Navigation jamming identification method based on spaceborne GNSS-R data, storage medium and equipment

Abstract

The application discloses a navigation interference identification method based on spaceborne GNSS-R data, a storage medium and equipment, first, spaceborne GNSS-R data is acquired, a time delay Doppler diagram of a channel is generated, and a training set, a verification set and a test set of semi-supervised learning are proportionally divided; twice different random enhancements are applied to each DDM image, two views are generated, and the two views are input into a neural network model, a contrast loss function is designed, contrast learning is performed on the two pretreated views, pre-training weights are loaded, a training process is performed, model parameters are optimized, and a model with minimum contrast loss is saved; a binary classification network head is designed as a classification output, training and verification are performed through DDM image data with RFI labels, an early stop mechanism is introduced, test set data is input into a trained RFI identification model for RFI identification, and development of spaceborne GNSS-R data in the navigation interference identification direction is promoted.

Description

Technical Field

[0001] This invention belongs to the technical field of navigation interference source identification and localization, and mainly relates to a navigation interference identification method, storage medium and device based on spaceborne GNSS-R data. Background Technology

[0002] With the rapid development of electromagnetic technology, radio frequency interference (RFI) has become increasingly prominent, especially in fields such as radar, satellite communications, and radio astronomy. RFI can severely affect signal quality and lead to a decline in system performance. RFI identification for Global Navigation Satellite Systems (GNSS) can utilize GNSS / Low Earth Orbit navigation observation data, ephemeris data, ADS-B, and AIS data, and deploy dedicated interference monitoring equipment for detection (Patent CN109359523B). Navigation interference identification based on GNSS-R data has significant application value due to its wide coverage, rapid data updates, and multi-satellite networking capabilities.

[0003] GNSS-R is a remote sensing method that uses reflected signals from navigation systems such as GNSS to retrieve physical parameters of the ocean, land surface, and vegetation. Spaceborne GNSS-R can acquire time-delay Doppler images (DDM) for calculating geophysical parameters such as soil moisture and sea surface wind speed. DDM data exhibits different signal characteristics when affected by radiofrequency interference (RFI), effectively characterizing navigation interference. However, DDM data is characterized by sparsity and irregularity, posing a significant challenge to RFI interference identification. Currently, the assessment of RFI in spaceborne GNSS-R data mainly relies on parameters such as the kurtosis of the DDM data's noise floor power, which has poor accuracy and weak robustness, limiting the accuracy of geophysical parameter retrieval.

[0004] Semi-supervised learning, as a learning method between supervised and unsupervised learning, can learn using a small amount of labeled data and a large amount of unlabeled data, making it particularly suitable for applications in GNSS-R where RFI annotation data is scarce. However, RFI identification based on time-delay Doppler image data using semi-supervised learning still faces many challenges. First, RFIs are diverse and complex, making them difficult to describe with simple mathematical models. Second, in practical applications, RFIs are often superimposed on the target signal, making them difficult to separate. Furthermore, the performance of semi-supervised learning algorithms is highly dependent on data quality and feature extraction methods; therefore, designing effective feature extraction methods and semi-supervised learning algorithms is crucial for achieving effective RFI identification. Summary of the Invention

[0005] This invention addresses the problems existing in the prior art by proposing a navigation interference identification method, storage medium, and device based on spaceborne GNSS-R data. First, spaceborne GNSS-R data is acquired, and channel delay Doppler maps are generated. These maps are then proportionally divided into training, validation, and test sets for semi-supervised learning. Two different random augmentations are applied to each DDM image to generate two views, which are input into the RFI identification model. A contrastive loss function is designed, and the two pre-processed views are compared and learned. Pre-trained weights are loaded, the training process is executed, model parameters are optimized, and the model with the minimum contrastive loss is saved. A binary classification network head is designed as the classification output. Training and validation are performed using RFI-labeled DDM image data, and an early stopping mechanism is introduced. Test set data is input into the trained RFI identification model for RFI identification. This invention combines semi-supervised learning to construct a DDM feature extraction model, further improving the ability to achieve high-precision RFI identification globally using spaceborne GNSS-R technology and promoting the development of machine learning in spaceborne GNSS-R interference identification.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a navigation interference identification method based on spaceborne GNSS-R data, comprising the following steps:

[0007] S1: Acquire spaceborne GNSS-R reflection observation data, generate channel delay Doppler maps based on the data, and divide the training set, validation set, and test set for semi-supervised learning according to the proportions; the channels include, but are not limited to, GPS, BeiDou, Galileo, and GLONASS satellite navigation systems, and the spaceborne GNSS-R reflection observation data includes at least DDM waveform data;

[0008] S2: Perform two different random enhancements on each DDM image to generate two data-enhanced views; the data enhancement methods include at least random color jitter, random grayscale conversion, and random horizontal flipping, wherein random color jitter is divided into random brightness, random contrast, random saturation, and random hue adjustment;

[0009] S3: Input the two data-enhanced views generated in step S2 into the RFI recognition model, set the contrast loss function for contrast learning training; load the pre-trained model weights, execute the training process, calculate the loss of the two image pairs after forward propagation, optimize the model parameters through backpropagation, and obtain the model with the minimum contrast loss.

[0010] S4: Using the binary classification network head as the classification output, load the model weights with the minimum contrast loss obtained in step S3, train and validate using DDM image data with RFI labels, and introduce an early stopping mechanism; input the test set data into the trained RFI recognition model for RFI recognition.

[0011] As an improvement of the present invention, in step S1, the spaceborne GNSS-R reflection observation data includes DDM waveform data. The dataset is randomly divided into an unsupervised learning dataset and a supervised learning dataset according to a ratio of 80% and 20%. The supervised learning dataset is further divided into a training set, a validation set, and a test set according to a ratio of 80%, 10%, and 10%.

[0012] As an improvement to the present invention, step S2, which enhances the random brightness in the random color jitter, specifically involves:

[0013]

[0014] in For random brightness factor, For the image in Position number The pixel values ​​of each channel;

[0015] The calculation process for random contrast enhancement is as follows:

[0016]

[0017] in For random contrast factor, This represents the average grayscale value of the image.

[0018] As another improvement of the present invention, step S2, the random saturation enhancement in random color jitter, specifically involves:

[0019] Convert the image to grayscale. :

[0020]

[0021] in , and The images are respectively in The pixel values ​​of the R, G, and B channels at the location.

[0022] Blending of RGB channels:

[0023]

[0024] in This is a random saturation factor.

[0025] As another improvement of the present invention, step S2, the random hue adjustment in random color jitter, specifically includes:

[0026] Convert the image from RGB space to HSV space, and then perform an addition operation on the hue channel H:

[0027]

[0028] Then convert the image back to RGB space;

[0029] The calculation process for random grayscale conversion is as follows:

[0030]

[0031]

[0032] in The probability of grayscale conversion. These are the pixel values ​​of the original image. These are the pixel values ​​of the image after grayscale processing.

[0033] As another improvement of the present invention, the RFI identification model in step S3 includes, but is not limited to, models with SwinTransformer, MobilenetV2, and ResNet as the backbone network, and the NTXentLoss function as the contrastive loss function, specifically:

[0034]

[0035] in, , These are the feature embedding vectors of two augmented views of the same DDM image. For cosine similarity, This is the temperature coefficient.

[0036] As a further improvement of the present invention, step S4 specifically includes the following steps:

[0037] S41: Use BCEWithLogitsLoss and dynamically set the weight coefficient of positive samples. Calculate the loss based on the ratio of positive to negative samples in the training data to alleviate the class imbalance problem. Use the AdamW optimizer and set the learning rate to 0.00005. Use OneCycleLR scheduling to dynamically adjust the learning rate to balance training speed and stability.

[0038] S42: Use Autocast+GradScaler for mixed-precision training, dynamically manage the numerical precision of different operators, convert some operators to FP16 (half-precision floating-point numbers) and keep the rest FP32 (single-precision floating-point numbers) to accelerate computation; GradScaler ensures the effectiveness of backpropagation by dynamically scaling gradients.

[0039] S43: Introduce an early stop mechanism, with the main metrics being Loss and Accuracy. It tolerates no improvement for 8 rounds and generates real-time trend charts of Loss and Accuracy for visual diagnosis to avoid overfitting.

[0040] S44: After each round of training, the model is evaluated on the validation set, the validation loss and accuracy are calculated, and the fine-tuned RFI recognition model weights are saved.

[0041] S45: Input the test set divided in step S1 into the RFI recognition model and calculate relevant indicators to evaluate the model performance.

[0042] As a further improvement of the present invention, the calculation process of the BCEWithLogitsLoss function in step S41 is as follows:

[0043]

[0044] in The weights output by the model. For the Sigmoid function, This is a real label;

[0045] In step S43, the Accuracy calculation process is as follows:

[0046] .

[0047] To achieve the above objectives, the present invention also adopts the following technical solution: a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a navigation interference identification method based on spaceborne GNSS-R data.

[0048] To achieve the above objectives, the present invention also adopts the following technical solution: a computer device, comprising:

[0049] Memory, used to store instructions;

[0050] A processor is configured to execute the instructions, causing the computer device to perform operations of a navigation interference identification method based on spaceborne GNSS-R data.

[0051] Compared with existing technologies, the present invention has the following advantages: The present invention discloses a navigation interference identification method, storage medium, and device based on spaceborne GNSS-R data. It uses a learning method that combines unsupervised and supervised learning, and utilizes the time-delay Doppler map in GNSS-R data for RFI identification. The difference from traditional RFI identification methods is that it uses complete time-delay Doppler map information combined with a lightweight and high-precision network model, the Swin Transformer, which enhances the generalization ability and robustness of the RFI identification method. In the application of spaceborne GNSS-R data, this method can effectively identify RFI received by GNSS signals. Attached Figure Description

[0052] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0053] Figure 2 These are time-delayed Doppler images under different scenarios, where

[0054] Figure 2 (a) is a classic time-delayed Doppler plot affected by RFI;

[0055] Figure 2 (b) is a time-delayed Doppler image of classical sea surface reflection;

[0056] Figure 2 (c) is a time-delayed Doppler image of classic ice surface reflection;

[0057] Figure 2 (d) is a time-delayed Doppler image suspected of being damaged by RFI;

[0058] Figure 3 This is a neural network structure diagram of swin_base_patch4_window7_224 in Embodiment 1 of the present invention. Detailed Implementation

[0059] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0060] Example 1

[0061] Navigation interference identification based on spaceborne GNSS-R data is a research direction with significant research importance and application value. Through in-depth research on feature extraction methods and semi-supervised learning algorithms for RFI interference signals, it is hoped that high-accuracy and robust navigation interference identification can be achieved, providing assurance for global navigation safety. Figure 1 As shown, a navigation interference identification method based on spaceborne GNSS-R data is described, which includes the following steps:

[0062] Step S1: Acquire onboard GNSS-R data, use the data to generate a time delay Doppler map of a certain channel (such as GPS, BeiDou, Galileo and GLONASS satellite navigation systems), and then divide it into training set, validation set and test set for semi-supervised learning according to a certain ratio;

[0063] The spaceborne GNSS-R reflection observation data includes DDM waveform data (Ddm_raw_data); of which, 80% of the data is used for unsupervised contrastive learning, 80% of the remaining data is used for supervised training, 10% is used for validation, and the remaining 10% is used for testing.

[0064] For example, one can download a day's L1 level reflectance data by visiting the official website of the Tianmu-1 remote sensing meteorological observation constellation, and obtain the time-delayed Doppler image of the C1 channel from 0:00 to 24:00. This embodiment selects time-delayed Doppler images under different scenarios, specifically as follows: Figure 2 As shown, Figure 2 (a) is a classic time-delayed Doppler image affected by RFI. Figure 2 (b) is a time-delayed Doppler image of classic sea surface reflection. Figure 2 (c) is a time-delayed Doppler image of classic ice surface reflection. Figure 2 (d) is a time-delay Doppler image suspected of being damaged by RFI, where the horizontal axis is the frequency shift axis and the vertical axis is the time delay axis. It is randomly divided into unsupervised learning dataset and supervised learning dataset according to the ratio of 80% and 20%. The supervised learning dataset is divided into training set, validation set and test set according to the ratio of 80%, 10% and 10%.

[0065] Step S2: Apply two different random augmentations to each DDM image to generate two views for comparative learning;

[0066] Different random enhancements, including random color jittering, random grayscale conversion, and random horizontal flipping, are applied to the images in the unsupervised contrastive learning data to obtain a set of image pairs for contrastive learning. Random color jittering involves random brightness, contrast, saturation, and hue adjustments. The calculation process for random brightness adjustment is as follows:

[0067]

[0068] in For random brightness factor, The input image.

[0069] The calculation process for random contrast adjustment is as follows:

[0070]

[0071] in For random contrast factor, This represents the average grayscale value of the image.

[0072] The calculation process for random saturation adjustment is as follows:

[0073] First, convert the image to grayscale. :

[0074]

[0075] Then the RGB channels are blended:

[0076]

[0077] in This is a random saturation factor.

[0078] Random tone adjustment first converts the image from RGB space to HSV space, then performs an addition operation on the hue channel H. The calculation process is as follows:

[0079]

[0080] Then the image is converted back to RGB space.

[0081] The random grayscale calculation process is as follows:

[0082]

[0083]

[0084] in This represents the probability of grayscale conversion.

[0085] In this embodiment, the random brightness adjustment factor, random contrast adjustment factor, random saturation adjustment factor, and random hue adjustment factor are set to 0.8, 0.8, 0.8, and 0.2, respectively, and the random grayscale factor is set to 0.2. Random enhancement processing is performed on two different DDMs to obtain a pair of images, which are then normalized to the range [0,1].

[0086] Step S3: Select a deep neural network model such as Swin Transformer, MobilenetV2, or ResNet as the backbone network; design a contrastive loss function to learn by comparing the two preprocessed views; load the pre-trained weights, execute the training process, optimize the model parameters, and save the model with the minimum contrastive loss.

[0087] In an embodiment of the present invention, the pre-trained model of the Swin Transformer network is the swin_base_patch4_window7_224 model provided by the HuggingFace website, which has high image classification accuracy.

[0088] The contrastive learning loss function uses the NTXentLoss function, which calculates the similarity between features of two views, maximizing the similarity of positive sample pairs (different views of the same image) while minimizing the similarity of negative sample pairs (views of different images); the calculation process is as follows:

[0089]

[0090] in, , These are the feature embedding vectors of two augmented views of the same DDM image. For cosine similarity, The temperature coefficient is set to 0.5 in this embodiment. Adam is introduced as the optimizer, with an initial learning rate set to 0.0001.

[0091] Load the pre-trained model weights, set the number of self-supervised learning rounds to 500, execute the training process, input image pairs into the network in each training round, calculate the loss of the two image pairs after forward propagation, optimize the parameters through backpropagation, calculate the output loss and optimize the parameters, record the optimal loss, and dynamically save the model parameters corresponding to the optimal loss.

[0092] Step S4: Design a binary classification network head as the classification output, load the contrast loss minimum model weights saved in the previous step, train and validate using DDM image data with RFI labels, and introduce an early stopping mechanism; input the test set data into the trained RFI recognition model for RFI recognition.

[0093] Load the model weights trained in step S3, and then... Figure 3 The neural network structure diagram shown optimizes an existing network by using multilayer perceptron for progressive dimensionality reduction. Each layer is equipped with BatchNorm, GELU, and Dropout, and the classification head is modified to adapt to binary classification tasks.

[0094] A class-balanced loss function, BCEWithLogitsLoss, is introduced to automatically adjust the loss weights based on the ratio of positive to negative samples, thereby alleviating the class imbalance problem. A lower learning rate (lr / 10) is used for the backbone parameters, while a normal learning rate (5e-5) is used for the classification head. The AdamW optimizer is used with a learning rate of 0.00005. OneCycleLR is used to implement periodic changes in the learning rate, balancing training speed and stability, and warm-start is combined to improve convergence speed. Mixed-precision training using Autocast + GradScaler accelerates the training process. The calculation process of the BCEWithLogitsLoss function is as follows:

[0095]

[0096] in The weights output by the model. For the Sigmoid function, This is a real label.

[0097] The training rounds are set to 100. The training process is executed, and the epoch change trends of Loss and Accuracy are recorded and plotted in real time. An early stopping mechanism is introduced. If Accuracy does not improve after eight rounds of training, the training process is terminated and the optimal model parameters are saved.

[0098] The test set is input into the RFI recognition model obtained in the previous step to perform RFI recognition testing, with accuracy, precision, and recall as evaluation criteria.

[0099] In a specific embodiment of the present invention, an RFI recognition model is trained using a Swing Transformer network, and the comparison between the evaluation metrics and the Tianmu-1 reflectance L1 level data quality identifier bit9 (RFI identifier) ​​is shown in Table 1:

[0100] Table 1. Comparison of RFI identification metrics between the two methods

[0101] RFI detection method Accuracy Precision Recall Swin Transformer 0.9981 1 0.9959 bit9 0.6630 0.9697 0.3937

[0102] As shown in Table 1, the RFI recognition model trained using the Swin Transformer network model has significantly higher accuracy and recall compared to the interference recognition method of Tianmu-1 itself, indicating that the former has a clear advantage over the latter in RFI detection.

[0103] In summary, the implementation scheme of this invention not only proposes a new scheme for RFI identification by spaceborne GNSS-R, effectively improving the accuracy of RFI identification, but also combines supervised learning with unsupervised learning, which enhances the generalization ability of the model and improves the model's identification performance.

[0104] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A navigation interference identification method based on spaceborne GNSS-R data, characterized in that, Includes the following steps: S1: Acquire spaceborne GNSS-R reflection observation data, generate channel delay Doppler maps based on the data, and divide the training set, validation set, and test set for semi-supervised learning according to the proportions; the channels include, but are not limited to, GPS, BeiDou, Galileo, and GLONASS satellite navigation systems, and the spaceborne GNSS-R reflection observation data includes at least DDM waveform data; S2: Perform two different random enhancements on each DDM image to generate two data-enhanced views; the data enhancement methods include at least random color jitter, random grayscale conversion, and random horizontal flipping, wherein random color jitter is divided into random brightness, random contrast, random saturation, and random hue adjustment; S3: Input the two data-enhanced views generated in step S2 into the RFI recognition model, set the contrast loss function for contrast learning training; load the pre-trained model weights, execute the training process, calculate the loss of the two image pairs after forward propagation, optimize the model parameters through backpropagation, and obtain the model with the minimum contrast loss. S4: Using the binary classification network head as the classification output, load the model weights with the minimum contrast loss obtained in step S3, train and validate using DDM image data with RFI labels, and introduce an early stopping mechanism; input the test set data into the trained RFI recognition model for RFI recognition. S41: Use BCEWithLogitsLoss and dynamically set the weight coefficients of positive samples to calculate the loss based on the ratio of positive to negative samples in the training data; S42: Use Autocast+GradScaler for mixed-precision training, dynamically manage the numerical precision of different operators, replace some operators with half-precision floating-point numbers, and keep the rest with single-precision floating-point numbers; GradScaler dynamically scales the gradient. S43: Introduce an early stop mechanism, with the main indicators being Loss and Accuracy. Maintain the rate of no improvement for 8 rounds and draw Loss and Accuracy trend charts in real time for visual diagnosis. S44: After each round of training, the model is evaluated on the validation set, the validation loss and accuracy are calculated, and the fine-tuned RFI recognition model weights are saved. S45: Input the test set divided in step S1 into the RFI recognition model, and calculate the evaluation index to evaluate the model performance.

2. The navigation interference identification method based on spaceborne GNSS-R data as described in claim 1, characterized in that: In step S1, the spaceborne GNSS-R reflection observation data includes DDM waveform data. The dataset is randomly divided into an unsupervised learning dataset and a supervised learning dataset according to a ratio of 80% and 20%. The supervised learning dataset is further divided into a training set, a validation set, and a test set according to a ratio of 80%, 10%, and 10%.

3. The navigation interference identification method based on spaceborne GNSS-R data as described in claim 1, characterized in that: In step S2, the random brightness enhancement in the random color jitter specifically involves: ； in For random brightness factor, For the image in Position number The pixel values ​​of each channel; The calculation process for random contrast enhancement is as follows: ； in For random contrast factor, This represents the average grayscale value of the image.

4. The navigation interference identification method based on spaceborne GNSS-R data as described in claim 3, characterized in that: In step S2, the random saturation enhancement in random color jitter specifically involves: Convert the image to grayscale. : ； in , and The images are respectively in Pixel values ​​of the R, G, and B channels at the location; Blending of RGB channels: ； in This is a random saturation factor.

5. The navigation interference identification method based on spaceborne GNSS-R data as described in claim 4, characterized in that: In step S2, the random hue adjustment in random color jitter specifically involves: converting the image from RGB space to HSV space and performing an addition operation on the hue channel H. ； Then convert the image back to RGB space.

6. The navigation interference identification method based on spaceborne GNSS-R data as described in claim 1, characterized in that: The RFI identification model in step S3 includes, but is not limited to, models with Swing Transformer, MobilenetV2, and ResNet as the backbone network, using the NTXentLoss function as the contrastive loss function, specifically: ； in, , These are the feature embedding vectors of two augmented views of the same DDM image. For cosine similarity, This is the temperature coefficient.

7. The navigation interference identification method based on spaceborne GNSS-R data as described in claim 6, characterized in that: In step S41, the calculation process of the BCEWithLogitsLoss function is as follows: ； in The weights output by the model. For the Sigmoid function, This is a real label; In step S43, the Accuracy calculation process is as follows: 。 8. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements the navigation interference identification method based on spaceborne GNSS-R data as described in any one of claims 1-7.

9. A computer device, characterized in that: include Memory, used to store instructions; A processor is configured to execute the instructions, causing the computer device to perform the operation of the navigation interference identification method based on spaceborne GNSS-R data as described in any one of claims 1-7.

Images