Non-supervision detection method, device and equipment for tangential direction of magnetic top layer and medium

Abstract

The invention discloses an unsupervised detection method, device and equipment for a tangential direction of a magnetic top layer and a medium. The method comprises the following steps: generating soft X-ray image data of a source domain and a target domain of the magnetic top layer of a target observation point; determining a true value label; respectively inputting the source domain soft X-ray image data and the target domain soft X-ray image data into a residual attention U-Net network to obtain a source prediction result and a target prediction result; based on the error between the source prediction result and the true value label and the error between the source prediction result and the target prediction result, carrying out first-round training on the residual attention U-Net network to obtain a preliminary network model; processing the target domain soft X-ray image data by using the preliminary network model to obtain a pseudo label of the target domain soft X-ray image data; performing two rounds of training on the preliminary network model based on the target domain soft X-ray image data and the pseudo tag to obtain a final network model; and detecting the soft X-ray image data of the target domain by using the final network model.

Description

Technical Field

[0001] The present invention relates to the field of semantic segmentation technology, and in particular to an unsupervised detection method, device, equipment and medium for the tangent direction of a magnetic top layer. Background Art

[0002] The magnetopause is an important region where the Earth's magnetosphere interacts with the solar wind. It is located outside the Earth's atmosphere and is the place where the Earth's magnetic field reconnects with the solar wind magnetic field. The study of the magnetopause has far-reaching scientific and application significance for understanding the microscopic mechanism of the interaction between the solar wind and the Earth's magnetosphere, as well as its impact on the Earth's space weather. Soft X-ray imaging technology is widely used in the observation of the magnetopause. This technology generates soft X-ray images of the magnetospheric system by detecting soft X-ray photons radiated when the solar wind interacts with particles in the Earth's space environment. These images contain a lot of important information about the large-scale structure of the magnetopause, such as the tangent direction of the magnetopause. Among them, SXI (Soft X-ray Imager) is a specific instrument or equipment that implements soft X-ray imaging technology.

[0003] However, due to the complexity of the SXI imaging principle and the influence of correlated noise, the generated SXI images often have a low signal-to-noise ratio, making it difficult to accurately identify the magnetopause structure from the image. Secondly, due to changes in the observation angle and the satellite orbital position, the magnetopause images captured from different viewpoints have different shapes, which increases the difficulty of image analysis and detection. In addition, current research relies on simulated data, but there may be differences between simulated data and real data, which affects the accuracy and reliability of the model in practical applications. At the same time, the annotation of real data also consumes a lot of time and energy. Summary of the invention

[0004] In view of this, the present invention provides a method, device, equipment and medium for unsupervised detection of the tangent direction of the top magnetic layer, which can directly detect the tangent direction of the top magnetic layer in the low signal-to-noise ratio SXI images obtained by different observation angles.

[0005] In the first aspect, the present invention provides an unsupervised detection method for the tangent direction of the top magnetic layer, the method comprising: using the Jorgensen&Sun model to generate source domain soft X-ray image data of the top magnetic layer of the target observation point; using the PPMLR-MHD model to generate target domain soft X-ray image data of the top magnetic layer of the target observation point; determining the tangent direction of the top magnetic layer corresponding to the source domain soft X-ray image data as a true value label; inputting the source domain soft X-ray image data and the target domain soft X-ray image data into a residual attention U-Net network respectively to obtain a source prediction result and a target prediction result; based on the error between the source prediction result and the true value label, and the error between the source prediction result and the target prediction result, performing a first round of training on the residual attention U-Net network to obtain a preliminary network model; using the preliminary network model to process the target domain soft X-ray image data to obtain a pseudo label of the target domain soft X-ray image data; based on the target domain soft X-ray image data and the pseudo label, performing a second round of training on the preliminary network model to obtain a final network model; and using the final network model to detect the target domain soft X-ray image data.

[0006] In the second aspect, the present invention provides an unsupervised detection device for the tangent direction of the top magnetic layer, the device comprising: a source data acquisition module, used to generate source domain soft X-ray image data of the top magnetic layer of the target observation point using the Jorgensen & Sun model; a target data acquisition module, used to generate target domain soft X-ray image data of the top magnetic layer of the target observation point using the PPMLR-MHD model; a true value label determination module, used to determine the tangent direction of the top magnetic layer corresponding to the source domain soft X-ray image data as the true value label; a result prediction module, used to input the source domain soft X-ray image data and the target domain soft X-ray image data into the residual attention U-Net network respectively to obtain the source domain soft X-ray image data. prediction results and target prediction results; a first-round training module, used to perform a first-round training on the residual attention U-Net network based on the error between the source prediction result and the true value label, and the error between the source prediction result and the target prediction result, to obtain a preliminary network model; a pseudo-label acquisition module, used to use the preliminary network model to process the target domain soft X-ray image data to obtain the pseudo-label of the target domain soft X-ray image data; a second-round training module, used to perform a second-round training on the preliminary network model based on the target domain soft X-ray image data and the pseudo-label to obtain the final network model; an image detection module, used to use the final network model to detect the target domain soft X-ray image data.

[0007] In a third aspect, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to execute a method for unsupervised detection of the direction of a magnetic top layer tangent according to the first aspect or any corresponding embodiment thereof.

[0008] In a fourth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute a method for unsupervised detection of a magnetic top layer tangent direction according to the first aspect or any corresponding embodiment thereof.

[0009] In a fifth aspect, the present invention provides a computer program product, comprising computer instructions, wherein the computer instructions are used to enable a computer to execute an unsupervised detection method for the direction of a magnetic top layer tangent according to the first aspect or any corresponding embodiment thereof.

[0010] The technical solution provided by one or more embodiments of the present invention first uses two different models to simulate the three-dimensional structure of the magnetosphere top under different solar wind densities, integrates along multiple observation angles to obtain two-dimensional integral images of the magnetosphere top at different viewing angles, and performs SXI simulation on all these integral images to obtain simulated soft X-ray image data of the magnetosphere top at a preset observation point. Secondly, the present invention designs a residual attention U-Net network for the detection task in view of the low signal-to-noise ratio characteristics of the SXI image. The present invention also proposes a two-stage training strategy for realizing unsupervised training, carefully constructs loss functions and training networks for different training stages, thereby gradually transferring the model from the source data set to the target data set, and realizing that the model can learn the data distribution of the target data set without true value annotation, and complete the detection task of the target data set. The present invention can not only effectively solve the problem of directly detecting different magnetosphere top tangent directions in SXI images with low signal-to-noise ratio, but also can use the existing true value data to directly fine-tune the model, and no longer requires true value labeling, thereby enhancing the practical application capability of the model in the future. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0012] Figure 1 is a schematic flow chart of an unsupervised detection method for the tangent direction of the magnetic top layer according to an embodiment of the present invention; Figure 2 is a schematic diagram of observation points selected on the satellite's May orbit and November orbit according to an embodiment of the present invention; Figure 3 is a two-dimensional integral image generated using the Jorgensen & Sun model and the PPMLR-MHD model according to an embodiment of the present invention; Figure 4According to an embodiment of the present invention Figure 3 The corresponding SXI simulation results of the 2D integral image in; Figure 5 is a schematic diagram of a residual attention U-Net network architecture according to an embodiment of the present invention; Figure 6 is a result of detecting the magnetosphere top tangent direction in the SXI image according to an embodiment of the present invention; Figure 7 is a structural block diagram of an unsupervised detection device for a magnetic top layer tangent direction according to an embodiment of the present invention; Figure 8 It is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0014] An embodiment of the present invention provides an unsupervised detection method for the tangent direction of the top magnetic layer, which achieves the effect of detecting the tangent direction of the top magnetic layer in multi-perspective SXI images without true value marking by using an improved U-Net network including a residual attention module and a two-stage training strategy.

[0015] According to an embodiment of the present invention, an embodiment of a method for unsupervised detection of the tangent direction of a magnetic top layer is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0016] In this embodiment, a method for unsupervised detection of the tangent direction of the magnetic top layer is provided. Figure 1 is a flow chart of an unsupervised detection method for the tangent direction of the magnetic top layer according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps: Step S101, using the Jorgensen & Sun model to generate source domain soft X-ray image data of the magnetic top layer of the target observation point.

[0017] In order to generate a simulated soft X-ray image of the magnetopause, a two-dimensional integral image of magnetospheric X-rays is first simulated. The Jorgensen & Sun model is used to simulate a boundary model that only includes the magnetopause and the bow shock, as well as an emission model of the region between these boundaries. The present invention performs a specific line of sight integration at a specific observation point to obtain a two-dimensional integral image, performs SXI simulation on these images, and generates the soft X-ray images required for the final detection, and the Jorgensen & Sun model data is used as the source domain soft X-ray image data.

[0018] Specifically, the Jorgensen & Sun model consists of a boundary model for the magnetopause and bow shock and an emission model for the region between the boundaries. The boundary model has the following form:

[0019] in, represents the angle with the X axis, represents the right-hand azimuth angle around the X axis from the Y axis. r0 represents the position of the subsolar point of the magnetopause. and Respectively represent the extent of the magnetotail opening in the xy plane and yz plane. The larger the value, the greater the opening of the magnetotail.

[0020] The X-ray emission model for the region between the boundaries consists of three regions separated by the magnetopause (MP) and the bow shock (BS), the solar wind region on the sunward (+X) side of the bow shock boundary, the magnetosheath between the bow shock and the magnetopause, and the magnetosphere inside the magnetopause. The model has the following form:

[0021] in, is a constant, , , , , are the variable parameter values ​​used to fit the radiation model.

[0022] The purpose of setting these variable parameters is to find the best match between the model and the PPMLR-MHD model. However, because the present invention uses the image generated by the Jorgensen & Sun model as the source domain data, and the source domain data only needs to provide the photon distribution characteristics in the tangent direction of the magnetosphere on the SXI image. There is no need for an absolute match between the two model data, so we only use these initial fitting parameters to generate images.

[0023] Using the above equations and initial parameters, a three-dimensional magnetopause model is generated at different solar wind densities. Then the specified viewing angle is integrated to obtain a two-dimensional image of the X-ray intensity, with an integration distance from the observation point to 80R E (where R E = 6371 km, which is the radius of the earth). The horizontal coordinates and vertical coordinates of the two-dimensional X-ray image in the present invention are respectively and express. and represents the angle in field of view coordinates, where The angle range is -8°~8°, The angular range is -13.5°~13.5°, and the accuracy is 0.1°.

[0024] Step S102, using the PPMLR-MHD model to generate target domain soft X-ray image data of the magnetic top layer of the target observation point.

[0025] The soft X-ray emission in the Earth's magnetic sheath and tip region under different solar wind conditions is simulated using PPMLR-MHD (global three-dimensional magnetosphere numerical model). Similar to step S102, the PPMLR-MHD model simulation data is used as the target domain soft X-ray image data.

[0026] Specifically, the PPMLR-MHD model uses an extended Lagrangian version to solve the MHD equations in the solution domain, which includes -300R E ≤X≤30R E and -150R E ≤y,z≤150R E , the grid spacing within the cube is uniformly maintained at least 0.1R E , and gradually decreases at the outside. The ionosphere is assumed to exhibit uniform Pedersen conductance (5S) and zero Hall conductance, as a simplified spherical shell. The present invention does not consider the effect of dipole tilt. The equation for the assumed three-dimensional X-ray volume emissivity is:

[0027] in, is the three-dimensional X-ray volume emissivity, represents the efficiency factor affected by various factors, including the size of the SWCX (Solar Wind Charge Exchange) cross section, the composition and abundance of heavy ions in the solar wind, and other related variables. The plasma number density and outer hydrogen are expressed as and express. and Among them, nsw represents the solar wind proton number density, usw represents the solar wind speed, and nth represents the particle thermal velocity. These three parameters are calculated using the PPMLR-MHD model.

[0028] The X-ray intensity along a particular line of sight is determined by the line integral of P using the following formula: .

[0029] In the integration process, other conditions are consistent with those in the Jorgensen model. In the present invention, 15 observation points in total are selected on the May orbit and the November orbit of the satellite. The specific observation points are selected as follows: Figure 2 As shown. Thus, we can obtain a two-dimensional X-ray integral diagram under different solar wind densities and different observation viewpoints. Figure 3 It shows that when the solar wind density is N=20 When the satellite is assumed to be at the apogee of the May orbit, the two-dimensional integral image is obtained using the Jorgensen & Sun model and the PPMLR-MHD model.

[0030] It should be noted that after obtaining the two-dimensional X-ray integral image, each image obtained above is regarded as a surface light source with uneven intensity distribution, and is simulated using the SXI simulator. The structural parameters set by the SXI simulator are the same as the SXI payload carried by the satellite. In this simulation, the SXI field of view is 16°×27°, and the simulation is performed every 0.5° in the horizontal and vertical directions. Each image consists of 271×161 point groups, so there are 271×161 intensity values ​​in the image. During the simulation, each intensity value can be input as a point light source, and the specific value of the number of rays emitted by each point light source can be calculated. The formula is as follows:

[0031] in, represents the intensity of the corresponding grid point, represents the area of ​​the grid points, T is the integration time, is the energy of each ray, is the solid angle covering the geometric light collecting area of ​​the point light source.

[0032] In the SXI simulation imaging process, two types of noise must be considered: instrument noise generated during SXI instrument simulation and cosmic background noise. Instrument noise is directly added during simulation, while cosmic background noise comes from diffuse X-rays from celestial bodies. Based on the diffuse background intensity shown in the X-ray band survey image of the Roentgen satellite, it can be estimated that the average soft X-ray background intensity within the SXI detection energy range (0.5~2.5keV) is 50 , which is also regarded as the cosmic background noise input in the present invention.

[0033] For example, the noisy two-dimensional X-ray integral images obtained using the Jorgensen & Sun model and the PPMLR-MHD model can be used as the input light source of SXI, and the final SXI image is a 55×33 photon matrix image. Figure 4 It is the photon map obtained after SXI simulation of the integral images produced by the above two models.

[0034] Step S103, determining the magnetosphere top tangent direction corresponding to the source domain soft X-ray image data as a true value label.

[0035] In supervised learning, the true value label is the correct output of the training data. In the present invention, the true value label specifically refers to the magnetosphere top tangent direction corresponding to each image in the source domain data set, indicating the true situation of the magnetosphere top tangent direction.

[0036] Step S104, the source domain soft X-ray image data and the target domain soft X-ray image data are respectively input into the residual attention U-Net network to obtain the source prediction result and the target prediction result.

[0037] The source domain soft X-ray image data is a soft X-ray image generated using the Jorgensen & Sun model, with corresponding true value labels. The target domain soft X-ray image data is generated using the PPMLR-MHD model, simulating soft X-ray images under real magnetospheric conditions, and these images do not have true value labels. The residual attention U-Net network is a deep learning network architecture designed for magnetopause detection tasks. When the residual attention U-Net network processes the source domain data and the target domain data, the outputs generated are called source prediction results and target prediction results.

[0038] Step S105, based on the error between the source prediction result and the true value label, and the error between the source prediction result and the target prediction result, the residual attention U-Net network is trained for the first round to obtain a preliminary network model.

[0039] The error between the source prediction result and the corresponding true value label is calculated, and in the absence of the true value label, the error is estimated by comparing the source prediction result with the target prediction result, which is usually quantified by a loss function. The first round of training is the first stage of model training. The purpose is to let the network learn the characteristics of the source domain data and try to reduce the error between the source prediction result and the true value label, while also taking into account the error between the source prediction result and the target prediction result, in order to prepare for subsequent domain adaptation.

[0040] Step S106, using the preliminary network model to process the target domain soft X-ray image data to obtain pseudo labels for the target domain soft X-ray image data.

[0041] The prediction results of the preliminary network model on the target domain data are called pseudo labels. The model trained with labeled data is used to predict the labels of unlabeled data. These predicted labels are called pseudo labels and can be used to guide the model to learn the characteristics of the target domain in the next stage.

[0042] Step S107, based on the target domain soft X-ray image data and pseudo labels, the preliminary network model is trained for two rounds to obtain the final network model.

[0043] The pseudo labels are used together with the target domain data for a second round of training. The second round of training usually involves fine-tuning the model parameters to improve the adaptability and accuracy of the model to the target domain data. Introduce or adjust the loss function, such as similarity loss or segmentation loss, to ensure that the model output is consistent with the pseudo labels while learning the specific features of the target domain data. The final network model obtained after the second round of training can more accurately detect the tangent direction of the collected simulated soft X-ray image data.

[0044] Step S108, using the final network model to detect the soft X-ray image data in the target domain.

[0045] The unsupervised detection method of the tangent direction of the top magnetic layer provided in this embodiment can process soft X-ray images from different observation angles. The method adopts an unsupervised learning strategy, allowing the model to be trained without a true value label. In view of the problem of low signal-to-noise ratio of SXI images, the method improves the detection accuracy of the tangent direction of the top magnetic layer in low signal-to-noise ratio images by designing a residual attention U-Net network. Through a two-stage unsupervised training strategy, the migration of the model from simulated data to real data is realized, which enhances the generalization ability of the model and the flexibility of practical application.

[0046] In some optional implementations, the residual attention U-Net network includes an encoder, a main classifier, and an auxiliary classifier; the main classifier and the auxiliary classifier share an encoder; the main classifier includes an encoding network and a complete decoding network; the auxiliary classifier includes a partial decoding network and a fully connected layer; The encoder includes a residual attention module, which consists of a channel attention module, a spatial attention module, and a residual module connected in sequence; The channel attention module is used to perform global average pooling and global maximum pooling operations on the input feature map; the spatial attention module includes a 1×1 convolution layer and an activation layer containing a sigmoid function; the residual connection layer includes a 1×1 convolution layer and a batch normalization layer.

[0047] In order to better adapt to the magnetopause detection task, the residual attention U-Net network proposed in this paper consists of a shared encoder, a main classifier and an auxiliary classifier, such as Figure 5As shown in the figure, the dual convolution module in the encoder includes two 3×3 convolutional layers, followed by batch normalization layers and ReLU activation layers, downsampling using the maximum pooling layer, and upsampling using the transposed convolutional layer in the decoding network, thereby changing the number of image channels during training. The final main classifier and auxiliary classifier share the same encoder structure. The auxiliary classifier outputs low-level features through the fully connected layer after passing through the partial decoding network, while the main classifier goes through the complete encoding and decoding process to obtain the final segmentation result. A Dropout layer is added to the decoding stage of each classifier, thereby introducing the required differences in the outputs of the two classifiers. At the same time, during the training process, in order to emphasize the model's attention to the details of the magnetosphere, a residual attention module is added to each layer of the shared encoder. The module mainly consists of three parts, namely the channel attention module, the spatial attention module, and the residual module.

[0048] Specifically, the channel attention module first performs global average pooling and global maximum pooling operations on the input feature map to obtain the global average pooling feature vector and the global maximum pooling feature vector, respectively. These feature vectors provide different feature descriptions and capture the importance of each channel in the original feature map. Subsequently, the feature vectors of these two operations are processed through their respective networks, each of which consists of a multi-layer perceptron (MLP) and a hidden layer. After traversing these networks, the output of the two branch results and the original feature map after passing through the sigmoid function is merged using element-wise multiplication. This operation amplifies the positions of channels with high attention weights in both branches, while weakening the positions with high attention weights in only one branch or low attention weights in both branches.

[0049] The spatial attention module consists of two components: a 1×1 convolutional layer and an activation layer containing a sigmoid function. The results obtained from the channel attention module are calculated through these layers. This output is then multiplied with the initial channel attention result to generate a channel-spatial attention map. The resulting attention feature map dynamically learns the importance of each channel and space. This adaptive learning mechanism improves the effectiveness of the model in the task of detecting the tangent direction of the magnetopause.

[0050] A residual module is integrated in each encoder layer, which includes a 1×1 convolution layer and a batch normalization layer. This module enables the network to directly obtain information from low-level features and transfer it to high-level feature representations. In this way, the network can better understand and utilize low-level details and local features.

[0051] The method provided in this embodiment designs a residual attention module. This module runs the channel attention module and the spatial attention module in parallel and adds a residual connection to improve the representation of the network. By adding the residual attention module to U-Net, the detection effect of target details is improved.

[0052] The above step S102 inputs the source domain soft X-ray image data and the target domain soft X-ray image data into the residual attention U-Net network respectively to obtain the source prediction result and the target prediction result, including the following steps: Step S1021, processing the source domain soft X-ray image data through the residual attention U-Net network to obtain a first source prediction result output by the main classifier and a second source prediction result output by the auxiliary classifier.

[0053] The source domain soft X-ray image data is input into the residual attention U-Net network. The network contains a main classifier and an auxiliary classifier with the same encoder structure. The main classifier is responsible for generating the first source prediction result, and the auxiliary classifier outputs the second source prediction result.

[0054] Step S1022, the target domain soft X-ray image data is processed through the residual attention U-Net network to obtain the first target prediction result output by the main classifier and the second target prediction result output by the auxiliary classifier.

[0055] Similarly, the target domain soft X-ray image data is also processed by the residual attention U-Net network.

[0056] The method provided in this embodiment obtains predictions from both source domain and target domain data through the residual attention U-Net network. By comparing and analyzing these prediction results, the model can be further trained and optimized.

[0057] In some optional implementations, the above step S103 performs a first round of training on the residual attention U-Net network based on the error between the source prediction result and the true value label, and the error between the source prediction result and the target prediction result to obtain a preliminary network model, including the following steps: Based on the following loss function, calculate the first loss function value , to update the parameters of the residual attention U-Net network:

[0058]

[0059]

[0060]

[0061]

[0062] in, The segmentation loss obtained by inputting the segmentation loss function into the first source prediction result; The segmentation loss obtained by inputting the segmentation loss function into the second source prediction result; Indicated by and Calculate the segmentation loss for the first round of training; is the main prediction adversarial loss, which characterizes the difference between the first source prediction result and the first target prediction result in the discriminator Dp; To assist in predicting the adversarial loss, characterize the difference between the second source prediction result and the second target prediction result in the discriminator Da; for and Adversarial loss after weight calculation; is a constant; among them, The prediction result for the first source, For the second source prediction result, Predict the results for the first goal, is the second target prediction result; Dp represents the discriminator used for primary prediction; Da represents the discriminator used for auxiliary prediction.

[0063] Among them, the segmentation loss function is as follows:

[0064] in, is one of the segmentation loss obtained by inputting the segmentation loss function of the first source prediction result and the segmentation loss obtained by inputting the segmentation loss function of the second source prediction result, is the binary cross entropy loss function; is the Dice loss function. To express and The coefficient of the weight ratio between them.

[0065] Specifically, is a real number between 0 and 1, used to represent and The relative importance of the total loss. When the total loss is completely determined by constitute, The influence of is 0; when When the total loss is completely determined by constitute, The impact is 0.

[0066] The dataset generated by the Jorgensen&Sun model is used as the source domain dataset, and the dataset generated by the PPMLR-MHD model is used as the target domain dataset. The images in the source domain are , the true value label is , the image in the target domain is ,but , . Where U and V represent the number of SXI images in the source and target domains, respectively. Each image in The corresponding true value labels are annotated in . Given an image in the source domain And the true value label corresponding to the image , an unlabeled true value image in the target domain , the goal is to learn a mapping function that transforms the input data Mapped to a unique segmentation result map .

[0067] The source domain knowledge is learned using the Jorgensen & Sun model data with true labels. Binary cross entropy loss and Dice loss are used as segmentation losses, and the formula is as follows:

[0068] Where N represents the total number of pixels, Indicates the label, Indicates the prediction result.

[0069]

[0070] Among them, X represents the pixel set of the prediction result, and Y represents the pixel set of the true value.

[0071] The Dice loss function is based on the Dice coefficient and is used to evaluate the overlap of two binary images or binary masks. The Dice loss quantifies the similarity between the predicted result and the true value, while the BCE loss evaluates the difference between the predicted target area and the background area. By combining the Dice loss and the BCE loss, the internal similarity and external dissimilarity of the predicted area can be optimized. Therefore, the segmentation loss adopts this combined loss function. Set to 0.5. The segmentation loss at this stage can be expressed as:

[0072] Then, the main classifier and the auxiliary classifier generate segmentation losses respectively, which are defined as and In this article, we set , the final segmentation loss is as follows: .

[0073] However, segmentation loss alone cannot achieve domain adaptation. Segmentation loss focuses on learning the photon distribution characteristics of the source dataset, with the goal of minimizing the gap between the predicted results and the true value labels of the training dataset. However, it cannot capture the difference between the two datasets, and is therefore insufficient to transfer the data knowledge from the Jorgensen & Sun model to the MHD model data knowledge and narrow the gap between the target domain and the source domain. Therefore, adversarial loss is added. and , to minimize the difference between the source and target domains.

[0074]

[0075]

[0076] in, and Represent two different discriminators for main prediction and auxiliary prediction, respectively.

[0077] The discriminator aims to determine the target domain prediction Prediction with source domain The degree of closeness in the semantic space. By optimizing the adversarial loss, the model is forced to minimize the difference between domains at the semantic level. The present invention uses a lower weight to add the two adversarial losses after calculation, thereby obtaining the final adversarial loss: .

[0078] The method provided in this embodiment performs supervised learning on the source domain data, and simultaneously utilizes the outputs of the two classifiers in the network to minimize the difference between the target domain and the source domain, thereby promoting the gradual transfer of knowledge from the source domain to the target domain, thereby improving the overall performance.

[0079] In some optional implementations, the above step S105 performs two rounds of training on the preliminary network model based on the target domain soft X-ray image data and pseudo labels, including the following steps: Step S1051, input the target domain soft X-ray image data into the preliminary network model for processing, and obtain the third target prediction result output by the main classifier and the fourth target prediction result output by the auxiliary classifier.

[0080] Step S1052, based on the error between the third target prediction result and the fourth target prediction result, the error between the third target prediction result and the pseudo label, and the error between the fourth target prediction result and the pseudo label, update the parameters of the preliminary network model.

[0081] This step updates the parameters of the preliminary network model based on the error between the predictions and the pseudo-labels. The errors are calculated by comparing the outputs of the primary and auxiliary classifiers (i.e., the third and fourth target predictions) with the pseudo-labels of the target domain data. These errors are used to calculate the loss function, which may include a combination of multiple losses, such as segmentation loss and similarity loss. By minimizing this loss function, the parameters of the network are adjusted to reduce the difference between the predictions and the pseudo-labels and improve the model's adaptability to the target domain data.

[0082] In some optional implementations, the above step S1052 updates the parameters of the preliminary network model based on the error between the third target prediction result and the fourth target prediction result, the error between the third target prediction result and the pseudo label, and the error between the fourth target prediction result and the pseudo label, including: Based on the following loss function, calculate the second loss function value , to update the parameters of the preliminary network model:

[0083]

[0084]

[0085] in, The segmentation loss obtained by inputting the segmentation loss function into the third target prediction result; The segmentation loss obtained by inputting the segmentation loss function into the fourth target prediction result; the similarity loss calculated for the third target prediction and the fourth target prediction; Indicated by and Calculated segmentation loss for two rounds of training; is a constant; predicting outcomes for the third objective; Predict the results for the fourth goal.

[0086] Specifically, after the first round of training, the preliminary network model is used to generate pseudo labels for the target domain data. Then, in the second round of training, the model is fine-tuned using only the target domain data to complete the detection of the tangent direction of the magnetosphere top on the target domain data. During the entire training phase of the target domain, a certain target domain image is input, and the network should produce a unique correct result. Therefore, while introducing the segmentation loss, the similarity loss is also introduced. This loss utilizes the difference between the output results of the main classifier and the auxiliary classifier, forcing the model to keep the output consistent.

[0087] KL loss (KLdivergence loss) is used as the similarity loss to limit the difference between the outputs of the two classifiers:

[0088] Similarly, in the second round of training, training the model on pseudo labels will also incur segmentation loss and The definition of segmentation loss is consistent with the first stage, and the final segmentation loss of the second round of training is:

[0089] Combine the above losses. For the similarity loss, set the weight of the regularization loss The total loss for the first and second rounds of training is as follows:

[0090] .

[0091] The method provided in this embodiment first generates pseudo labels for the target domain data using the model trained in the previous round. The initial parameters of the segmentation network are directly obtained from the model parameters of the first round of training, and then the model is further improved in the target domain. By using pseudo labels, the model can focus on learning specific knowledge in the target domain. Considering the inherent noise in pseudo labels, the difference between the output of the main classifier and the auxiliary classifier is used to effectively prevent the model from overfitting the noise present in the pseudo labels.

[0092] In some optional implementations, the SXI image obtained by the method proposed in the present invention using the PPMLR-MHD model is tested to test the detection effect of the magnetopause tangent direction of the method without true value annotation.

[0093] The source dataset includes SXI images with cosmic background noise generated by the Jorgensen & Sun model and the corresponding ground truth. In contrast, the target domain dataset used to simulate the real magnetospheric conditions only contains SXI images with cosmic background noise generated by the PPMLR-MHD model. These datasets all contain 15 satellite observation viewpoints, with solar wind density N ranging from 5cm -3 、12cm -3 , 20cm -3 、25cm -3 Up to 35cm -3. Due to the uncertainty of the magnetospheric environment, each 2D X-ray integral image was subjected to 100 independent SXI simulations. Both the source and target datasets contain 7,500 images, and the datasets are randomly divided into training, validation, and test sets in a ratio of 8:1:1. The input end consists of the original photon matrix to ensure detection accuracy, and all input photon matrices can be visualized. The specific process of training includes: Step 1: Initialize the residual attention U-Net network model.

[0094] Step 2: Input both the source and target datasets into the network model. The training uses the stochastic gradient descent (SGD) optimizer with a batch size of 64 and an initial learning rate of 0.0001 to train the segmentation model. The discriminator uses the Adam optimizer with an initial learning rate of 0.0001. Both the segmentation network and the adversarial network use a dynamically changing learning rate decay strategy, and then multiply the learning rate by a coefficient ,in, is the current iteration number, is the total number of iterations. In the first round of training, the total number of iterations was set to 2000. First, the first 60% of the iterations were trained without adversarial loss to reduce the initial prediction noise and better learn source domain specific knowledge, and then adversarial training was introduced. In the second round of training, similarity loss was added and an additional 500 iterations were performed to further fine-tune the model. The main software environment uses python = 3.8.18, pytorch = 1.11.0, torchvision = 0.13.1. When training the model, the training iteration round with the minimum total loss in each stage is recorded. After the training is completed, the network model with the minimum total loss obtained from the second round of training is saved as a pth file.

[0095] Step 3: Create a network model, load the trained pth file, and read the pre-trained weights. Input the target domain test data set into the network model, obtain the magnetosphere top tangent direction results in each SXI image, and save it as a binary image. The obtained detection result image is as follows Figure 6 As shown, from top to bottom are the MHD two-dimensional X-ray integral image, the corresponding SXI simulation image, and the final magnetospheric top tangent direction detection result.

[0096] In this embodiment, a device for unsupervised detection of the tangent direction of the magnetic top layer is also provided, which is used to implement the above-mentioned embodiments and preferred implementation modes, and the descriptions that have been made will not be repeated. As used below, the term "module" can implement a combination of software and / or hardware of a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0097] This embodiment provides an unsupervised detection device for the tangent direction of the magnetic top layer, such as Figure 7 As shown, including: A source data acquisition module 701 is used to generate source domain soft X-ray image data of the magnetic top layer of the target observation point using the Jorgensen & Sun model; The target data acquisition module 702 is used to generate the target domain soft X-ray image data of the magnetic top layer of the target observation point by using the PPMLR-MHD model; A true value label determination module 703 is used to determine the magnetosphere top tangent direction corresponding to the source domain soft X-ray image data as a true value label; A result prediction module 704 is used to input the source domain soft X-ray image data and the target domain soft X-ray image data into the residual attention U-Net network to obtain a source prediction result and a target prediction result; A first round training module 705 is used to perform a first round of training on the residual attention U-Net network based on the error between the source prediction result and the true value label, and the error between the source prediction result and the target prediction result to obtain a preliminary network model; A pseudo label acquisition module 706 is used to process the target domain soft X-ray image data using the preliminary network model to obtain a pseudo label for the target domain soft X-ray image data; A second round training module 707, for performing a second round of training on the preliminary network model based on the target domain soft X-ray image data and pseudo labels to obtain a final network model; The image detection module 708 is used to detect the soft X-ray image data in the target domain using the final network model.

[0098] In some optional implementations, the residual attention U-Net network includes an encoder, a main classifier, and an auxiliary classifier; the main classifier and the auxiliary classifier share an encoder; the main classifier includes an encoding network and a complete decoding network; the auxiliary classifier includes a partial decoding network and a fully connected layer; The encoder includes a residual attention module, which consists of a channel attention module, a spatial attention module, and a residual module connected in sequence; The channel attention module is used to perform global average pooling and global maximum pooling operations on the input feature map; the spatial attention module includes a 1×1 convolution layer and an activation layer containing a sigmoid function; the residual connection layer includes a 1×1 convolution layer and a batch normalization layer.

[0099] The above-mentioned result prediction module 704 includes: The first result prediction unit is used to process the source domain soft X-ray image data through the residual attention U-Net network to obtain the first source prediction result output by the main classifier and the second source prediction result output by the auxiliary classifier.

[0100] The second result prediction unit is used to process the target domain soft X-ray image data through the residual attention U-Net network to obtain the first target prediction result output by the main classifier and the second target prediction result output by the auxiliary classifier.

[0101] In some optional implementations, the above-mentioned one-round training module 705 includes: The first loss calculation unit is used to calculate the first loss function value based on the following loss function: , to update the parameters of the residual attention U-Net network:

[0102]

[0103]

[0104]

[0105]

[0106] in, The segmentation loss obtained by inputting the segmentation loss function into the first source prediction result; The segmentation loss obtained by inputting the segmentation loss function into the second source prediction result; Indicated by and Calculate the segmentation loss for the first round of training; is the main prediction adversarial loss, which characterizes the difference between the first source prediction result and the first target prediction result in the discriminator Dp; To assist in predicting the adversarial loss, characterize the difference between the second source prediction result and the second target prediction result in the discriminator Da; for and Adversarial loss after weight calculation; is a constant; among them, The prediction result for the first source, For the second source prediction result, Predict the results for the first goal, is the second target prediction result; Dp represents the discriminator used for primary prediction; Da represents the discriminator used for auxiliary prediction.

[0107] Among them, the segmentation loss function is as follows:

[0108] in, is one of the segmentation loss obtained by inputting the segmentation loss function of the first source prediction result and the segmentation loss obtained by inputting the segmentation loss function of the second source prediction result, is the binary cross entropy loss function; is the Dice loss function. To express and The coefficient of the weight ratio between them.

[0109] In some optional implementations, the two-round training module 707 includes: A result acquisition unit, inputting the target domain soft X-ray image data into the preliminary network model for processing, and obtaining the third target prediction result output by the main classifier and the fourth target prediction result output by the auxiliary classifier; A parameter updating unit is used to update the parameters of the preliminary network model based on the error between the third target prediction result and the fourth target prediction result, the error between the third target prediction result and the pseudo label, and the error between the fourth target prediction result and the pseudo label.

[0110] In some optional implementations, the parameter updating unit includes: The second loss calculation unit is used to calculate the second loss function value based on the following loss function: , to update the parameters of the preliminary network model:

[0111]

[0112]

[0113] in, The segmentation loss obtained by inputting the segmentation loss function into the third target prediction result; The segmentation loss obtained by inputting the segmentation loss function into the fourth target prediction result; the similarity loss calculated for the third target prediction and the fourth target prediction; Indicated by and Calculated segmentation loss for two rounds of training; is a constant; predicting outcomes for the third objective; Predict the results for the fourth goal.

[0114] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0115] The unsupervised detection device for the tangent direction of the magnetic top layer in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0116] The embodiment of the present invention also provides a computer device having the above Figure 7 Unsupervised detection device of the tangent direction of the magnetic top layer shown.

[0117] See also Figure 8 , Figure 8 is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention, such as Figure 8 As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 8 A processor 10 is taken as an example.

[0118] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0119] The memory 20 stores instructions executable by at least one processor 10, so that at least one processor 10 executes the method shown in the above embodiment.

[0120] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0121] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0122] The computer device also includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means. Figure 8 The example of connecting through bus is taken in the following.

[0123] The input device 30 can receive input digital or character information, and generate key signal input related to the user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, an indicator bar, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 may include a display device, an auxiliary lighting device (e.g., an LED) and a tactile feedback device (e.g., a vibration motor), etc. The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display and a plasma display. In some optional embodiments, the display device may be a touch screen.

[0124] The embodiment of the present invention also provides a computer-readable storage medium. The method according to the embodiment of the present invention can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0125] A part of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the existence of the computer program instruction in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc., and accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium accessible to the computer.

[0126] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. An unsupervised detection method for the tangent direction of the magnetic top layer, characterized in that: The method comprises: The Jorgensen & Sun model is used to generate source domain soft X-ray image data of the magnetic top layer at the target observation point; Generate target domain soft X-ray image data of the magnetic top layer of the target observation point by using the PPMLR-MHD model; Determine the magnetosphere top tangent direction corresponding to the source domain soft X-ray image data as a true value label; Inputting the source domain soft X-ray image data and the target domain soft X-ray image data into a residual attention U-Net network respectively to obtain a source prediction result and a target prediction result; Based on the error between the source prediction result and the true value label, and the error between the source prediction result and the target prediction result, the residual attention U-Net network is trained for the first round to obtain a preliminary network model; Processing the target domain soft X-ray image data using the preliminary network model to obtain pseudo labels for the target domain soft X-ray image data; Based on the target domain soft X-ray image data and pseudo labels, the preliminary network model is trained for two rounds to obtain a final network model; The final network model is used to detect the target domain soft X-ray image data.

2. The method according to claim 1, characterized in that: The residual attention U-Net network includes an encoder, a main classifier and an auxiliary classifier; the main classifier and the auxiliary classifier share the encoder; the main classifier includes an encoding network and a complete decoding network; the auxiliary classifier includes a partial decoding network and a fully connected layer; The step of inputting the source domain soft X-ray image data and the target domain soft X-ray image data into a residual attention U-Net network respectively to obtain a source prediction result and a target prediction result includes: Processing the source domain soft X-ray image data through the residual attention U-Net network to obtain a first source prediction result output by the main classifier and a second source prediction result output by the auxiliary classifier; The target domain soft X-ray image data is processed through the residual attention U-Net network to obtain a first target prediction result output by the main classifier and a second target prediction result output by the auxiliary classifier.

3. The method according to claim 2, characterized in that Based on the error between the source prediction result and the true value label, and the error between the source prediction result and the target prediction result, the residual attention U-Net network is trained for the first round to obtain a preliminary network model, including: Based on the following loss function, calculate the first loss function value , to update the parameters of the residual attention U-Net network: in, The segmentation loss obtained by inputting the segmentation loss function into the first source prediction result; The segmentation loss obtained by inputting the segmentation loss function into the second source prediction result; Indicated by and Calculate the segmentation loss for the first round of training; is the main prediction adversarial loss, which characterizes the difference between the first source prediction result and the first target prediction result in the discriminator Dp; To assist in predicting the adversarial loss, characterize the difference between the second source prediction result and the second target prediction result in the discriminator Da; for and Adversarial loss after weight calculation; is a constant; among them, The prediction result for the first source, For the second source prediction result, Predict the results for the first goal, is the second target prediction result; Dp represents the discriminator used for primary prediction; Da represents the discriminator used for auxiliary prediction.

4. The method according to claim 3, characterized in that The segmentation loss function is as follows: in, is one of the segmentation loss obtained by inputting the segmentation loss function of the first source prediction result and the segmentation loss obtained by inputting the segmentation loss function of the second source prediction result, is the binary cross entropy loss function; is the Dice loss function; To express the and stated The coefficient of the weight ratio between them.

5. The method according to claim 2, characterized in that: The step of performing two rounds of training on the preliminary network model based on the target domain soft X-ray image data and the pseudo labels includes: Inputting the target domain soft X-ray image data into the preliminary network model for processing, obtaining a third target prediction result output by the main classifier and a fourth target prediction result output by the auxiliary classifier; Based on the error between the third target prediction result and the fourth target prediction result, the error between the third target prediction result and the pseudo label, and the error between the fourth target prediction result and the pseudo label, the parameters of the preliminary network model are updated.

6. The method according to claim 5, characterized in that The updating of parameters of the preliminary network model based on the error between the third target prediction result and the fourth target prediction result, the error between the third target prediction result and the pseudo label, and the error between the fourth target prediction result and the pseudo label comprises: Based on the following loss function, calculate the second loss function value , to update the parameters of the preliminary network model: in, The segmentation loss obtained by inputting the segmentation loss function into the third target prediction result; The segmentation loss obtained by inputting the segmentation loss function into the fourth target prediction result; the similarity loss calculated for the third target prediction and the fourth target prediction; Indicated by and Calculated segmentation loss for the second round of training; is a constant; predicting outcomes for the third objective; Predict the results for the fourth goal.

7. The method according to any one of claims 2 to 5, characterized in that: The encoder includes a residual attention module, which is composed of a channel attention module, a spatial attention module and a residual module connected in sequence; The channel attention module is used to perform global average pooling and global maximum pooling operations on the input feature map; the spatial attention module includes a 1×1 convolution layer and an activation layer containing a sigmoid function; the residual connection layer includes a 1×1 convolution layer and a batch normalization layer.

8. An unsupervised detection device for the tangent direction of the magnetic top layer, characterized in that: The device comprises: Source data acquisition module, used to generate source domain soft X-ray image data of the magnetic top layer of the target observation point using the Jorgensen & Sun model; A target data acquisition module, used for generating target domain soft X-ray image data of the magnetic top layer of the target observation point by using the PPMLR-MHD model; A true value label determination module, used to determine the magnetosphere top tangent direction corresponding to the source domain soft X-ray image data as a true value label; A result prediction module, used to input the source domain soft X-ray image data and the target domain soft X-ray image data into a residual attention U-Net network respectively to obtain a source prediction result and a target prediction result; A first round training module, used to perform a first round of training on the residual attention U-Net network based on the error between the source prediction result and the true value label, and the error between the source prediction result and the target prediction result, to obtain a preliminary network model; a pseudo label acquisition module, configured to process the target domain soft X-ray image data using the preliminary network model to obtain a pseudo label for the target domain soft X-ray image data; A two-round training module, used for performing two rounds of training on the preliminary network model based on the target domain soft X-ray image data and pseudo labels to obtain a final network model; An image detection module is used to detect the soft X-ray image data in the target domain using the final network model.

9. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method according to any one of claims 1 to 7 by executing the computer instructions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 7.

Images