CN118332443A - Pulsar data radio frequency interference detection method - Google Patents

Abstract

The invention discloses a pulsar data radio frequency interference detection method, and belongs to the fields of image segmentation technology and radioastronomy. The invention comprises the following steps: acquiring a plurality of pulsar data; converting the acquired pulsar data into pulsar time-frequency images, manufacturing labels, and dividing the pulsar time-frequency images with the labels into training sets, verification sets and test sets; constructing a pulsar data radio frequency interference detection model based on the encoder-decoder structure; training and verifying the pulsar data radio frequency interference detection model according to the training set and the verification set to obtain a trained pulsar data radio frequency interference detection model; detecting the pulsar time-frequency image to be detected according to the trained pulsar data radio frequency interference detection model, and obtaining a classification result of the pulsar time-frequency image. The invention has better radio frequency interference detection performance in real data, can detect radio frequency interference more comprehensively and accurately, and achieves good balance in detection efficiency and detection effect.

Description

Pulsar data radio frequency interference detection method

Technical Field

The invention relates to a pulsar data radio frequency interference detection method, and belongs to the fields of image segmentation technology and radioastronomy.

Background

Radio observation is an important means for studying the universe, and can observe celestial bodies and phenomena which cannot be seen by an optical telescope under the condition that visible light cannot penetrate. However, with the continued advancement of astronomical research and technology, although the sensitivity of radio astronomical equipment has been significantly improved, researchers have been able to observe over a wider range of frequencies. At the same time, however, due to the rapid development of human communication technology, signals and noise generated by human activities occupy more and more frequency bands, and the interference to these frequency bands becomes more and more, which seriously affects the quality of radio astronomical observation data. Therefore, in order to guarantee the accuracy and effectiveness of radioastronomical studies, it is important to accurately detect the radio frequency interference (Radio Frequency Interference, RFI) from complex radioobservation data.

The radio frequency interference detection method is mainly divided into two major categories, namely a traditional method and a deep learning method. The traditional radio frequency interference detection method is mainly based on a threshold value, but the selection of the threshold value is greatly influenced by human experience factors, so that the requirement of manual intervention is increased, and the data processing efficiency is greatly reduced. In order to overcome the limitations of the conventional method, the detection of radio frequency interference by using a deep learning technology based on computer vision is also gradually explored. The existing deep learning detection method is mainly based on a convolutional neural network. They detect radio frequency interference in a time-frequency image by extracting features of the radio frequency interference using a convolution operation on the time-frequency image. However, when the existing deep learning models are applied to detecting radio frequency interference in actual observation data of a radio telescope, false detection and missed detection often occur, and further manual detection is required. In addition, as radio astronomical data continues to grow, there is also an urgent need for a model that runs faster and detects more accurately in order to improve the efficiency and quality of data processing.

Disclosure of Invention

The invention provides a pulsar data radio frequency interference detection method, which realizes the interference detection of real pulsar data and achieves better balance between data processing efficiency and detection effect.

The technical scheme of the invention is as follows:

According to a first aspect of the present invention, there is provided a pulsar data radio frequency interference detection method, comprising: acquiring a plurality of pulsar data; converting the acquired pulsar data into pulsar time-frequency images, manufacturing labels, and dividing the pulsar time-frequency images with the labels into training sets, verification sets and test sets; constructing a pulsar data radio frequency interference detection model based on the encoder-decoder structure; training and verifying the pulsar data radio frequency interference detection model according to the training set and the verification set to obtain a trained pulsar data radio frequency interference detection model; detecting the pulsar time-frequency image to be detected according to the trained pulsar data radio frequency interference detection model to obtain a classification result of the pulsar time-frequency image; wherein the classification comprises: interference, non-interference.

Training the training set as input of a pulsar data radio-frequency interference detection model after data enhancement operation and reset operation; and performing a reset operation on the verification set and the test set, and then performing verification and test as input of a pulsar data radio-frequency interference detection model.

The encoder-decoder structure-based pulsar data radio frequency interference detection model is constructed, and specifically comprises the following steps: the encoder stage comprises two convolution layers, 5 stacked MSFE modules and CA modules, and downsampling operation is introduced after the first 4 stacked MSFE modules and CA modules; the decoder stage includes 5 convolutional layers, an up-sampling operation is introduced before the first 4 convolutional layers, and the output results of all pixel points are mapped to a probability range between 0 and 1 using a Sigmoid activation function after the 5 th convolutional layer.

The downsampling operation in the encoding phase uses a2 x 2 convolution and the upsampling operation in the decoding phase uses a2 x 2 transpose convolution.

The MSFE module specifically comprises: firstly, copying an input feature map, then dividing the feature map into four branches, wherein the first branch uses a depth separable convolution extraction feature with the size of 3 multiplied by 3, the second branch uses a depth separable convolution extraction feature with the size of 1 multiplied by 7, then uses a depth separable convolution extraction feature with the size of 7 multiplied by 1, the third branch uses a depth separable convolution extraction feature with the size of 1 multiplied by 11, then uses a depth separable convolution with the size of 11 multiplied by 1 to extract the feature, and the fourth branch uses a depth separable convolution extraction feature with the size of 5 multiplied by 5; and then, splicing the results obtained by the four branches, carrying out batch normalization after splicing, and finally, carrying out element-level addition operation on the batch normalization and the feature map copied before, so as to obtain output features.

The pulsar data radio frequency interference detection model adopts a binary cross entropy loss function to measure the difference between the model output and the real label.

According to a second aspect of the present invention, there is provided a pulsar data radio frequency interference detection apparatus, comprising: the acquisition module is used for acquiring a plurality of pulsar data; the dividing module is used for converting the acquired pulsar data into pulsar time-frequency images and manufacturing labels, and dividing the pulsar time-frequency images with the labels into training sets, verification sets and test sets; the construction module is used for constructing a pulsar data radio-frequency interference detection model based on the encoder-decoder structure; the acquisition module is used for training and verifying the pulsar data radio frequency interference detection model according to the training set and the verification set to acquire a trained pulsar data radio frequency interference detection model; the detection module is used for detecting the pulsar time-frequency image to be detected according to the trained pulsar data radio frequency interference detection model, and obtaining a classification result of the pulsar time-frequency image; wherein the classification comprises: interference, non-interference.

According to a third aspect of the present invention, there is provided a processor for running a program, wherein the program is run to perform the pulsar data radio frequency interference detection method of any one of the above.

According to a fourth aspect of the present invention, there is provided a computer readable storage medium comprising a stored program, wherein the program when run controls a device in which the computer readable storage medium is located to perform any one of the above-described pulsar data radio-frequency interference detection methods.

The beneficial effects of the invention are as follows: aiming at the problems that the interference in the real observation data is generally complex and the detection effect is not ideal, the invention provides a pulsar data radio frequency interference detection method, which has better radio frequency interference detection performance in the real data, can more comprehensively and accurately detect the radio frequency interference, reduces errors, improves edge details, reduces useful signal loss and does not need further manual inspection compared with the prior art. In addition, the invention achieves good balance in detection efficiency and detection effect, and can meet the requirement of processing a large amount of data. Specifically, the pulsar data radio frequency interference detection model of the invention extracts multi-scale characteristics and local information of radio frequency interference in real data through convolution kernels of various scales in an encoder stage, and then uses an attention mechanism to enable the model to focus on global information so as to focus on the characteristics which are more helpful to radio frequency interference detection; and in the decoder stage, the model utilizes a convolution operation of 3 multiplied by 3 to reconstruct and up-sample the feature map, and finally, the detection result of the radio frequency interference is obtained.

Drawings

FIG. 1 is a block diagram of a pulsar data radio frequency interference detection model proposed by the present invention;

FIG. 2 is a block diagram of a multi-scale feature extraction Module (MSFE);

FIG. 3 is a block diagram of a coordinate attention mechanism (CA);

FIG. 4 is a graph of a loss function of the pulsar data during training of the radio frequency interference detection model according to the present invention;

FIG. 5 is a time-frequency image of real data observed by a 40 m radio telescope of the Yunnan astronomical station;

FIG. 6 is a real label corresponding to the time-frequency image of FIG. 5;

FIG. 7 is a diagram of the detection result of the pulsar data RF interference detection model to FIG. 5 according to the present invention;

FIG. 8 is a graph of the detection results of the R-Net model versus FIG. 5;

FIG. 9 is a graph of the detection results of the U-Net model versus FIG. 5;

fig. 10 is a graph of the RFI-Net model versus the test results of fig. 5.

Detailed Description

The invention will be further described with reference to the drawings and examples, but the invention is not limited to the scope.

Example 1: as shown in fig. 1 to 10, according to a first aspect of an embodiment of the present invention, there is provided a pulsar data radio frequency interference detection method, including: acquiring a plurality of pulsar data; converting the acquired pulsar data into pulsar time-frequency images, manufacturing labels, and dividing the pulsar time-frequency images with the labels into training sets, verification sets and test sets; constructing a pulsar data radio frequency interference detection model based on the encoder-decoder structure; training and verifying the pulsar data radio frequency interference detection model according to the training set and the verification set to obtain a trained pulsar data radio frequency interference detection model; detecting the pulsar time-frequency image to be detected according to the trained pulsar data radio frequency interference detection model to obtain a classification result of the pulsar time-frequency image; wherein the classification comprises: interference, non-interference. The pulsar time-frequency image to be tested is from the test set or the obtained pulsar time-frequency image of the non-present test set.

Further, training the training set as input of a pulsar data radio-frequency interference detection model after data enhancement operation and resize operation; and performing a reset operation on the verification set and the test set, and then performing verification and test as input of a pulsar data radio-frequency interference detection model.

Further, the encoder-decoder structure-based pulsar data radio frequency interference detection model is constructed, specifically: the encoder stage comprises two convolution layers, 5 stacked MSFE modules and CA modules, and downsampling operation is introduced after the first 4 stacked MSFE modules and CA modules; the decoder stage includes 5 convolutional layers, an up-sampling operation is introduced before the first 4 convolutional layers, and the output results of all pixel points are mapped to a probability range between 0 and 1 using a Sigmoid activation function after the 5 th convolutional layer.

Further, the 2x2 convolution is used for the downsampling operation in the encoding phase and the 2x2 transpose convolution is used for the upsampling operation in the decoding phase.

Further, the MSFE module specifically includes: firstly, copying an input feature map, then dividing the feature map into four branches, wherein the first branch uses a depth separable convolution extraction feature with the size of 3 multiplied by 3, the second branch uses a depth separable convolution extraction feature with the size of 1 multiplied by 7, then uses a depth separable convolution extraction feature with the size of 7 multiplied by 1, the third branch uses a depth separable convolution extraction feature with the size of 1 multiplied by 11, then uses a depth separable convolution with the size of 11 multiplied by 1 to extract the feature, and the fourth branch uses a depth separable convolution extraction feature with the size of 5 multiplied by 5; and then, splicing the results obtained by the four branches, carrying out batch normalization after splicing, and finally, carrying out element-level addition operation on the batch normalization and the feature map copied before, so as to obtain output features.

Further, the pulsar data radio frequency interference detection model adopts a binary cross entropy loss function to measure the difference between the model output and the real label.

According to a second aspect of an embodiment of the present invention, there is provided a pulsar data radio frequency interference detection apparatus, including: the acquisition module is used for acquiring a plurality of pulsar data; the dividing module is used for converting the acquired pulsar data into pulsar time-frequency images and manufacturing labels, and dividing the pulsar time-frequency images with the labels into training sets, verification sets and test sets; the construction module is used for constructing a pulsar data radio-frequency interference detection model based on the encoder-decoder structure; the acquisition module is used for training and verifying the pulsar data radio frequency interference detection model according to the training set and the verification set to acquire a trained pulsar data radio frequency interference detection model; the detection module is used for detecting the pulsar time-frequency image to be detected according to the trained pulsar data radio frequency interference detection model, and obtaining a classification result of the pulsar time-frequency image; wherein the classification comprises: interference, non-interference. For portions of the foregoing that are not detailed for each module, reference may be made to the relevant description of the embodiments.

According to a third aspect of an embodiment of the present invention, there is provided a processor, configured to execute a program, where the program executes any one of the above methods for detecting pulsar data radio frequency interference.

According to a fourth aspect of an embodiment of the present invention, there is provided a computer readable storage medium, including a stored program, where the program, when executed, controls a device in which the computer readable storage medium is located to perform any one of the above-mentioned methods for detecting pulsar data radio-frequency interference.

Still further, the application provides for the following optional implementation of:

Step 1: psrfits files (PSR J0332+5434, J00358+5413, J0437-4715) from three different pulsar data observed by a 40-meter radio telescope of a Yunnan astronomical platform are selected, and 1384 psrfits files are totally selected; respectively converting the data into a pulsar time-frequency image (selecting data of a first polarized channel of a four-dimensional array (nsub, npol, nchan, nbin) stored in psrfits files, and then averaging nbin dimensions to obtain a time-frequency image of the (nsub, nchan) dimension, wherein nsub represents the number of sub-integrals and represents the time dimension, npol represents the number of polarized channels, nchan represents the number of frequency channels and represents the frequency dimension, and nbin represents the number of phase points); 1384 pulsar time-frequency images and corresponding labels are randomly divided into a training set, a verification set and a test set according to the proportion of 70%, 15% and 15%, and finally 968 pulsar time-frequency images and corresponding labels participate in model training, 207 pulsar time-frequency images and corresponding labels participate in model verification, and 208 pulsar time-frequency images and corresponding labels participate in performance evaluation of the model. In the above, the label is produced as follows: the invention prepares the real labels corresponding to the 1384 pulsar time-frequency images according to aoflagger algorithm: aoflagger is that a variety of marking strategies can be selected, the 1384 pulsar time-frequency images can be marked by selecting aoflagger different strategies, and then the marked results can be manually checked and selected therefrom for model training and evaluation. In the label, white areas indicate interference and black areas indicate no interference.

In order to enhance the generalization capability of the model, data enhancement operation and the size operation are carried out on the training set, wherein the data enhancement operation mainly comprises random horizontal overturn and random vertical overturn, the verification set and the test set only carry out the size operation, and the time-frequency image size when all the data sets are processed and input as the network model is 256 multiplied by 256.

Step 2: constructing a pulsar data radio frequency interference detection model, wherein the detailed structure is shown in figure 1; the pulsar data radio frequency interference detection model comprises two convolution layers, 5 stacked MSFE modules+CA modules and 5 convolution layers which are connected in sequence.

The model is described as follows:

The invention is based on an encoder-decoder architecture, where first two 3 x3 convolutions are used to perform a preliminary extraction of radio frequency interference features for subsequent use. And then the primarily extracted radio frequency interference feature map is sent to a multi-scale feature extraction (MSFE) module to extract multi-scale information of radio frequency interference and obtain a corresponding feature map. These feature maps are then fed into a coordinate attention mechanism (CA) to enhance the model's understanding of the spatial structure of the input data, and by introducing coordinate information, the model can better understand the relationship between the different locations. This attentiveness mechanism can help the model focus better on important areas in the input data, thereby capturing better the spatial structure of the input data. An encoder is constructed by stacking five such multi-scale feature extraction (MSFE) modules described above and a coordinate attention mechanism (CA). The downsampling in the encoding process uses 2x 2 convolutions, which can reduce the computational burden while reducing the loss of spatial and semantic information. Downsampling (downsampling) with each 2x 2 convolution halves the height (H) and width (W) of the time-frequency image feature map, while doubling the number of channels of the feature map.

In the decoder stage, features are further extracted by stacking 43 x 3 convolutions and up-sampling (up sample) using transpose convolutions (Transposed convolution) to gradually restore the resolution of the image, doubling the height (H) and width (W) of the feature map while halving the number of channels. At the final output layer, a1×1 convolutional layer is used to reduce the number of channels to 1. Then, the output results of all the pixel points are mapped to a probability range between 0 and 1 using the Sigmoid activation function. The output binary segmentation mask is obtained by setting the threshold to 0.5: a pixel having a value greater than 0.5 is set to 1 for interference, while a pixel having a value less than 0.5 is set to 0 for no interference.

Further, the MSFE module, as shown in fig. 2, is composed of a series of depth separable convolutions (d) of different scale convolution kernel sizes; utilizing depth separable convolutions may save more computing resources than ordinary convolutions. The specific operation when the input data feature map passes through the multi-scale feature extraction module is as follows: first make a copy of the input feature map and then divide the feature map into four branches, the first branch using a3 x 3 size depth separable convolution to extract features, the second branch first using a 1x 7 size depth separable convolution to extract features and then using a 7 x 1 depth separable convolution to extract features, the third branch first using a 1x 11 size depth separable convolution to extract features and then using an 11 x 1 size depth separable convolution to extract features, the fourth branch using a 5 x5 size depth separable convolution to extract features. And then, splicing the results obtained by the four branches, carrying out batch normalization after splicing, and finally, carrying out element-level addition operation on the batch normalization and the feature map copied before, so as to obtain output features. All operations performed in the multi-scale feature extraction module do not change the size of the feature map channel number. It should be noted that, in this module, a pair of convolution kernels of 1×n1 and n1×1 sizes are used to simulate a convolution kernel of n1×n1 size and a pair of convolution kernels of 1×n2 and n2×1 sizes to simulate a convolution kernel of n2×n2 size, and the purpose of using two pairs of convolution de-extraction features of different sizes is to reduce memory overhead and to better de-detect stripe disturbances in the data, because many of the disturbance shapes in the data are represented as stripe shapes.

Further, the specific structure of the CA module is shown in fig. 3. The CA module aims to enhance the expression capability of learning features of the network model, and can output tensors with the same size after converting and changing any intermediate feature tensors in the network.

When the feature map of the data passes through the CA module, in order to acquire the attention on the width and the height of the image and encode the accurate position information, the input feature map is divided into two directions of the width and the height to be respectively subjected to global average pooling, and the feature maps in the two directions of the width and the height are respectively obtained; then the feature images of the width and height directions of the global receptive field are spliced together, then the feature images are sent to a shared 1X 1 convolution, the dimension of the feature images is reduced to the original C/r, and then the feature images subjected to batch normalization processing are sent to a nonlinear activation function (relu) to obtain the feature images with the shape of 1X (W+H) X C/r(W is the width of the feature map, H is the height of the feature map, C is the number of channels, r is the reduction factor); the characteristic diagram is then displayedRespectively carrying out convolution conversion with the other two convolution kernels of 1 multiplied by 1 according to the original height and widthAnd) Respectively obtaining feature graphs with the same number of channels as the original feature graphs, and then performing Sigmoid activation function on the feature graphs obtained) The attention weights of the feature images in the height are obtained respectivelyAnd an attention weight in the width direction. The formula is as follows:

；

；

in the method, in the process of the invention, Is the attention weight in the height direction,Is the attention weight in the width direction,Is a Sigmoid activation function that is activated by,AndAre all a1 x 1 convolution transform,AndRespectively is a characteristic diagramTwo tensors resolved along the two spatial dimensions of height and width.

According to the calculated attention weight in the height directionAnd an attention weight in the width directionAnd obtaining the characteristic diagram with the attention weight in the width and height directions finally through multiplication weighted calculation on the original characteristic diagram.

The training process of the model is further described as follows:

The detection of radio frequency interference is to judge whether a pixel point is interference or not on a pixel level, which is essentially a pixel-by-pixel classification problem, so that a binary cross entropy loss function (Binary Cross Entropy Loss, BCEloss) can be adopted to measure the difference between the model output and the real label and help the model learn the correct classification boundary. By minimizing the binary cross entropy loss function, the model can better predict the classification of the sample and improve the performance of the model in the classification task. BCEloss can be expressed by the following formula:

；

Wherein, Is the total number of samples to be processed,Is the firstBinary tag values of 0 or 1 for each sample,Is the model pairPredicted values of the individual samples.

Model parameters were updated using an adam optimizer, setting the learning rate to 0.001 and the number of rounds of training (epoch) to 500.

The loss function image of the pulsar data radio frequency interference detection model training is shown in fig. 4, and the loss function is approximately converged around 250 rounds of model training.

Super-parametric optimization with respect to model and model selection:

The training set is used to train the model, and then the super parameter optimization is carried out according to the performance of the verification set model, and the model with the best performance in the verification set is saved.

Model evaluation with respect to model:

The test set is used to evaluate the performance of the model after it is saved. The method comprises the steps of selecting four commonly used evaluation indexes of radio frequency interference detection, namely precision, recall (recall), F1 fraction and cross ratio (iou), to evaluate the model, and further giving training duration (Train duration) results under different models.

The various evaluation index scores for the four models are shown in table 1 below:

TABLE 1

The table above lists the scores of various indexes of the invention and U-Net, RFI-Net and R-Net models, and the table shows that the pulsar data radio frequency interference detection model provided by the invention has the values of 84.12%, 81.24%, 82.66% and 71.63% on four common evaluation indexes for radio frequency interference detection, namely precision, recall (recall), F1 score and cross-over ratio (iou), and the performance of the four evaluation indexes is superior to that of other models, especially compared with R-Net, the performance of the pulsar data radio frequency interference detection model is obviously improved. Further, while the training time of R-Net is the shortest, its performance of the evaluation index is the worst in training time. The model training time length is about 20 minutes slower than that of the R-Net, but is respectively shorter than that of the U-Net and the RFI-Net by one more hours and two more hours, and the model is higher than that of the U-Net and the RFI-Net in score, so that the invention achieves a good balance between the detection efficiency and the detection effect.

An example of a time-frequency image from a test set is shown in fig. 5, where the horizontal axis represents the number of frequency channels representing the frequency axis and the vertical axis represents the number of sub-products representing the time axis. Fig. 6 is a real label corresponding to the time-frequency image of fig. 5, fig. 7 shows the detection result of the pulsar data radio frequency interference detection model proposed by the present invention to fig. 5, and fig. 8, 9 and 10 are the detection results of R-Net, U-Net and RFI-Net to fig. 5, respectively. Comparing these figures can see that the detection result of the invention is very close to the real label, and other models have larger degree of false detection and missing detection.

Therefore, from the visual result, the deep learning model provided by the invention realizes better detection effect on radio frequency interference in real pulsar data of the 40-meter radio telescope of the Yunnan astronomical station, achieves good balance between the detection effect and the detection efficiency, and can meet the requirement of high-speed data processing when the future data volume is more and more.

While the present invention has been described in detail with reference to the drawings, the present invention is not limited to the above embodiments, and various changes can be made without departing from the spirit of the present invention within the knowledge of those skilled in the art.

Claims (9)

1. The pulsar data radio frequency interference detection method is characterized by comprising the following steps of:

Acquiring a plurality of pulsar data;

converting the acquired pulsar data into pulsar time-frequency images, manufacturing labels, and dividing the pulsar time-frequency images with the labels into training sets, verification sets and test sets;

Constructing a pulsar data radio frequency interference detection model based on the encoder-decoder structure;

Training and verifying the pulsar data radio frequency interference detection model according to the training set and the verification set to obtain a trained pulsar data radio frequency interference detection model;

detecting the pulsar time-frequency image to be detected according to the trained pulsar data radio frequency interference detection model to obtain a classification result of the pulsar time-frequency image; wherein the classification comprises: interference, non-interference.

2. The pulsar data radio frequency interference detection method according to claim 1, wherein the training set is trained as an input of a pulsar data radio frequency interference detection model after data enhancement operation and resize operation; and performing a reset operation on the verification set and the test set, and then performing verification and test as input of a pulsar data radio-frequency interference detection model.

3. The method for detecting the radio frequency interference of the pulsar data according to claim 1, wherein the constructing a pulsar data radio frequency interference detection model based on the encoder-decoder structure comprises the following steps: the encoder stage comprises two convolution layers, 5 stacked MSFE modules and CA modules, and downsampling operation is introduced after the first 4 stacked MSFE modules and CA modules; the decoder stage includes 5 convolutional layers, an up-sampling operation is introduced before the first 4 convolutional layers, and the output results of all pixel points are mapped to a probability range between 0 and 1 using a Sigmoid activation function after the 5 th convolutional layer.

4. A pulsar data radio frequency interference detection method according to claim 3, wherein the downsampling operation in the encoding phase uses a2 x2 convolution and the upsampling operation in the decoding phase uses a2 x2 transpose convolution.

5. The method for detecting radio frequency interference of pulsar data according to claim 3, wherein the MSFE module specifically comprises: firstly, copying an input feature map, then dividing the feature map into four branches, wherein the first branch uses a depth separable convolution extraction feature with the size of 3 multiplied by 3, the second branch uses a depth separable convolution extraction feature with the size of 1 multiplied by 7, then uses a depth separable convolution extraction feature with the size of 7 multiplied by 1, the third branch uses a depth separable convolution extraction feature with the size of 1 multiplied by 11, then uses a depth separable convolution with the size of 11 multiplied by 1 to extract the feature, and the fourth branch uses a depth separable convolution extraction feature with the size of 5 multiplied by 5; and then, splicing the results obtained by the four branches, carrying out batch normalization after splicing, and finally, carrying out element-level addition operation on the batch normalization and the feature map copied before, so as to obtain output features.

6. The method of claim 1, wherein the pulsar data radio frequency interference detection model uses a binary cross entropy loss function to measure the difference between the model output and the real tag.

7. A pulsar data radio frequency interference detection device, comprising:

The acquisition module is used for acquiring a plurality of pulsar data;

the dividing module is used for converting the acquired pulsar data into pulsar time-frequency images and manufacturing labels, and dividing the pulsar time-frequency images with the labels into training sets, verification sets and test sets;

the construction module is used for constructing a pulsar data radio-frequency interference detection model based on the encoder-decoder structure;

the acquisition module is used for training and verifying the pulsar data radio frequency interference detection model according to the training set and the verification set to acquire a trained pulsar data radio frequency interference detection model;

The detection module is used for detecting the pulsar time-frequency image to be detected according to the trained pulsar data radio frequency interference detection model, and obtaining a classification result of the pulsar time-frequency image; wherein the classification comprises: interference, non-interference.

8. A processor, wherein the processor is configured to run a program, wherein the program, when run, performs the pulsar data radio frequency interference detection method of any one of claims 1-6.

9. A computer readable storage medium, characterized in that the computer readable storage medium comprises a stored program, wherein the program when run controls a device in which the computer readable storage medium is located to perform the pulsar data radio-frequency interference detection method according to any one of claims 1-6.