Radio Telescope RF Interference Suppression Method Based on Semi-Supervised Deep Learning

Abstract

The present invention discloses a radio telescope radio frequency interference suppression method based on semi-supervised deep learning, including accurately extracting two-dimensional time-frequency data from a standardized FITS file, performing multi-layer two-dimensional wavelet transform on the obtained two-dimensional time-frequency data, decomposing low-frequency components and high-frequency components corresponding to each layer (including horizontal high frequency, vertical high frequency, and diagonal high frequency), and performing RFI elimination on the decomposed low-frequency components and high-frequency components of each layer respectively through a DnCNN denoising method based on semi-supervised learning combined with a channel attention mechanism; the low-frequency components and high-frequency components of each layer after RFI elimination are used as input, and the components are merged through an inverse wavelet transform to reconstruct interference-free time-frequency two-dimensional data with effectively suppressed RFI signals. The present invention can effectively suppress RFI while significantly enhancing the signal-to-noise ratio and maintaining high data fidelity.

Description

Technical Field

[0001] The present invention relates to the field of radio astronomy technology, and in particular to a method for suppressing radio frequency interference of a radio telescope based on semi-supervised deep learning. Background Art

[0002] Existing high-sensitivity radio telescopes such as the Five-hundred-meter Aperture Spherical radio Telescope (FAST) can already detect weak radio astronomy signals, but such equipment is particularly sensitive to radio frequency interference (RFI), which seriously affects the quality of observation data. RFI comes from a wide range of sources, including communication base stations, television broadcasting, satellite communications, and aircraft. Therefore, research on efficient RFI suppression technology is crucial to improving the signal-to-noise ratio of observation data.

[0003] In order to systematically improve the quality of radio astronomy observation data, a phased RFI suppression strategy is usually adopted, which is mainly divided into three stages: the first stage is active prevention, that is, physically isolating RFI sources by selecting station sites and planning quiet zones. Although the effect is remarkable, it is limited by area and sensitivity, and is not enough to deal with dynamic interference sources. The second stage is pre-correlation processing. For single-antenna observations, reference antennas and adaptive filtering technology are used to identify and suppress RFI; for array antennas, spatial filtering is used. However, the effects of both are limited by the prior knowledge of the location of the interference source. Accurately obtaining this information in a complex environment becomes a key challenge. The third stage is post-correlation processing. For the captured observation data, a variety of efficient RFI suppression methods are proposed, which can be roughly divided into four categories: the first category is the threshold statistical method, which uses statistical characteristics (such as mean, root mean square) to set the threshold, identify and process RFI that exceeds the threshold. For example, the Pulsar Phase and Standard Deviation (PPSD) method proposed by Song et al. The key is to identify abnormal phase and standard deviation signals. The efficiency of this type of method is limited by empirical thresholds and may not be effective for new or unknown RFI; the second category is the model-driven method, which uses the physical model of pulsars to protect and restore pulsar signals and reduce the impact of RFI. For example, the singular vector decomposition (SVD) proposed by Pen et al. can identify and suppress interference signals that present specific patterns, but it cannot achieve the best suppression effect for interference that does not have obvious repetitive patterns or has complex and changeable patterns. Shan et al. used compressed sensing (CS) technology to effectively eliminate false peaks caused by RFI in multi-peak pulsar profiles and improve the signal-to-noise ratio. This method usually requires accurate prior knowledge of the sparse nature of the signal, and in practical applications, additional signal processing steps may be required to ensure it. The third category is signal image processing methods, which combine signal processing technology and image processing technology to analyze and process RFI. For example, Lin et al. used wavelet transform technology to effectively remove RFI from astronomical signals through frequency band identification and directional suppression strategies, thereby improving the accuracy and efficiency of interference suppression. The selection of parameters and thresholds in this method usually relies on experience and statistical methods, and its adaptability to complex interference needs to be improved. The fourth category is machine learning methods, which automatically identify and suppress RFI by learning data features. This type of method has been widely used in recent years. Among them, Sun et al. proposed a robust convolutional neural network (CNN) to effectively identify and suppress RFI in SKA1-LOW simulation data and LOFAR and MeerKAT actual observation data, verifying the practicality and effectiveness of the model in different data sources.Zhang et al. further proposed an improved RFI-DRUnet model based on RFI-Unet to repair the dynamic spectrum under the influence of radio frequency interference, improving the quality and accuracy of spectrum data recovery. It can be seen that the fourth type of machine learning method has higher processing efficiency in dealing with complex and changeable RFI. However, in the field of RFI suppression technology, the following challenges still exist:

[0004] (1) Due to the diversity and uncertainty of RFI sources, it is extremely complicated to construct a comprehensive RFI suppression method, which requires the integration of multiple RFI suppression methods.

[0005] (2) When processing large-scale radio astronomy signals, existing RFI detection and suppression methods face performance bottlenecks, limited accuracy and high resource requirements, making it difficult to achieve efficient real-time analysis.

[0006] However, these methods still cannot achieve the best effect in enhancing the signal-to-noise ratio, resulting in the lack of prominent features of useful signals, which affects subsequent signal analysis and processing. At the same time, in the process of suppressing interference, the original signal may be distorted to a certain extent, reducing the fidelity of the data, thereby affecting the quality and scientific value of the observation data. Therefore, how to significantly enhance the signal-to-noise ratio and maintain high data fidelity while effectively suppressing RFI is an urgent problem to be solved in the current field of radio telescope RFI suppression technology. Summary of the invention

[0007] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a radio telescope radio frequency interference suppression method based on semi-supervised deep learning, which can effectively suppress RFI while significantly enhancing the signal-to-noise ratio and maintaining high data fidelity.

[0008] The radio telescope radio frequency interference suppression method based on semi-supervised deep learning of the present invention comprises the following steps:

[0009] Step 1: Data preprocessing: Open the original standardized FITS file, extract the list of all Header-Data Units (HDU), each HDU encapsulates the corresponding data set and its associated metadata, parse the second HDU, obtain the 5-dimensional data it contains (i.e., the number of sub-integration blocks, the number of sampling points in the sub-integration blocks, the number of polarization channels, the number of frequency channels, and the data point value), reshape the 5-dimensional data into the required 2D time-frequency data, normalize it, output the 2D time-frequency data, and save it as a *.npy file;

[0010] Step 2, multi-layer two-dimensional wavelet transform: perform multi-layer two-dimensional wavelet transform on the two-dimensional time-frequency data obtained by preprocessing, and recursively decompose it into three levels, each level contains 1 low-frequency component and 3 high-frequency components, and the next level of components is decomposed from the previous level of low-frequency components, and 3 low-frequency components and 9 high-frequency components are obtained;

[0011] Step 3, DnCNN method to eliminate interference: The dataset is divided into labeled samples and unlabeled samples. The self-training semi-supervised learning method is used to preliminarily train the DnCNN model in combination with the labeled samples. The size of the three low-frequency components and nine high-frequency components of the three levels decomposed by the multi-layer two-dimensional wavelet transform in step 2 is set to 525×525×1, which are used as the original data input respectively. The first layer uses 64 3×3 convolution kernels to perform convolution operations on the original data, and outputs a 64-channel 525×525×64 feature map, which is processed by the ReLU activation function to enhance nonlinearity. From the 2nd to the 16th layer, each layer uses 64 3×3 convolution kernels for convolution operations. For each spatial position (x, y), weighted summation of the 64 channel values ​​of the original image, 64 new feature maps are generated and stacked into an output of 525×525×64, followed by batch normalization, and nonlinearity is introduced through ReLU activation; after convolution, batch normalization and ReLU activation, the 5th and 12th layers introduce the channel attention mechanism to perform refined weighted processing on the feature maps; the 17th layer uses a 3×3 convolution kernel to process the feature data passed from the 16th layer to generate a 525×525×1 interference estimation map, and the original data is subtracted from the interference estimation map, so that the clean data of 3 low-frequency components and 9 high-frequency components without interference of size 525×525×1 after RFI suppression are obtained;

[0012] Step 4, inverse wavelet transform to merge components: Take the 3 low-frequency components and 9 high-frequency components without interference outputted in step 3 as input, select the wavelet basis consistent with the multi-layer two-dimensional wavelet transform in step 2, start from the high-frequency components (detail coefficients) of the highest layer, gradually add the high-frequency components (detail coefficients) to the low-frequency components (approximate coefficients) through the inverse wavelet filter, merge layer by layer until all the components at all levels are merged, and reconstruct the interference-free time-frequency two-dimensional data with the RFI signal effectively suppressed.

[0013] The above-mentioned method for suppressing radio frequency interference of radio telescopes based on semi-supervised deep learning, wherein: the second HDU described in step 1 contains sub-integration information of the observation data, covering the number of channels, number of samples and sub-integration length.

[0014] The above-mentioned method for suppressing radio frequency interference of radio telescopes based on semi-supervised deep learning, wherein: the low-frequency component described in step 2 is an approximate coefficient.

[0015] The above-mentioned method for suppressing radio frequency interference of radio telescopes based on semi-supervised deep learning, wherein: the high-frequency component described in step 2 is a detail coefficient, including horizontal high frequency, vertical high frequency and diagonal high frequency.

[0016] The above-mentioned radio telescope radio frequency interference suppression method based on semi-supervised deep learning, wherein the method of introducing the channel attention mechanism described in step 3 is to use the 5th or 12th layer of the DnCNN model structure as the input feature after convolution, batch normalization and ReLU activation, and compress the spatial dimension of each channel from 525×525 to 1×1 through global average pooling and global maximum pooling operations, respectively, to generate two sets of 1×1×64 feature data, and then the two sets of data are processed by two series of 1×1 convolutional layers, and the ReLU activation function is added in the middle to increase nonlinearity, and two channel attention weights are generated. The two weights are activated by the Sigmoid function and converted into attention coefficients, and then multiplied element by element with the 525×525×64 input features to generate two attention-weighted feature maps, and the two weighted feature maps are added to obtain a comprehensive 525×525×64 output feature.

[0017] Compared with the prior art, the present invention has obvious beneficial effects. It can be seen from the above technical scheme that: the present invention accurately extracts two-dimensional time-frequency data from a standardized FITS file, performs multi-layer two-dimensional wavelet transform on the obtained two-dimensional time-frequency data, decomposes low-frequency components and high-frequency components corresponding to each layer (including horizontal high frequency, vertical high frequency, and diagonal high frequency), aiming to reveal the characteristics of the signal in different directions and scales, which is crucial for subsequent RFI identification; adopts a DnCNN denoising method based on semi-supervised learning combined with a channel attention mechanism to perform RFI elimination on the decomposed low-frequency components and high-frequency components of each layer respectively; uses the low-frequency components (approximate coefficients) and high-frequency components of each layer (detail coefficients, including horizontal, vertical, and diagonal) after RFI elimination as input, merges the components through inverse wavelet transform, and reconstructs interference-free time-frequency two-dimensional data with RFI signals effectively suppressed. The present invention integrates the two-dimensional wavelet transform of signal image processing technology and the DnCNN denoising method of machine learning method, and combines the RFI suppression method of channel attention mechanism, accurately identifies and suppresses RFI through multi-scale and multi-directional signal decomposition, and adopts a semi-supervised learning strategy to effectively integrate limited labeled data and a large amount of unlabeled data, significantly reducing the demand for a large amount of labeled data, while enhancing the generalization and adaptability of the model, and is suitable for practical application scenarios where labeled data is scarce. In addition, the present invention avoids the complex threshold setting in the traditional threshold method, reduces the uncertainty introduced by human factors, and thus improves the accuracy and stability of the processing. By selecting samples such as B1907+02, B1929+10, J1850-0124 and J1905+0400 for experimental comparative analysis, the results show that compared with the rfifind method and the original two-dimensional wavelet transform method, the present invention is superior to the rfifind method and the original two-dimensional wavelet transform method in improving the signal-to-noise ratio and data fidelity, thereby verifying its feasibility and practicality in RFI suppression. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is the principle diagram of the present invention;

[0019] Figure 2 This is the DnCNN network structure diagram;

[0020] Figure 3 This is the network structure diagram of the channel attention mechanism;

[0021] Figure 4 Training flowchart for semi-supervised learning;

[0022] Figure 5-1 This is the rfifind method integrated contour map of the B1929+10 pulsar;

[0023] Figure 5-2This is the integrated contour map of the B1929+10 pulsar using the two-dimensional wavelet transform method;

[0024] Figure 5-3 The integrated contour diagram of the method of the present invention for the B1929+10 pulsar;

[0025] Figure 6-1 This is the time-frequency diagram of the original data of the B1929+10 pulsar;

[0026] Figure 6-2 This is the time-frequency diagram of the B1929+10 pulsar using the rfifind method;

[0027] Figure 6-3 This is the time-frequency diagram of the B1929+10 pulsar using the two-dimensional wavelet transform method;

[0028] Figure 6-4 This is the time-frequency diagram of the method of the present invention for the B1929+10 pulsar. DETAILED DESCRIPTION

[0029] In conjunction with the accompanying drawings and preferred embodiments, the specific implementation methods, features and effects of the radio telescope radio frequency interference suppression method based on semi-supervised deep learning proposed in the present invention are described in detail as follows.

[0030] Example:

[0031] See also Figure 1 , a radio telescope radio frequency interference suppression method based on semi-supervised deep learning of the present invention comprises the following steps:

[0032] Step 1: Data preprocessing: Open the original standardized FITS file and extract the list of all Header-Data Units (HDU). Each HDU encapsulates the corresponding data set and its associated metadata. The second HDU (index 1) contains the sub-integration information of the observed data, including the number of channels, number of samples, and sub-integration length. Parse the second HDU to obtain the 5-dimensional data it contains (i.e., the number of sub-integration blocks, the number of sampling points in the sub-integration blocks, the number of polarization channels, the number of frequency channels, and the data point value). Reshape the 5-dimensional data into the required 2D time-frequency data, normalize it, output the 2D time-frequency data, and save it as a *.npy file.

[0033] Step 2, multi-layer two-dimensional wavelet transform: perform multi-layer two-dimensional wavelet transform on the two-dimensional time-frequency data obtained by preprocessing, and recursively decompose it into three levels, each level contains 1 low-frequency component (approximate coefficient) and 3 high-frequency components (detail coefficient, including horizontal high frequency, vertical high frequency and diagonal high frequency), and the next level of components is decomposed from the previous level of low-frequency components, and 3 low-frequency components and 9 high-frequency components are obtained;

[0034] Step 3: DnCNN method to eliminate interference: The data set is divided into labeled samples and unlabeled samples. The self-training semi-supervised learning method is used to preliminarily train the DnCNN model with labeled samples (such as Figure 2 As shown in the figure, the size of the three low-frequency components and nine high-frequency components of the three levels decomposed by the multi-layer two-dimensional wavelet transform in step 2 is set to 525×525×1, which are used as the original data input respectively. The first layer uses 64 3×3 convolution kernels to perform convolution operation on the original data, and outputs a 64-channel 525×525×64 feature map, which is then processed by the ReLU activation function to enhance nonlinearity. From the 2nd to the 16th layer, each layer uses 64 3×3 convolution kernels for convolution operation, and the 64 channel values ​​of each spatial position (x, y) are weighted summed to generate 64 new feature maps and stacked. The output is 525×525×64, followed by batch normalization, and then nonlinearity is introduced through ReLU activation; after convolution, batch normalization and ReLU activation, the 5th and 12th layers introduce the channel attention mechanism to perform refined weighted processing on the feature map; the 17th layer uses a 3×3 convolution kernel to process the feature data passed from the 16th layer to generate a 525×525×1 interference estimation map, and the original data is subtracted from the interference estimation map, so that the clean data of 3 low-frequency components and 9 high-frequency components without interference of size 525×525×1 after RFI suppression are obtained;

[0035] Among them, the introduction of the channel attention mechanism (such as Figure 3 The method shown in FIG. 1 is to use the 5th or 12th layer of the DnCNN model structure as the input feature after convolution, batch normalization and ReLU activation, and compress the spatial dimension of each channel from 525×525 to 1×1 through global average pooling and global maximum pooling operations, respectively, to generate two sets of 1×1×64 feature data, and then the two sets of data are processed by two series of 1×1 convolution layers, and the ReLU activation function is added in the middle to increase the nonlinearity, and two channel attention weights are generated. The two weights are activated by the Sigmoid function and converted into attention coefficients, and then multiplied element by element with the 525×525×64 input features to generate two attention-weighted feature maps, and the two weighted feature maps are added to obtain a comprehensive 525×525×64 output feature.

[0036] Step 4, inverse wavelet transform to merge components: Take the 3 low-frequency components and 9 high-frequency components without interference outputted in step 3 as input, select the wavelet basis consistent with the multi-layer two-dimensional wavelet transform in step 2, start from the high-frequency components (detail coefficients) of the highest layer, gradually add the high-frequency components (detail coefficients) to the low-frequency components (approximate coefficients) through the inverse wavelet filter, merge layer by layer until all the components at all levels are merged, and reconstruct the interference-free time-frequency two-dimensional data with the RFI signal effectively suppressed.

[0037] Experimental example: Performance verification of the present invention

[0038] On the actual FAST observation data set, a comparative analysis was conducted with the rfifind program and the two-dimensional wavelet transform method to verify the advantages of the present invention in improving the signal-to-noise ratio and data fidelity.

[0039] 3.1 Experimental Environment

[0040] Hardware environment: 4 independent physical computing nodes, 2 Intel Core i7-9700K @ 3.6GHz CPU, 1 Intel Core i7-1065G7 @ 1.5GHz CPU and 1 Intel Core i5-9300H @ 2.4 GHz CPU, with a total of 32 CPU cores (total RAM is 68G, total disk space is 3T).

[0041] Software environment: centos7 system, Anaconda3-4.2.0, python3.8, CUDA11.0, pytorch1.7.1+cu110.

[0042] 3.2 Data Decomposition and Model Training

[0043] The FRB20201124A dataset in the FRB (Fast Radio Burst, FRB) key scientific project was selected as the training sample set. The samples of this dataset have completed the data preprocessing steps of the method of the present invention, and contain 200 samples of 4096×4096 pixels. The Daubechies8 (db8) wavelet function is used to perform 3-layer wavelet decomposition on each sample. The decomposition results include low-frequency approximate components (LL) and high-frequency detail components, the latter covering three components in the horizontal (HL), longitudinal (LH) and diagonal directions (HH). The original data is decomposed into an approximate coefficient layer and two detail coefficient layers, [L0L0, [H1L1, L1H1, H1H1], [H2L2, L2H2, H2H2]], which are low-frequency components and high-frequency components in different directions (the subscript represents the layer number). Among them, L0L0, H1L1, and H2L2 often contain the main components of the interference signal and need to be paid special attention to. Table 1 lists the data scale of these three features.

[0044] Table 1 Data size of different feature data

[0045]

[0046] Based on the above decomposed data samples, a semi-supervised learning strategy of self-training is adopted to train the DnCNN model for the low-frequency components (approximation coefficients) and high-frequency components (detail coefficients) of each layer, aiming to reduce the dependence on labeled data when the labeled data is limited and improve the performance of the model for RFI suppression tasks. The specific training process (such as Figure 4 As shown in Figure 1, the dataset is first divided into two parts: 100 labeled samples with "clean" data for supervised learning, and the other 100 as unlabeled data for semi-supervised iteration. After the model is initially trained with labeled data, the unlabeled data is predicted and the results are added to the training set as pseudo-labels. This process is repeated until the model performance reaches the expected level. Finally, the model is output.

[0047] In order to effectively remove data interference, RFI suppression models are constructed for L0L0, H1L1, and H2L2 respectively. Table 2 lists the training parameters of each feature, where the learning rate (lr) and optimizer are set uniformly, while the batch size (Batch_Size) setting needs to consider resource constraints (including data set size and hardware resource limitations), as well as model complexity and training objectives. Resource constraints determine the amount of data that can be processed, and model complexity and training objectives determine how to optimally utilize this data. By weighing these factors, the batch size (Batch_Size) suitable for the current task can be found to achieve the best training results.

[0048] Table 2 Training parameters of different models

[0049]

[0050] 3.2 Experimental Results Analysis

[0051] After model training, data of four known pulsars observed by the 3021 project in the FAST early science project were selected as test sets: B1907+02, B1929+10, J1850_0124, and J1905+0400, with 100 observation samples each. In order to comprehensively evaluate the performance of the semi-supervised RFI suppression method proposed in this paper, it was connected to the pulsar search pipeline based on the "Standardization of Processing Pulsar Search Mode Data" software package (Pulsar Exploration and Search Toolkit, PRESTO), and the performance was compared with the existing interference removal program rfifind of the PRESTO pipeline and the traditional two-dimensional wavelet transform method. All methods were run under the same software and hardware environment.

[0052] PRESTO is widely used in the field of pulsar detection and astrophysical signal processing. Its workflow effectively processes RFI through the rfifind program and generates a mask (*.mask) file as a key input parameter, and then performs folding processing on the target data (generating a prepfold file). In order to accurately measure the performance of the method, the proposed method and the two-dimensional wavelet transform method are integrated into PRESTO to replace the original rfifind program in its process for comparative analysis. Table 3 shows the parameter settings used in the prepfold processing in this paper.

[0053] Table 3 Parameters used in Prepfold

[0054]

[0055] The evaluation indicators used are Average Signal-to-Noise Ratio (ASNR) and Structural Similarity Index Measure (SSIM). ASNR Represents the arithmetic mean of the signal-to-noise ratio of all data samples of the same pulsar, depending on the signal-to-noise ratio data value provided by the prefold module in PRESTO snr . As shown in formula (1).

[0056] (1)

[0057] in, N is the number of data samples, represents the signal-to-noise ratio value of the i-th data, Indicates the maximum value function. Average It represents the average value of the similarity between the processed signal and the original signal. The calculation formula is shown in formula (2).

[0058] (2)

[0059] Among them, x and y represent the original signal and the processed signal respectively. and are the means of x and y respectively, and are the variances of x and y, respectively, is the covariance of x and y, c1 and c2 are two small constants used to avoid the denominator being zero, usually and , where L is the dynamic range of the signal value, k1 and k2 are small constants, usually k1=0.01 and k2=0.03.

[0060] Table 4 lists in detail the performance comparison of the three methods on the two evaluation indicators of ASNR and PSNR. As can be seen from the table, the method of the present invention shows higher ASNR and SSIM values ​​on all test samples, indicating its advantages in improving signal-to-noise ratio and fidelity. In particular, when processing B1929+10 and J1905+0400 data, the ASNR and SSIM improvements of the method of the present invention are most obvious, which may be attributed to the fact that the characteristics of RFI in these data are more compatible with the design of the method of the present invention. In contrast, the rfifind method and the two-dimensional wavelet transform method do not perform as well as the method of the present invention on some samples.

[0061] Table 4 Comparison of three methods

[0062]

[0063] Figure 5-1 , Figure 5-2 , Figure 5-3 The integral profiles of the B1929+10 pulsar using the rfifind method, two-dimensional wavelet transform, and the method of the present invention are shown respectively. Compared with rfifind and two-dimensional wavelet transform, the method of Example 1 exhibits better interference suppression capability and makes the integral profile smoother. Figure 6-1 , Figure 6-2 , Figure 6-3 and Figure 6-4The time-frequency diagram comparison of the B1929+10 pulsar is shown. There is obvious interference at a specific position in the original data. After rfifind and two-dimensional wavelet transform processing, the interference is reduced but still not completely suppressed. The method of Example 1 almost completely eliminates the interference and effectively retains the key features of the original signal. The above analysis shows that the method of Example 1 has certain advantages over the other two methods.

[0064] From Table 4 and Figure 5-3 , Figure 6-4 It can be seen that through sufficient model training, the method of Example 1 can effectively identify and suppress multiple RFI sources in complex astronomical signals while maintaining high data fidelity. In addition, with the help of semi-supervised learning strategy, the model shows good generalization ability even when the labeled data is limited.

[0065] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A radio telescope radio frequency interference suppression method based on semi-supervised deep learning, characterized in that: The following steps are involved: Step 1, data preprocessing: open the original standardized FITS file, extract the list of all header data units HDU, each HDU encapsulates the corresponding data set and its associated metadata, parse the second HDU, obtain its 5-dimensional data including the number of sub-integration blocks, the number of sampling points in the sub-integration blocks, the number of polarization channels, the number of frequency channels and the data point value, reshape the 5-dimensional data into the required 2D time-frequency data, normalize it, output the 2D time-frequency data, and save it as a *.npy file; Step 2, multi-layer two-dimensional wavelet transform: perform multi-layer two-dimensional wavelet transform on the two-dimensional time-frequency data obtained by preprocessing, and recursively decompose it into three levels, each level contains 1 low-frequency component and 3 high-frequency components, and the next level of components is decomposed from the previous level of low-frequency components, and 3 low-frequency components and 9 high-frequency components are obtained; Step 3, DnCNN method to eliminate interference: The data set is divided into labeled samples and unlabeled samples. The self-training semi-supervised learning method is used to preliminarily train the DnCNN model in combination with labeled samples. The size of the three low-frequency components and nine high-frequency components of the three levels decomposed by the multi-layer two-dimensional wavelet transform in step 2 is set to 525×525×1, which are used as the original data input respectively. The first layer uses 64 3×3 convolution kernels to perform convolution operations on the original data, and outputs a 64-channel 525×525×64 feature map, which is processed by the ReLU activation function to enhance nonlinearity. From the 2nd to the 16th layer, each layer uses 64 3×3 convolution kernels for convolution operations. For each spatial position x, The 64 channel values ​​of y are weighted and summed to generate 64 new feature maps and stacked into an output of 525×525×64, followed by batch normalization and nonlinearity introduced through ReLU activation. After convolution, batch normalization and ReLU activation, the 5th and 12th layers introduce the channel attention mechanism to perform refined weighted processing on the feature maps. The 17th layer uses a 3×3 convolution kernel to process the feature data passed from the 16th layer to generate a 525×525×1 interference estimation map, and the original data is subtracted from the interference estimation map, so that the clean data of 3 low-frequency components and 9 high-frequency components without interference of size 525×525×1 after RFI suppression are obtained. Step 4, inverse wavelet transform to merge components: Take the 3 low-frequency components and 9 high-frequency components without interference outputted in step 3 as input, select the wavelet basis consistent with the multi-layer two-dimensional wavelet transform in step 2, start from the high-frequency components of the highest layer, gradually add the high-frequency components to the low-frequency components through the wavelet inverse filter, merge layer by layer until all the components of all layers are merged, and reconstruct the interference-free time-frequency two-dimensional data with the RFI signal effectively suppressed.

2. The method for suppressing radio frequency interference of a radio telescope based on semi-supervised deep learning as claimed in claim 1, characterized in that: The second HDU described in step 1 contains the sub-integration information of the observed data, including the number of channels, number of samples and sub-integration length.

3. The method for suppressing radio frequency interference of a radio telescope based on semi-supervised deep learning as claimed in claim 1, characterized in that: The low frequency components described in step 2 are approximate coefficients.

4. The method for suppressing radio frequency interference of a radio telescope based on semi-supervised deep learning as claimed in claim 1, characterized in that: The high-frequency components described in step 2 are detail coefficients, including horizontal high frequencies, vertical high frequencies, and diagonal high frequencies.

5. The method for suppressing radio frequency interference of a radio telescope based on semi-supervised deep learning according to claim 1, characterized in that: The method for introducing the channel attention mechanism described in step 3 is to use the 5th or 12th layer of the DnCNN model structure as the input feature after convolution, batch normalization and ReLU activation, and compress the spatial dimension of each channel from 525×525 to 1×1 through global average pooling and global maximum pooling operations, respectively, to generate two sets of 1×1×64 feature data, and then the two sets of data are processed by two series of 1×1 convolution layers, and the ReLU activation function is added in the middle to increase the nonlinearity, and two channel attention weights are generated. The two weights are activated by the Sigmoid function and converted into attention coefficients, and then multiplied element by element with the 525×525×64 input features to generate two attention-weighted feature maps. The two weighted feature maps are added to obtain a comprehensive output feature of 525×525×64.

Images