Nuclear power plant surrounding ground feature element extraction method based on deep learning

Abstract

The invention belongs to the technical field of remote sensing, and particularly relates to a nuclear power plant surrounding ground feature element extraction method based on deep learning, and the method comprises the steps: 1, making a nuclear power plant surrounding ground feature deep learning sample; step 2, based on the HRNet model, constructing a Mama attention residual module, an inner convolution residual module and an asynchronous convolution residual module through Mama, inner convolution and heterogeneous convolution, and constructing an improved HRNet model; 3, training of a U-Net model, a Deeplabv3 model, a HRNet model, a Swin Transformer model and an improved HRNet model is completed through the sample set training data and the verification data; 4, evaluating land element extraction results of different deep learning models; and 5, extracting elements of the surrounding land of the large-scale nuclear power plant by using the trained improved HRNet model. The method can effectively improve the extraction precision of complex ground feature elements around the nuclear power plant, remarkably improves the calculation efficiency of the model, and is suitable for efficient monitoring and classification of multiple ground feature elements around the nuclear power plant in a large range.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing technology, and in particular relates to a method for extracting ground features around a nuclear power plant based on deep learning. Background Art

[0002] The monitoring of the surrounding features of nuclear power plants plays an important role in the safe operation and environmental protection of nuclear power plants. During the operation of nuclear power plants, there may be potential impacts on the surrounding natural environment and infrastructure (such as trees, lakes, roads, buildings and tanks). Therefore, accurate and efficient monitoring of these elements is of great significance for the operation and maintenance of nuclear power plants.

[0003] The monitoring of land features around nuclear power plants mostly relies on on-site inspections and manual analysis. Such methods not only require a lot of manpower and material resources, but are also inefficient when faced with a wide monitoring area. More importantly, the results of manual analysis are easily affected by subjective factors, which may lead to deviations in monitoring results and fail to provide comprehensive and accurate monitoring data. Especially in high-risk environments such as nuclear power plants, manual methods also have potential safety hazards.

[0004] In recent years, with the advancement of remote sensing technology, environmental monitoring around nuclear power plants has gradually shifted from manual operations to automated analysis based on high-resolution images. Remote sensing images acquired by satellites and drones can cover a large area and update data regularly, becoming an important tool for environmental monitoring. Compared with traditional image processing methods, automated models can more accurately extract environmental elements around nuclear power plants, such as roads, buildings, trees, and water bodies, by learning features from large amounts of data. This type of method does not rely on manually set rules and can extract more detailed features from images, improving the accuracy and efficiency of monitoring.

[0005] Although deep learning models have shown great potential in remote sensing image processing, existing models still have some limitations in extracting environmental elements around nuclear power plants. For example, when faced with complex environmental backgrounds, the accuracy and stability of extracting specific elements may be disturbed. The current commonly used models perform well in identifying regular objects such as buildings and roads, but often get confused when identifying natural objects (such as trees and lakes). This requires improvements to existing models to improve the model's extraction accuracy on different environmental elements. Summary of the invention

[0006] The purpose of the present invention is to provide a method for extracting land features around nuclear power plants based on deep learning. The method can more efficiently and accurately identify key elements such as trees, lakes, roads, buildings and tanks, effectively improve the extraction accuracy of complex land features around nuclear power plants, and significantly improve the calculation efficiency of the model. It is suitable for efficient monitoring and classification of a large range and multiple land features around nuclear power plants.

[0007] The technical solution to achieve the purpose of the present invention is:

[0008] A method for extracting ground features around a nuclear power plant based on deep learning, the method comprising:

[0009] Step 1: Preparation of deep learning samples of land elements around nuclear power plants;

[0010] Step 2: Based on the HRNet model, the Mamba attention residual module, the inner convolution residual module, and the asynchronous convolution residual module are constructed through Mamba, inner convolution, and heterogeneous convolution to construct an improved HRNet model.

[0011] Step 3: Use the sample set training data and verification data to complete the training of U-Net, Deeplabv3, HRNet, SwinTransformer and improved HRNet models respectively;

[0012] Step 4: Evaluation of the results of land element extraction using different deep learning models;

[0013] Step 5: Use the trained improved HRNet model to extract large-scale land elements around nuclear power plants.

[0014] The step 1 is specifically as follows: using Gaofen-2 remote sensing images to construct a sample set of land elements around the nuclear power plant, using intelligent manual interactive sample annotation, using data enhancement technology to process the original data, and after annotation and enhancement, cropping the image, and after cropping, dividing the sample into a training set, a validation set, and a test set.

[0015] The sample set of land elements around the nuclear power plant in step 1 includes: vegetation, lakes, roads, buildings and tanks around the nuclear power plant.

[0016] The size of the image cropped in step 1 is 1024×1024.

[0017] The step 2 is specifically as follows: Mamba attention constructs a Mamba attention residual module by dynamically adjusting weights to replace the ordinary residual module in the Stem of HRNet; the inner convolution operator reverses the properties of the convolution, generates a convolution kernel for each position, groups them by channel, shares channel parameters within the group, and constructs an inner convolution residual module to replace the ordinary residual module in the stage-1 of the main network of HRNet; heterogeneous convolution performs input features Figure 1 Some channels apply k×k convolution kernels, and the remaining channels apply 1×1 convolution kernels, and P is used to control the proportion of convolution kernels to k. An asynchronous convolution residual module is constructed to replace the ordinary residual module in stage-4 of the HRNet main network.

[0018] The step 3 is specifically as follows: training is performed using the improved HRNet model and the pre-divided training set and validation set; during the training process, the model generates a predicted value through forward propagation, and uses a binary cross entropy loss function to calculate the error between the predicted value and the true value; the error updates the model parameters through back propagation, and uses the gradient descent method to adjust the model's weights and biases to minimize the loss function value; the model is evaluated on the validation set to prevent overfitting.

[0019] The step 4 is specifically as follows: using the model trained in step 3 to extract the results in the test set respectively, and using the following formula to quantify the prediction results:

[0020]

[0021] In the formula,

[0022] Recall: Recall rate, which represents the ratio of the number of correct targets identified by the model to the number of actual targets, and is used to measure the model's ability to identify positive samples;

[0023] Precision: Precision is the ratio of the number of correct targets identified by the model to the number of all targets identified by the model. It is used to measure how many of the samples predicted as positive by the model are correct.

[0024] F1-score: represents the harmonic mean of precision and recall, which is used to comprehensively evaluate the accuracy and recognition ability of the model;

[0025] IoU: An indicator used to measure the degree of overlap between the predicted result and the actual target area;

[0026] TP: predicted as a positive example, the actual one is a positive example; FP: predicted as a positive example, the actual one is a negative example; FN: predicted as a negative example, the actual one is a positive example.

[0027] The step 5 is specifically as follows: the remote sensing image of the land elements around the large-scale nuclear power plant is cropped into grid blocks of uniform size, and each grid block is predicted one by one using the trained improved HRNet model to obtain the corresponding land element classification prediction results; and the land elements of the large area around the nuclear power plant are extracted by splicing the prediction results of each grid block into an overall image.

[0028] The beneficial technical effects of the present invention are:

[0029] The present invention provides a method for extracting features of land objects around nuclear power plants based on deep learning. On the basis of the HRNet model, the Mamba attention module, heterogeneous convolution (HetConv) and inner convolution (Involution) are used to improve the basic residual structure, and the basic residual structure in HRNet is replaced by the improved module, thereby realizing a lightweight design of the model and improving the accuracy and computational efficiency of feature extraction.

[0030] 1. More accurate feature extraction: The present invention provides a method for extracting features of land objects around nuclear power plants based on deep learning. By introducing the Mamba attention module, the model can dynamically adjust weights at different scales and accurately capture the detailed features of complex land objects (such as trees, lakes, roads, buildings, and tanks) around nuclear power plants. Compared with traditional models, the Mamba module improves the ability to extract complex features, especially for small and difficult to distinguish targets.

[0031] 2. More efficient computing performance: The present invention provides a method for extracting features of land objects around nuclear power plants based on deep learning. Combining heterogeneous convolution (HetConv) and inner convolution (Involution) technology, the model achieves a lightweight design and significantly reduces the number of parameters and computational complexity. While ensuring accuracy, the efficiency of the model is improved, which is particularly suitable for large-scale nuclear power plant surrounding environment monitoring tasks, so that the model has higher computing performance when processing large-scale images.

[0032] 3. More efficient computing performance: The present invention provides a method for extracting features of land objects around nuclear power plants based on deep learning. Combining heterogeneous convolution (HetConv) and inner convolution (Involution) technology, the model achieves a lightweight design and significantly reduces the number of parameters and computational complexity. While ensuring accuracy, the efficiency of the model is improved, which is particularly suitable for large-scale nuclear power plant surrounding environment monitoring tasks, so that the model has higher computing performance when processing large-scale images. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flow chart of a method for extracting ground features around a nuclear power plant based on deep learning provided by the present invention;

[0034] Figure 2 1 is a structural diagram of an improved HRNet model constructed in an embodiment of the present invention;

[0035] Figure 3 Schematic diagram of the HetConv heterogeneous convolution structure constructed in an embodiment of the present invention;

[0036] Figure 4 Four residual network structure diagrams in the embodiments of the present invention are shown below: Figure 4(a) is the structure diagram of the common residual module in stage 1 of the HRNet main network. Figure 4 (b) is the structure diagram of Mamba attention residual module. Figure 4 (c) is the structure diagram of the involution residual module. Figure 4 (d) is the structure diagram of the asynchronous convolution residual module;

[0037] Figure 5 Some prediction results under different network models: Figure 5 (a) is the original image. Figure 5 (b) is the label map, Figure 5 (c) is a partial prediction result diagram of the U-Net model. Figure 5 (d) is a partial prediction result diagram of the Deeplabv3 model. Figure 5 (e) is a partial prediction result diagram of the HRNet model. Figure 5 (f) is a partial prediction result diagram of the Swin Transformer model. Figure 5 (g) is a partial prediction result diagram of the improved HRNet model;

[0038] Figure 6 The result map of using the trained improved HRNet model to extract land elements around the nuclear power plant on a large scale: Figure 6 (a) is the original image. Figure 6 (b) is the overall prediction result diagram of the improved HRNet model. DETAILED DESCRIPTION

[0039] The present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0040] like Figure 1 As shown, the present invention provides a method for extracting ground features around a nuclear power plant based on deep learning, which specifically includes the following steps:

[0041] Step 1: Preparation of deep learning samples of land elements around nuclear power plants:

[0042] Gaofen-2 remote sensing images were used to construct a sample set of land elements around nuclear power plants, mainly including vegetation, lakes, roads, buildings and tanks around nuclear power plants. In the sample production process, an intelligent manual interactive sample annotation method was adopted to improve the annotation efficiency and accuracy. In order to further expand the data volume, data enhancement techniques (such as rotation, flipping, scaling, etc.) were used to expand the original data. After annotation and enhancement, the images were cropped to 1024×1024 size pictures. Finally, the processed samples were divided into training set, validation set and test set to ensure the effectiveness and generalization ability of model training.

[0043] Step 2: Based on the HRNet model, Mamba attention residual module, involution residual module, asynchronous convolution residual module are constructed through Mamba, Involution, and HetConv to build an improved HRNet model.

[0044] The Mamba module is used to capture detailed features of different scales, HetConv improves the recognition efficiency of roads and tanks, and Involution enhances the ability to extract boundaries of complex elements.

[0045] Mamba attention can effectively capture the details of surrounding elements at different scales by dynamically adjusting weights. Therefore, we use it to build a Mamba attention residual module and replace the ordinary residual block in the Stem of the HRNet structure, such as Figure 4 As shown in (b), it is used in the Stem of HRNet to better and finer capture target features. Considering that the HRNet model is relatively large. Involution reverses the properties of convolution, making the operator spatially specific and channel invariant. The inner convolution operator generates a convolution kernel for each position. In order to reduce parameters, it is grouped by channel, and the channel parameters within the group are shared. The receptive field of the convolution kernel is increased, thereby improving the performance of the network. Therefore, Involution is used to construct the inner convolution residual module, as shown in Figure 4 As shown in (c), replace the ordinary residual module in stage-1 of the HRNet main network, such as Figure 4 (a) shows. HetConv can be used to input features Figure 1 Some channels apply k×k convolution kernels, and the remaining channels apply 1×1 convolution kernels, and use P to control the proportion of convolution kernels to k. Here we use heterogeneous convolution with k=3 and P=2 as shown in Figure 3 As shown in Figure 2, the efficiency of the convolutional neural network is improved without losing accuracy. The asynchronous convolution residual module is constructed using HetConv, as shown in Figure 2. Figure 4 (d) and used to replace the ordinary residual block in stage-4 of the HRNet main network. These three improved residual modules can ensure that the input and output feature maps are C×H×W. Solid can complete the replacement of the above ordinary residual module position. The final improved HRNet model structure is as follows Figure 2 As shown in the figure, the Mamba attention residual module, the inward residual module and the asynchronous convolution residual module are used to replace the ordinary residual blocks in the original HRNet structure Stem, Stage-1 and Stage-4 respectively.

[0046] Step 3: Use the sample set training data and verification data to complete the training of U-Net, Deeplabv3, HRNet, SwinTransformer and improved HRNet models respectively.

[0047] The experiments of the present invention are carried out on the Ubuntu 20.01 operating system built with Docker. The network models are all trained using NVIDIA TESLAV100 32G, and the network model codes are all in the PyTorch framework. During the training process, the optimizer we used is SDG. In the process of training the model, the binary cross entropy loss function is used to guide and regulate the back propagation process of the model, so that the model parameters are continuously updated to tend to the optimal result. Specifically, the binary cross entropy loss function provides an error signal for each iteration by calculating the difference between the predicted value of the model prediction output and the true value of the true label. The error is transmitted backward through the back propagation algorithm, and then the weights and biases of the model are adjusted according to the gradient descent method to minimize the loss function value. The formula of the binary cross entropy loss function is:

[0048]

[0049] Where: y is the true value after binarization, is the predicted value after binarization.

[0050] In this step, the training data and validation data of the sample set were used to complete the training of U-Net, Deeplabv3, HRNet, Swin Transformer and improved HRNet models. The training method includes data preparation, model initialization, setting hyperparameters, iterative training and validation, and model optimization. Specifically, the output results are calculated through forward propagation, the error is calculated using the loss function, back propagation is performed to update the parameters, and the model performance is evaluated on the validation set to prevent overfitting. Finally, the model is tuned and regularized to achieve the best performance.

[0051] Step 4: Evaluation of land element extraction results using different deep learning models:

[0052] In order to demonstrate the advancedness of the improved HRNet, the model trained in step 3 is used to extract the results in the test set. The prediction results are quantified using the following formula:

[0053]

[0054] In the formula,

[0055] Recall: It represents the ratio of the number of correct targets identified by the model to the number of actual targets, and is used to measure the model's ability to recognize positive samples.

[0056] Precision: It represents the ratio of the number of correct targets identified by the model to the total number of targets identified by the model. It is used to measure how many of the samples predicted as positive by the model are correct.

[0057] F1-score: The harmonic average of precision and recall, used to comprehensively evaluate the accuracy and recognition ability of the model.

[0058] IoU: It is an indicator used to measure the degree of overlap between the predicted result and the actual target area.

[0059] True Positive (TP): predicted as a positive example, the actual example is a positive example; False Positive (FP): predicted as a positive example, the actual example is a negative example; False Negative (FN): predicted as a negative example, the actual example is a positive example.

[0060] Since the present invention targets five types of land elements, the overall mRecall (meanRecall), mPrecision (meanPrecision), mF1-score (meanF1-score), and mIoU (meanIoU) are calculated based on the Recall, Precision, F1-score, and IoU of each type of land element. The overall mRecall (meanRecall), mPrecision (meanPrecision), mF1-score (meanF1-score), and mIoU (meanIoU) are obtained by adding the Recall, Precision, F1-score, and IoU of each type of land element, and calculating their average values. Specifically, these overall indicators are the average values ​​of the corresponding indicators of each category, which are used to comprehensively evaluate the performance of the model in all categories.

[0061] The evaluation indicators of the test set in different network models are shown in Table 1. It can be seen from Table 1 that the improved HRNet exceeds some other convolutional neural networks in terms of performance. Figure 5 The results of land element extraction around the nuclear power plant show that the five types of land elements predicted by the improved HRNet are basically consistent with the labels.

[0062] Table 1 Average evaluation index of 5 categories of test sets in different network models

[0063]

[0064] Step 5: Use the trained improved HRNet model to extract large-scale land elements around the nuclear power plant:

[0065] Using the trained improved HRNet model, we first cropped the large-scale remote sensing image into 1024×1024 grid blocks, and then predicted each grid block one by one to obtain the corresponding land feature classification prediction results. By splicing the prediction results of each grid block into an overall image, we can achieve continuous and accurate extraction of land features in a large area around the nuclear power plant. This method can not only maintain the model's recognition accuracy for small area details, but also achieve efficient and complete feature detection in large-scale scenarios, such as Figure 6 The results shown demonstrate the superior performance of this method in monitoring land elements around large-scale nuclear power plants.

[0066] The above results show that the method of the present invention has been trained and verified by the model, and the improved HRNet model is used to perform large-scale land element classification prediction, and the accurate monitoring of the surrounding environment of the nuclear power plant is achieved through image segmentation results.

[0067] The present invention is described in detail above with reference to the accompanying drawings and embodiments, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of ordinary technicians in the field without departing from the purpose of the present invention. The contents not described in detail in the present invention can adopt the existing technology.

Claims

1. A method for extracting ground features around a nuclear power plant based on deep learning, characterized in that: The method comprises: Step 1: Preparation of deep learning samples of land elements around nuclear power plants; Step 2: Based on the HRNet model, the Mamba attention residual module, the inner convolution residual module, and the asynchronous convolution residual module are constructed through Mamba, inner convolution, and heterogeneous convolution to construct an improved HRNet model. Step 3: Use the sample set training data and verification data to complete the training of U-Net, Deeplabv3, HRNet, SwinTransformer and improved HRNet models respectively; Step 4: Evaluation of the results of land element extraction using different deep learning models; Step 5: Use the trained improved HRNet model to extract large-scale land elements around nuclear power plants.

2. According to claim 1, a method for extracting ground features around a nuclear power plant based on deep learning is characterized in that: The step 1 is specifically as follows: using Gaofen-2 remote sensing images to construct a sample set of land elements around the nuclear power plant, using intelligent manual interactive sample annotation, using data enhancement technology to process the original data, and after annotation and enhancement, cropping the image, and after cropping, dividing the sample into a training set, a validation set, and a test set.

3. The method for extracting ground features around a nuclear power plant based on deep learning according to claim 2, characterized in that: The sample set of land elements around the nuclear power plant in step 1 includes: vegetation, lakes, roads, buildings and tanks around the nuclear power plant.

4. The method for extracting ground features around a nuclear power plant based on deep learning according to claim 2, characterized in that: The size of the image cropped in step 1 is 1024×1024.

5. The method for extracting ground features around a nuclear power plant based on deep learning according to claim 1, characterized in that: The step 2 is specifically as follows: Mamba attention constructs a Mamba attention residual module by dynamically adjusting weights to replace the ordinary residual module in the Stem of HRNet; the inner convolution operator reverses the properties of the convolution, generates a convolution kernel for each position, groups them by channels, shares channel parameters within the group, and constructs an inner convolution residual module to replace the ordinary residual module in the stage-1 of the HRNet main network; heterogeneous convolution applies k×k convolution kernels to part of the channels of the input feature map, and 1×1 convolution kernels to the remaining channels, and uses P to control the proportion of the convolution kernel to k, to construct an asynchronous convolution residual module to replace the ordinary residual module in the stage-4 of the HRNet main network.

6. The method for extracting ground features around a nuclear power plant based on deep learning according to claim 1, characterized in that: The step 3 is specifically as follows: training is performed using the improved HRNet model and the pre-divided training set and validation set; during the training process, the model generates a predicted value through forward propagation, and uses a binary cross entropy loss function to calculate the error between the predicted value and the true value; the error updates the model parameters through back propagation, and uses the gradient descent method to adjust the model's weights and biases to minimize the loss function value; the model is evaluated on the validation set to prevent overfitting.

7. The method for extracting ground features around a nuclear power plant based on deep learning according to claim 1, characterized in that: The step 4 is specifically as follows: using the model trained in step 3 to extract the results in the test set respectively, and using the following formula to quantify the prediction results: In the formula, Recall: Recall rate, which represents the ratio of the number of correct targets identified by the model to the number of actual targets, and is used to measure the model's ability to identify positive samples; Precision: Precision is the ratio of the number of correct targets identified by the model to the number of all targets identified by the model. It is used to measure how many of the samples predicted as positive by the model are correct. F1-score: represents the harmonic mean of precision and recall, which is used to comprehensively evaluate the accuracy and recognition ability of the model; IoU: An indicator used to measure the degree of overlap between the predicted result and the actual target area; TP: predicted as positive, actually positive; FP: predicted as positive example, but actually negative example; FN: The prediction is a negative example, but the actual example is a positive example.

8. The method for extracting ground features around a nuclear power plant based on deep learning according to claim 1, characterized in that: The step 5 is specifically as follows: the remote sensing image of the land elements around the large-scale nuclear power plant is cropped into grid blocks of uniform size, and each grid block is predicted one by one using the trained improved HRNet model to obtain the corresponding land element classification prediction results; and the land elements of the large area around the nuclear power plant are extracted by splicing the prediction results of each grid block into an overall image.

Images