Description

Technical Field

[0001] This invention relates to nuclear radiation monitoring technology, specifically to an online nuclear radiation monitoring method based on convolutional neural networks and multi-source data fusion. Background Technology

[0002] In nuclear accident emergencies or searches for unknown radioactive sources, rapidly and accurately acquiring radiation field information and locating source terms is crucial. Unmanned aerial vehicles (UAVs) equipped with radiation imaging sensors have become important tools in this field. However, existing technologies face three major challenges: Poor image quality: Radiation images are affected by the randomness of high-energy particles and have severe Poisson-Gaussian mixed noise. Traditional denoising methods are prone to causing the loss of key texture details (such as particle trajectories).

[0003] Low positioning accuracy: The positioning capability of a single sensor is limited, and the time-space synchronization and deep fusion of multi-source sensors (LiDAR, vision, GNSS, radiometer) are insufficient, making it difficult to achieve accurate radiation field mapping and source term positioning in complex environments.

[0004] Weak predictive ability: The radiation dose dynamic change prediction model does not fully consider the specificity and suddenness of radiation data, and the prediction accuracy and timeliness cannot meet the needs of emergency decision-making.

[0005] Therefore, there is an urgent need for a comprehensive technical solution that can solve image denoising, high-precision positioning, and dynamic prediction in one integrated manner. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, the nuclear radiation online monitoring method based on convolutional neural networks and multi-source data fusion provided by this invention solves the problems of poor image quality, low positioning accuracy, and weak prediction capabilities when using existing UAVs equipped with radiation imaging sensors for online nuclear radiation monitoring.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for online monitoring of nuclear radiation based on convolutional neural networks and multi-source data fusion is provided, comprising the following steps: S1. Use a drone to collect lidar point cloud data, radiation dose rate and radiation image of the monitoring area, and use a dual-branch convolutional neural network to denoise the radiation image to obtain a denoised image. S2. Based on the lidar point cloud and denoised image, the visual odometry pose and lidar odometry pose are determined using SLAM mapping and trajectory optimization methods, and the two poses are fused to obtain the UAV trajectory position. S3. Spatiotemporal registration of radiation dose rate and UAV trajectory position is performed to form a dataset, and based on the dataset, Kriging interpolation is used to obtain the continuous radiation field of the monitoring area. S4. Based on the continuous radiation field, the L-BFGS-B algorithm is used to minimize the loss function with the radiation source location as the optimization variable to obtain the location coordinates of the radiation source. S5. Perform data cleaning on the radiation dose in the continuous radiation field to obtain the dose rate time series, and extract multi-dimensional features to obtain the feature vector composed of multi-dimensional features. S6. Input the feature vectors into multiple prediction models to obtain multiple predicted radiation doses, and then weight and fuse the three predicted radiation doses to obtain the predicted dose rate of the radiation source.

[0008] Furthermore, the dual-branch convolutional neural network includes a noise learning unit, a texture learning unit, and a feature fusion unit; The radiometric image is processed sequentially through a multi-scale convolutional module, a receptive field module, and a spatial attention module of the noise learning unit to obtain a noise feature map. ; After the radiometric image is input into the texture learning unit, features are extracted through a lightweight residual network containing three 3×3 convolutional layers. The output of the third convolutional layer is then added to the radiometric image to obtain the texture feature map. ; Feature fusion unit is used to analyze radiometric images, and After fusion, the image is passed through a 1×1 convolutional layer to output the final denoised image. The expression for fusion is: in, and These are the weighting coefficients; Features after fusion; This is a radiation image.

[0009] The beneficial effects of the above technical solution are as follows: the dual-branch convolutional neural network achieves the synergistic optimization of noise separation and texture preservation. The noise learning unit uses a multi-scale receptive field and spatial attention mechanism to accurately identify complex noise patterns, while the texture learning unit effectively protects key details such as radioactive particles through a lightweight residual structure. Finally, adaptive weighted fusion achieves a high texture preservation rate while maintaining the similarity of image structure, providing a high-quality visual data foundation for subsequent radiation monitoring.

[0010] Furthermore, based on the dataset, methods for obtaining the continuous radiation field of the monitoring area using kriging interpolation include: S31. The anisotropic variogram is used to calculate the variogram values ​​between all known point pairs in the dataset and between each point of the dose rate to be interpolated and all known points. The expression for the anisotropic variogram is: in, Given the variogram values, for a known pair of points i and j, the spatial vector h = p i -p j p i and p j Let i and j be the spatial coordinates of a known pair of points, respectively. For a known point i to point 0 of the dose rate to be interpolated, the spatial vector h = p i -p0, p0 are the spatial coordinates of point 0 of the dose rate to be interpolated; Let h be the value of the spatial vector in the xyz direction; , and For the range in the xyz directions; S32. Based on the variogram values ​​between all known point pairs and between each point of the dose rate to be interpolated and all known points, calculate the weight value corresponding to each point of the dose rate to be interpolated using the Kriging equations: in, , , and These are the variation function values ​​for known point 1 and known point 1, known point 2, known point n, and point 0 of the dose rate to be interpolated, respectively; , , and These are the variation function values ​​of known point 2 and known point 1, known point 2, known point n, and point 0 of the dose rate to be interpolated, respectively; , , and These are the variation function values ​​of known point n and known point 1, known point 2, known point n, and point 0 of the dose rate to be interpolated, respectively; , and These are the weight values ​​of known point 1, known point 2, and known point n, respectively, and point 0 of the dose rate to be interpolated. These are Lagrange multipliers used to constrain the sum of weights to 1; S33. Based on the radiation dose rates of all known points in the dataset and the weight values ​​of the points for which the dose rate to be interpolated is located, calculate the point dose rate for which the dose rate to be interpolated is located: in, The dose rate at point 0 of the dose rate to be interpolated; Let be the radiation dose rate at the i-th known point; The weight values ​​are the i-th known point and the point 0 of the dose rate to be interpolated; S34. After all points 0 of the dose rates to be interpolated are interpolated, the continuous radiation field of the monitoring area is obtained.

[0011] The beneficial effects of the above technical solution are as follows: This solution can accurately characterize the directional diffusion characteristics of the nuclear radiation field in space through the anisotropic variogram, and achieves accurate reconstruction from sparse discrete monitoring data to a three-dimensional continuous radiation field by combining the Kriging interpolation method. The advantage of this method is that it can effectively integrate multi-source heterogeneous monitoring data, provide optimal unbiased estimation through spatial autocorrelation analysis, significantly improve the accuracy and reliability of radiation field modeling, and provide accurate spatial distribution for subsequent precise location of radiation sources and dose prediction.

[0012] Furthermore, step S4 further includes: S41. Based on the continuous radiation field, calculate the radiation source at the known point using steps S31 to S33. Dose rate at the site ; S42, Based on dose rate Construct a loss function with the location of the radiation source as the optimization variable: in, The loss function; The coordinates of the radiation source location to be optimized; For a known point At point A, the theoretical dose caused by the radiation source; A is the measurement point. The source strength to be sought; This is the regularization term, where λ is the weighting coefficient. for The gradient; Let L2 be the norm of the vector; S43. The L-BFGS-B algorithm is used to minimize the loss function to obtain the coordinates of the radiation source location. .

[0013] The beneficial effects of the above technical solution are as follows: By constructing a loss function that integrates the physical model and measured data, the radiation source localization problem is transformed into an optimization problem, and the L-BFGS-B algorithm is used to achieve high-precision inversion of the radiation source location. This method effectively overcomes the localization ambiguity problem of traditional methods through matching and optimization between the theoretical dose model and the measured radiation field. Combined with gradient regularization constraints, it ensures the stability of the solution under complex radiation environments, and can provide reliable source term localization capabilities for nuclear emergency response.

[0014] Furthermore, before proceeding to step S5, the number of radiation sources is determined. When there are multiple radiation sources, K-means clustering is used to group the continuous radiation field to obtain the data points corresponding to each group. For each group, the corresponding methods in steps S41 to S43 are used to obtain the location coordinates of each radiation source. Then, steps S5 and S6 are executed for each group to obtain the predicted dose of each radiation source.

[0015] Furthermore, methods for multi-dimensional feature extraction from dose rate time series include: A sliding window of fixed length is used to extract the mean within the dose rate time series. Standard deviation Maximum value and minimum value : , = ) =max({ '…, '}) =min({ '…, '}) in, The dose rate time series within the sliding window is the value at time point o; t is the current time point. 'This represents the dose rate time series value at time point t-23; the fixed length is equal to 23.' Calculate the average gradient of dose rate change within a preset duration adjacent to the current time. : in, and These are the values ​​of time points r and r-1 in the dose rate time series, respectively; the preset duration is 5. Calculate periodic time characteristics ,in For working day indicator variables; Compare the drone trajectory location with the dose rate time series Perform spatial alignment and calculate the Euclidean distance between the current position of the UAV and the radiation source. : in, Current location of the drone The Euclidean distance to the radiation source's location coordinates; θrel is the azimuth angle of the current UAV position relative to the radiation source; Calculate the percentage of peak values and persistence : , in, It is a counting function; This represents the 0th dose rate value within the sliding window; These represent the mean and standard deviation of the historical background dose, respectively. Calculate the effective decay constant and meteorological influencing factors : in, nuclide The dose contribution weight; nuclide The decay constant; nuclide The dose contribution weight; This is a nuclide index used to traverse all nuclides to calculate the total dose rate; To monitor the prevailing wind direction in the area; and These represent the wind speed and wind direction in the monitored area.

[0016] The feature vector is constructed by using multi-dimensional features. : .

[0017] The beneficial effects of the above technical solution are as follows: K-means clustering effectively solves the problem of identification and separation in multi-radiation source scenarios; combined with multi-dimensional feature engineering, a comprehensive feature vector covering statistical characteristics, spatiotemporal variations, physical mechanisms, and environmental influences is constructed, achieving a comprehensive characterization of the evolution law of the radiation field. This feature system not only fully explores the spatiotemporal correlation characteristics of dose rate data, but also incorporates prior physical knowledge such as nuclide attenuation and meteorological diffusion, providing high-information-density input features for subsequent multi-model fusion prediction, significantly improving the accuracy and generalization ability of dose rate prediction in complex radiation environments.

[0018] Furthermore, the prediction results obtained by the XGBoost prediction model for: in, The number of training samples; For the model to the first Predicted dose rate for each sample; For the first The true dose rate of each sample; The loss function; These are parameters for controlling model complexity. These are the weight coefficients for the L2 regularization term; This is a vector of weights for all leaves in the entire model; T is the total number of decision trees. These are the weight coefficients for the L1 regularization term; For the first The leaf weight vector of each tree; The prediction results of the LSTM-BP neural network prediction model for: in, Mean square error, , This refers to the number of samples in a batch. The weighting coefficient for the spectral loss term is a preset positive number; This is a Fourier transform function used to extract spectral features from the dose rate sequence; The square of the L2 norm of the vector; and These are the sequence vectors of the predicted values ​​and the sequence vectors of the true values, respectively. Each of the batches The sequence vector of predicted values ​​and the sequence vector of true values ​​for each sample; The prediction results of the random forest prediction model for: Where Z is the total number of decision trees in the random forest model, z is the index of the decision tree, z=1,2,…,Z; Representing the z-th decision tree model, this function receives input features and outputs a predicted value; This is the multi-dimensional feature vector of the input model.

[0019] The beneficial effects of the above technical solution are as follows: By leveraging the complementary strengths of XGBoost, LSTM-BP neural network, and random forest, a powerful multi-model fusion prediction framework is constructed. XGBoost excels at capturing complex nonlinear relationships between features, LSTM-BP neural network effectively learns the temporal dynamic characteristics of dose rate, and random forest provides a stable baseline prediction. The weighted fusion of the three models ensures prediction accuracy while enhancing the model's robustness and generalization ability, significantly improving the accuracy and reliability of radiation dose rate prediction and providing strong technical support for nuclear radiation early warning.

[0020] Furthermore, the expression for the predicted dose rate of the radiation source obtained by weighted fusion of the three predicted radiation doses is as follows: in, , and For the prediction results The corresponding weights; Let be the dose rate gradient at the current time point t; Dose rate gradient The standard deviation of historical data; and These represent the mean and standard deviation of historical dose rate data, respectively.

[0021] The beneficial effects of this invention are as follows: This solution effectively separates and removes noise from radiation images while preserving key texture details through a dual-branch convolutional neural network. Combined with multi-source data fusion localization technology, it achieves centimeter-level precise positioning of radiation sources. Furthermore, it utilizes a multi-model fusion prediction method to construct a dynamic early warning mechanism, significantly improving the real-time performance, accuracy, and reliability of nuclear radiation monitoring. This system possesses lightweight embedded deployment capabilities, enabling it to adapt to the complex and ever-changing field environment during nuclear accident emergency response, laying the foundation for the rapid detection and location of radiation anomaly sources.

[0022] This solution uses a dual-branch CNN to process radiation images in real time (parameter count < 0.5M, inference speed ≥ 25FPS). It can compress the original image data by 80% before transmitting it back from the drone to the local device, significantly saving network bandwidth resources while ensuring image accuracy.

[0023] This invention extracts multi-dimensional features before prediction, introducing a series of radiation-specific characteristics such as peak percentage, persistence, effective decay constant, and wind influence factor, enabling the model to deeply understand the physical changes in the radiation field. Through a time-varying weight fusion mechanism, the model can adaptively adjust the confidence level of each sub-model according to the characteristics of the current data (stable, gradual, or abrupt), thus achieving faster response and higher prediction accuracy for sudden release events, significantly outperforming methods using a single model or fixed-weight fusion. Attached Figure Description

[0024] Figure 1 This is a flowchart of an online nuclear radiation monitoring method based on convolutional neural networks and multi-source data fusion.

[0025] Figure 2This is a schematic diagram of the system in which the online nuclear radiation monitoring method of this scheme is applied. Detailed Implementation

[0026] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0027] refer to Figure 1 The figure shows a flowchart of an online nuclear radiation monitoring method based on convolutional neural networks and multi-source data fusion; such as Figure 1 As shown, the method S includes steps S1 to S6.

[0028] In step S1, a drone is used to collect lidar point cloud data, radiation dose rate and radiation image of the monitoring area, and a dual-branch convolutional neural network is used to denoise the radiation image to obtain a denoised image. In this scheme, when performing denoising, it is preferable to perform dynamic range logarithmic compression, gamma correction and pixel value normalization on the original image in sequence to obtain a preprocessed image as the radiation image input to the dual-branch convolutional neural network.

[0029] In one embodiment of the present invention, the dual-branch convolutional neural network includes a noise learning unit, a texture learning unit, and a feature fusion unit; The radiometric image is processed sequentially through a multi-scale convolutional module, a receptive field module, and a spatial attention module of the noise learning unit to obtain a noise feature map. Its detailed implementation process is as follows: Multi-scale convolutional modules with kernel sizes of 1×1 and 3×3 (hole rate) The three output feature maps are concatenated along the channel dimension using 5×5 convolution operations. The receptive field module uses 3×3 dilated convolutions with dilation rates of 1, 3, and 5 to process the concatenated feature maps in parallel, and then adds and fuses the results. The spatial attention module performs global average pooling and global max pooling on the fused feature map, concatenates the results, and then passes them through a 7×7 convolutional layer and a sigmoid activation function to generate a spatial attention weight map. This weight map is then multiplied pixel-by-pixel with the fused feature map to output the noise feature map. .

[0030] After the radiometric image is input into the texture learning unit, features are extracted through a lightweight residual network containing three 3×3 convolutional layers. The output of the third convolutional layer is then added to the radiometric image to obtain the texture feature map. ; Feature fusion unit is used to analyze radiometric images, and After fusion, the image is passed through a 1×1 convolutional layer to output the final denoised image. The expression for fusion is: in, and These are the weighting coefficients; Features after fusion; This is a radiation image.

[0031] In step S2, based on the lidar point cloud and the denoised image, the visual odometry pose and the lidar odometry pose are determined using SLAM mapping and trajectory optimization methods, and the two poses are fused to obtain the UAV trajectory position; the detailed implementation process of step S2 is as follows: Point cloud data is processed using the LOAM algorithm. Feature extraction and laser SLAM calculation are performed to obtain the laser odometry pose. ; The denoised image is processed using the ORB-SLAM3 algorithm. Visual feature extraction and visual SLAM calculation are performed to obtain the visual odometry pose. ; The pose is fused based on the Extended Kalman Filter (EKF). and This yields a preliminary fused pose estimate; The graph optimization library g2o was used to perform global optimization on the initially fused pose trajectory, resulting in a high-precision optimized UAV trajectory. .

[0032] In step S3, the radiation dose rate and the UAV trajectory position are spatiotemporally registered to form a dataset, and based on the dataset, Kriging interpolation is used to obtain the continuous radiation field of the monitoring area.

[0033] In a preferred embodiment of this scheme, the method for obtaining the continuous radiation field of the monitoring area based on the dataset using kriging interpolation includes: S31. The anisotropic variogram is used to calculate the variogram values ​​between all known point pairs in the dataset and between each point of the dose rate to be interpolated and all known points. The expression for the anisotropic variogram is: in, Given the variogram values, for a known pair of points i and j, the spatial vector h = p i -p j p i and p j Let i and j be the spatial coordinates of a known pair of points, respectively. For a known point i to point 0 of the dose rate to be interpolated, the spatial vector h = p i -p0, p0 are the spatial coordinates of point 0 of the dose rate to be interpolated; Let h be the value of the spatial vector in the xyz direction; , and For the range in the xyz directions; S32. Based on the variogram values ​​between all known point pairs and between each point of the dose rate to be interpolated and all known points, calculate the weight value corresponding to each point of the dose rate to be interpolated using the Kriging equations: in, , , and These are the variation function values ​​for known point 1 and known point 1, known point 2, known point n, and point 0 of the dose rate to be interpolated, respectively; , , and These are the variation function values ​​of known point 2 and known point 1, known point 2, known point n, and point 0 of the dose rate to be interpolated, respectively; , , and These are the variation function values ​​of known point n and known point 1, known point 2, known point n, and point 0 of the dose rate to be interpolated, respectively; , and These are the weight values ​​of known point 1, known point 2, and known point n, respectively, and point 0 of the dose rate to be interpolated. These are Lagrange multipliers used to constrain the sum of weights to 1; S33. Based on the radiation dose rates of all known points in the dataset and the weight values ​​of the points for which the dose rate to be interpolated is located, calculate the point dose rate for which the dose rate to be interpolated is located: in, The dose rate at point 0 of the dose rate to be interpolated; Let be the radiation dose rate at the i-th known point; The weight values ​​are the i-th known point and the point 0 of the dose rate to be interpolated; S34. After all points 0 of the dose rates to be interpolated are interpolated, the continuous radiation field of the monitoring area is obtained.

[0034] In step S4, based on the continuous radiation field, the L-BFGS-B algorithm is used to minimize the loss function with the radiation source location as the optimization variable, thereby obtaining the location coordinates of the radiation source; the detailed implementation process of step S4 is as follows: S41. Based on the continuous radiation field, calculate the radiation source at the known point using steps S31 to S33. Dose rate at the site ; S42, Based on dose rate Construct a loss function with the location of the radiation source as the optimization variable: in, The loss function; The coordinates of the radiation source location to be optimized; For a known point At point A, the theoretical dose caused by the radiation source; A is the measurement point. The source strength to be sought; This is the regularization term, where λ is the weighting coefficient. for The gradient; Let L2 be the norm of the vector; S43. The L-BFGS-B algorithm is used to minimize the loss function to obtain the coordinates of the radiation source location. .

[0035] Before proceeding to step S5, the number of radiation sources is determined. When there are multiple radiation sources, K-means clustering is used to group the continuous radiation field to obtain the data points corresponding to each group. For each group, the corresponding methods in steps S41 to S43 are used to obtain the location coordinates of each radiation source. After that, steps S5 and S6 need to be executed for each group to obtain the predicted dose of each radiation source.

[0036] In step S5, the radiation dose data in the continuous radiation field is cleaned to obtain a dose rate time series, and multi-dimensional features are extracted to obtain a feature vector composed of multi-dimensional features. The detailed implementation process of data cleaning is as follows: Missing value handling: Dose rate time series data Given dataset D, extract time-series dose rate data {d} t The missing data ({} and its corresponding timestamp {t}) are filled using time series linear interpolation, with the following formula: Outlier handling: Calculate the median of the sequence based on the improved 3σ principle (applicable to non-normal distributions). and interquartile range ;like If a value is found to be outlier, it is replaced with the median, resulting in the cleaned sequence. '.

[0037] In a preferred embodiment of this scheme, the method for multi-dimensional feature extraction of the dose rate time series includes: A sliding window of fixed length is used to extract the mean within the dose rate time series. Standard deviation Maximum value and minimum value : , = ) =max({ '…, '}) =min({ '…, '}) in, The dose rate time series within the sliding window is the value at time point o; t is the current time point. 'This represents the dose rate time series value at time point t-23; Calculate the average gradient of dose rate change within a preset duration adjacent to the current time. : in, and These are the values ​​of the dose rate time series at time points r and r-1, respectively; Calculate periodic time characteristics ,in The weekday indicator variable is essentially a categorical variable code, represented by a number from 0 to 6, which represents each day of the week. For example, I_weekday=1 indicates that the current date is a weekday (usually Monday to Friday), and I_weekday=0 indicates that the current date is a non-working day (i.e., Saturday, Sunday, and public holidays).

[0038] Compare the drone trajectory location with the dose rate time series Perform spatial alignment and calculate the Euclidean distance between the current position of the UAV and the radiation source. : in, Current location of the drone The Euclidean distance to the radiation source's location coordinates; θrel is the azimuth angle of the current UAV position relative to the radiation source; Calculate the percentage of peak values and persistence : , in, It is a counting function; This represents the 0th dose rate value within the sliding window; These represent the mean and standard deviation of the historical background dose, respectively. Calculate the effective decay constant and meteorological influencing factors : in, nuclide The dose contribution weight; nuclide The decay constant; nuclide The dose contribution weight; This is a nuclide index used to traverse all nuclides to calculate the total dose rate; To monitor the prevailing wind direction in the area; and These are the wind speed and wind direction in the monitored area, respectively. The feature vector is constructed by using multi-dimensional features. : .

[0039] In step S6, the feature vectors are input into multiple prediction models to obtain multiple predicted radiation doses, and the three predicted radiation doses are weighted and fused to obtain the predicted dose rate of the radiation source. The multiple prediction models are the XGBoost prediction model, the LSTM-BP neural network prediction model, and the random forest prediction model.

[0040] The XGBoost model has a maximum tree depth of 5, a learning rate of 0.1, and a number of trees of 100, and is trained using a GPU-accelerated histogram algorithm. The LSTM-BP neural network is trained using the Adam optimizer, with mean squared error (MSE) as the main loss function, 50 training epochs, and a batch size of 32. The random forest model has a number of trees of 200 and a maximum depth of 8.

[0041] In one embodiment of the present invention, the prediction result obtained by the XGBoost prediction model for: in, The number of training samples; For the model to the first Predicted dose rate for each sample; For the first The true dose rate of each sample; The loss function; These are parameters for controlling model complexity. These are the weight coefficients for the L2 regularization term; This is a vector of weights for all leaves in the entire model; T is the total number of decision trees. These are the weight coefficients for the L1 regularization term; For the first The leaf weight vector of each tree; The prediction results of the LSTM-BP neural network prediction model for: in, Mean square error, , This refers to the number of samples in a batch. The weighting coefficient for the spectral loss term is a preset positive number; This is a Fourier transform function used to extract spectral features from the dose rate sequence; The square of the L2 norm of the vector; and These are the sequence vectors of the predicted values ​​and the sequence vectors of the true values, respectively. Each of the batches The sequence vector of predicted values ​​and the sequence vector of true values ​​for each sample; The prediction results of the random forest prediction model for: Where Z is the total number of decision trees in the random forest model, z is the index of the decision tree, z=1,2,…,Z; Representing the z-th decision tree model, this function receives input features and outputs a predicted value; This is the multi-dimensional feature vector of the input model.

[0042] The expression for the predicted dose rate of a radiation source obtained by weighted fusion of three predicted radiation doses is as follows: in, , and For the prediction results The corresponding weights; Let be the dose rate gradient at the current time point t; Dose rate gradient The standard deviation of historical data; and These represent the mean and standard deviation of historical dose rate data, respectively.

[0043] In application, the radiation online monitoring method of this solution can be deployed on a ground station server or a cloud platform. For details, please refer to [link / reference needed]. Figure 2 In this scheme, steps S1 and S2 are located on the UAV system, while steps S3 to S6 are located at the ground center.

[0044] In summary, the radiation online monitoring method proposed in this scheme innovatively combines advanced deep learning image processing, precise multi-source sensor fusion positioning, and multi-dimensional feature extraction, effectively solving key pain points in nuclear radiation monitoring such as poor image quality, low positioning accuracy, and weak prediction capabilities. It has high practical value and promising prospects for widespread application.

Claims

1. A method for online monitoring of nuclear radiation based on convolutional neural networks and multi-source data fusion, characterized in that, Including the following steps: S1. Use a drone to collect lidar point cloud data, radiation dose rate and radiation image of the monitoring area, and use a dual-branch convolutional neural network to denoise the radiation image to obtain a denoised image. S2. Based on the lidar point cloud and denoised image, the visual odometry pose and lidar odometry pose are determined using SLAM mapping and trajectory optimization methods, and the two poses are fused to obtain the UAV trajectory position. S3. Spatiotemporal registration of radiation dose rate and UAV trajectory position is performed to form a dataset, and based on the dataset, Kriging interpolation is used to obtain the continuous radiation field of the monitoring area. S4. Based on the continuous radiation field, the L-BFGS-B algorithm is used to minimize the loss function with the radiation source location as the optimization variable to obtain the location coordinates of the radiation source. S5. Perform data cleaning on the radiation dose in the continuous radiation field to obtain the dose rate time series, and extract multi-dimensional features to obtain the feature vector composed of multi-dimensional features. S6. Input the feature vectors into multiple prediction models to obtain multiple predicted radiation doses, and then weight and fuse the three predicted radiation doses to obtain the predicted dose rate of the radiation source.

2. The method for online monitoring of nuclear radiation according to claim 1, characterized in that, The dual-branch convolutional neural network includes a noise learning unit, a texture learning unit, and a feature fusion unit; The radiometric image is processed sequentially through a multi-scale convolutional module, a receptive field module, and a spatial attention module of the noise learning unit to obtain a noise feature map. ; After the radiometric image is input into the texture learning unit, features are extracted through a lightweight residual network containing three 3×3 convolutional layers. The output of the third convolutional layer is then added to the radiometric image to obtain the texture feature map. ; Feature fusion unit is used to analyze radiometric images, and After fusion, the image is passed through a 1×1 convolutional layer to output the final denoised image. The expression for fusion is: in, and These are the weighting coefficients; Features after fusion; This is a radiation image.

3. The method for online monitoring of nuclear radiation according to claim 1, characterized in that, Methods for obtaining the continuous radiation field of the monitoring area based on the dataset and using kriging interpolation include: S31. The anisotropic variogram is used to calculate the variogram values ​​between all known point pairs in the dataset and between each point of the dose rate to be interpolated and all known points. The expression for the anisotropic variogram is: in, Given the variogram values, for a known pair of points i and j, the spatial vector h = p i -p j p i and p j Let i and j be the spatial coordinates of a known pair of points, respectively. For a known point i to point 0 of the dose rate to be interpolated, the spatial vector h = p i -p0, p0 are the spatial coordinates of point 0 of the dose rate to be interpolated; Let h be the value of the spatial vector in the xyz direction; , and For the range in the xyz directions; S32. Based on the variogram values ​​between all known point pairs and between each point of the dose rate to be interpolated and all known points, calculate the weight value corresponding to each point of the dose rate to be interpolated using the Kriging equations: in, , , and These are the variation function values ​​for known point 1 and known point 1, known point 2, known point n, and point 0 of the dose rate to be interpolated, respectively; , , and These are the variation function values ​​of known point 2 and known point 1, known point 2, known point n, and point 0 of the dose rate to be interpolated, respectively; , , and These are the variation function values ​​of known point n and known point 1, known point 2, known point n, and point 0 of the dose rate to be interpolated, respectively; , and These are the weight values ​​of known point 1, known point 2, and known point n, respectively, and point 0 of the dose rate to be interpolated. These are Lagrange multipliers used to constrain the sum of weights to 1; S33. Based on the radiation dose rates of all known points in the dataset and the weight values ​​of the points for which the dose rate to be interpolated is located, calculate the point dose rate for which the dose rate to be interpolated is located: in, The dose rate at point 0 of the dose rate to be interpolated; Let be the radiation dose rate at the i-th known point; The weight values ​​are the i-th known point and the point 0 of the dose rate to be interpolated; S34. After all points 0 of the dose rates to be interpolated are interpolated, the continuous radiation field of the monitoring area is obtained.

4. The method for online monitoring of nuclear radiation according to claim 3, characterized in that, Step S4 further includes: S41. Based on the continuous radiation field, calculate the radiation source at the known point using steps S31 to S33. Dose rate at the site ; S42, Based on dose rate Construct a loss function with the location of the radiation source as the optimization variable: in, The loss function; The coordinates of the radiation source location to be optimized; For a known point At point A, the theoretical dose caused by the radiation source; A is the measurement point. The source strength to be sought; This is the regularization term, where λ is the weighting coefficient. for The gradient; Let L2 be the norm of the vector; S43. The L-BFGS-B algorithm is used to minimize the loss function to obtain the coordinates of the radiation source location. .

5. The method for online monitoring of nuclear radiation according to claim 4, characterized in that, Before proceeding to step S5, the number of radiation sources is also determined. When there are multiple radiation sources, K-means clustering is used to group the continuous radiation field according to the continuous radiation field, and data points are obtained for each group. For each group, the corresponding methods of steps S41 to S43 are used to obtain the location coordinates of each radiation source; then, steps S5 and S6 are performed for each group to obtain the predicted dose of each radiation source.

6. The method for online monitoring of nuclear radiation according to claim 3, characterized in that, Methods for multi-dimensional feature extraction of dose rate time series include: A sliding window of fixed length is used to extract the mean within the dose rate time series. Standard deviation Maximum value and minimum value : 、 = ) =max({ '…, '})、 =min({ '…, '}) in, The dose rate time series within the sliding window is the value at time point o; t is the current time point. 'This represents the dose rate time series value at time point t-23; the fixed length is equal to 23.' Calculate the average gradient of dose rate change within a preset duration adjacent to the current time. : in, and These are the values ​​of time points r and r-1 in the dose rate time series, respectively; the preset duration is 5. Calculate periodic time characteristics ,in For working day indicator variables; Compare the drone trajectory location with the dose rate time series Perform spatial alignment and calculate the Euclidean distance between the current position of the UAV and the radiation source. : in, Current location of the drone Coordinates of the radiation source location The Euclidean distance; θrel is the azimuth angle of the current UAV position relative to the radiation source; Calculate the percentage of peak values and persistence : ， in, It is a counting function; This represents the 0th dose rate value within the sliding window; These represent the mean and standard deviation of the historical background dose, respectively. Calculate the effective decay constant and meteorological influencing factors : in, nuclide The dose contribution weight; nuclide The decay constant; nuclide The dose contribution weight; This is a nuclide index used to traverse all nuclides to calculate the total dose rate; To monitor the prevailing wind direction in the area; and These are the wind speed and wind direction in the monitored area, respectively. The feature vector is constructed by using multi-dimensional features. : 。 7. The method for online monitoring of nuclear radiation according to claim 6, characterized in that, Multiple prediction models were used, including the XGBoost prediction model, the LSTM-BP neural network prediction model, and the random forest prediction model.

8. The method for online monitoring of nuclear radiation according to claim 7, characterized in that, The prediction results obtained by the XGBoost prediction model for: in, The number of training samples; For the model to the first Predicted dose rate for each sample; For the first The true dose rate of each sample; The loss function; These are parameters for controlling model complexity. These are the weight coefficients for the L2 regularization term; This is a vector of weights for all leaves in the entire model; T is the total number of decision trees. These are the weight coefficients for the L1 regularization term; For the first The leaf weight vector of each tree; The prediction results of the LSTM-BP neural network prediction model for: in, Mean square error, , This refers to the number of samples in a batch. The weighting coefficient for the spectral loss term is a preset positive number; This is a Fourier transform function used to extract spectral features from the dose rate sequence; The square of the L2 norm of the vector; and These are the sequence vectors of the predicted values ​​and the sequence vectors of the true values, respectively. Each of the batches The sequence vector of predicted values ​​and the sequence vector of true values ​​for each sample; The prediction results of the random forest prediction model for: Where Z is the total number of decision trees in the random forest model, z is the index of the decision tree, z=1,2,…,Z; Representing the z-th decision tree model, this function receives input features and outputs a predicted value; This is the multi-dimensional feature vector of the input model.

9. The method for online monitoring of nuclear radiation according to claim 8, characterized in that, The expression for the predicted dose rate of a radiation source obtained by weighted fusion of three predicted radiation doses is as follows: in, , and For the prediction results The corresponding weights; This represents the dose rate gradient at the current time point t. Dose rate gradient The standard deviation of historical data; and These represent the mean and standard deviation of historical dose rate data, respectively.

Images