CN116894207A - An intelligent radiation source identification method based on Swin Transformer and transfer learning - Google Patents

Abstract

The invention relates to the technical field of signal processing, in particular to an intelligent radiation source identification method based on Swin transform and transfer learning. The use of self-attention mechanisms allows it to better learn the radiation source fingerprint features from new data using a priori knowledge. Even if the data characteristics are changed in the new environment, the general characteristics and modes of the new environment and the original environment can be captured in the mode, and the reusability of the model is improved. The method of using sliding window makes updating network parameters reduce complexity of calculation, and raises training speed and reasoning speed of model.

Description

An intelligent radiation source identification based on Swin Transformer and transfer learning method

Technical field

The invention relates to the field of signal processing technology, and specifically relates to an intelligent radiation source identification method based on Swin Transformer and transfer learning.

Background technique

With the continuous development of communication technology, the number and types of communication equipment continue to increase, and communication protocols continue to change, which makes it increasingly difficult to obtain target information on radiation sources of interest efficiently and accurately. As an important part of communication countermeasures, radiation source identification technology has become increasingly critical. In fact, in the actual electromagnetic environment, the type of radiation source, power, antenna, distance between the radiation source and receiver, modulation information, channel, etc. will affect the fingerprint characteristics of the radiation source, making it difficult to identify the radiation source.

Commonly used radiation source identification methods are roughly divided into two categories, including identification methods based on artificial features and identification methods based on deep learning. Traditional radiation source identification methods based on manual feature extraction or machine learning are very difficult to extract effective features with high discriminability from the detected radiation source signals. Moreover, due to the large amount of signal data information, it is difficult to manually process the data quickly and accurately to distill useful features. Information. The identification method based on deep learning can directly use the collected signal data to achieve end-to-end identification of individual radiation sources, which not only eliminates the manual fingerprint extraction operation, but also has a higher classification accuracy. However, in actual scenarios, effective data are very scarce, most signal data are unlabeled, and available and efficient training sets are difficult to obtain, which cannot meet the data training needs of intelligent radiation source identification based on deep learning. In addition, the current application scenarios of electromagnetic target recognition technology are narrow and cannot achieve cross-scenario and cross-platform recognition. There is an urgent need to find an effective feature recognition method to improve the accuracy of electromagnetic target recognition in scenarios received by different locations and platforms as much as possible. reliability.

Contents of the invention

The purpose of the present invention is to provide an intelligent radiation source identification method based on SwinTransformer and transfer learning. Aiming at the problem of scene migration and identification in cross-receiver scenarios, the unsupervised feature extraction model is used to characterize the electromagnetic target data when the annotation is missing. Extraction, so as to achieve better target recognition results, that is, to expand the application scenarios of individual radiation source identification and realize scene-adaptive intelligent identification of radiation sources.

In order to achieve the above technical objectives and achieve the above technical effects, the present invention is implemented through the following technical solutions:

An intelligent radiation source identification method based on Swin Transformer and transfer learning, including the following steps:

S1: Input the original radiation source signal as the input sequence, perform data enhancement through a combined signal preprocessing method, and generate a series of samples;

S2: Convert each sample generated in step S1 into a vector representation and add position encoding;

S3: Input the vector representation obtained in step S2 into the encoder GPT. The encoder consists of multiple identical layers, each layer including a multi-head attention mechanism and a feedforward neural network;

S4: Input the output of the encoder into the decoder. The decoder is also composed of multiple identical layers. Each layer contains a multi-head attention mechanism, a feedforward neural network and an encoder-decoder attention mechanism;

S5: Convert the output of the decoder into target sequence samples.

Further, the signal preprocessing method includes: retaining all the information of the orthogonal component (Q path) and the in-phase component (I path) of the dual-channel signal in complex form using short-time Fourier transform (STFT) , high-order moments and high-order spectral parameter features (integrated bispectral features) and wavelet transform (Wavelet transform, WT) to perform diversified operations on the input sequence.

Further, the diversification operations specifically include:

For original dual-channel time domain data, data enhancement includes but is not limited to time domain translation, stretching, compression and adding noise; building a network through one-dimensional convolutional layers reduces computational complexity and shortens training time;

For the signal two-dimensional sample sequence after transformation domain or modulation domain transformation, data enhancement includes but is not limited to translation, rotation, scaling, flipping and adding noise to the two-dimensional sample sequence; increasing the diversity of the data set, improving the generalization ability of the model and Accuracy;

The original sample size is set to w×h, the crop size is set to c _w ×c _h , and the number of crops is set to n. Then the number of pictures that can be randomly cropped is:

(wc _w +1)×(hc _h +1)

Flip includes vertical rotation and horizontal rotation; rotation in four directions can expand the number of images four times; the number of samples that can be expanded by one sample is:

4×2×(wc _w +1)×(hc _h +1)

Further, the attention mechanism is a ShiftedWindow local attention mechanism, which specifically includes: first, performing a dot product between the query vector and the key vector to obtain a score vector; then dividing the score vector by an adjustable scaling factor to avoid The score is too large or too small; then the scaled score vector is input into the softmax function to obtain the attention weight; finally, the attention weight and the value vector are weighted and averaged to obtain the final context vector (Attention). This vector Each input position can be assigned a weight to better capture important information in the input sequence;

The formula used to calculate self-attention is:

Q, K, V represent the query vector, key vector and value vector respectively, and d _k represents the adjustable scaling factor.

Furthermore, the Shifted Window local attention mechanism is a hierarchical Transformer, which is calculated by the shifted window method; the shifted window scheme limits the self-attention calculation to non-overlapping local windows while allowing cross-window connections. , thereby improving efficiency;

Shifted Window calculates self-attention within a local window, and the windows are evenly divided in a non-overlapping manner; assuming that each window is divided into sub-windows of M×M size, the calculation of the global MSA module and the window based on the h×w size division sample The complexity is:

Ω(MSA)＝4hwC ² +2(hw) ² C

Ω(W-MSA)＝4hwC ² +2M ² hwC

The moving window method allows communication between windows and has contextual information; at the same time, relative position coding B is added to Q, K, and V in the original formula for calculating self-attention;

Further, the softmax function maps a vector to a probability distribution and converts the output of the neural network into a category probability; the formula of the softmax function is as follows, x _i is the i-th element in the input vector, and n is the length of the vector:

In two-dimensional data processing tasks, the fully connected layer can connect every pixel in the input data to every pixel in the output data, so that every pixel in the output data can be common to all pixels in the input data. Decide.

Beneficial effects of the present invention:

The present invention uses a combined signal preprocessing method for data enhancement, and uses short-time Fourier transform and one-dimensional vector and two-dimensional vector enhancement methods of time domain data to increase the sample size of the data set by several orders of magnitude, thereby expanding The amount of data allows the model to get more information. The use of self-attention mechanism makes it better to use prior knowledge to learn radiation source fingerprint features from new data. Even if the data characteristics change in the new environment, the common characteristics and patterns of the new environment and the original environment can be captured in this way, improving the reusability of the model. Using the sliding window method reduces the computational complexity of updating network parameters and improves the training speed and inference speed of the model. The computational complexity is reduced to the original (M/h) ² , that is, the sub-window, which brings higher efficiency and Training speed. A classification method in which the output of the model is mapped to a specific class of one-layer neural network, using the softmax function in a dual-channel time domain one-dimensional data processing task. Fully connected layers are used in two-dimensional data processing tasks in the changing domain (frequency domain, etc.).

1. Expand the application scenarios of individual radiation source identification: Traditional radiation source identification methods are usually limited to individual identification in specific scenarios, and may not be accurately identified during scene migration or cross-receiver scenarios. The method based on SwinTransformer and transfer learning can enhance the generalization ability of the model and adapt to the identification of radiation sources in different scenarios, thereby expanding the scope of individual radiation source identification applications.

2. Realize scene-adaptive intelligent identification of radiation sources: When the annotation of radiation source data is missing, it is difficult for traditional supervised learning methods to achieve better identification results. By using an unsupervised feature extraction model, this method can extract features from electromagnetic target data in the absence of annotations. This allows the model to effectively perform intelligent identification of radiation sources when faced with incompletely annotated data, achieving scene adaptation.

3. Data enhancement and diversification operations: Data enhancement is a common method. By performing a series of transformation operations on the original data, the diversity of the data set can be increased, thereby improving the generalization ability and robustness of the model. In this method, the signal preprocessing method uses a variety of operations, including time domain translation, stretching, compression, noise addition, and two-dimensional sample sequence translation, rotation, scaling and flipping. These operations can simulate the changes in radiation source signals in different actual scenarios, allowing the model to better adapt to radiation source identification tasks under various circumstances.

4. Use the Shifted Window local attention mechanism: The attention mechanism is an important part of the Transformer model, which can provide different weights for inputs at different positions to better capture important information in the input sequence. The Shifted Window local attention mechanism is a special attention mechanism. By limiting the self-attention calculation to non-overlapping local windows, it can reduce the amount of calculation and allow communication between windows, so that the model can obtain local and global information at the same time. contextual information.

5. Shift window scheme based on hierarchical Transformer: In order to further improve calculation efficiency and model interpretability, a shift window scheme based on hierarchical Transformer is adopted in this method. This scheme reduces computational complexity by limiting self-attention calculations to local windows, while allowing mutual communication between windows, helping to extract more informative features.

6. Use of softmax function: The softmax function is often used to convert the neural network output into a category probability distribution, so that the output of the model more meets the requirements of the classification task. In this method, by using the softmax function, the output of the model can be converted into probability values of each category to facilitate classification decisions.

7. The use of fully connected layers: The fully connected layer is one of the commonly used layers in neural networks. It connects each pixel in the input data to each pixel in the output data, so that each pixel in the output data can be determined jointly by all pixels in the input data. In the task of processing two-dimensional data, the fully connected layer can capture the spatial information in the input data and improve the accuracy of the classification task. In this method, the use of fully connected layers helps to extract fine-grained features of the input data while preserving the spatial relationship between pixels.

Of course, any product implementing the present invention does not necessarily need to achieve all the above-mentioned advantages at the same time.

Description of the drawings

In order to explain the technical solutions of the embodiments of the present invention more clearly, the drawings needed to describe the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic diagram of the overall process of the present invention;

Figure 2 is a schematic diagram of the model architecture of the present invention;

Figure 3 is a schematic diagram of the recognition loss accuracy curve of the resnet50 training set and test set according to the embodiment of the present invention;

Figure 4 is a schematic diagram of the sample recognition confusion matrix trained by resnet50 on the migrated verification set according to the embodiment of the present invention;

Figure 5 is a schematic diagram of the loss accuracy curve of the training set and test set based on Swin Transformer according to the embodiment of the present invention;

Figure 6 is a schematic diagram of using Swin Transformer to identify the confusion matrix of the migrated verification set according to the embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

Example 1

An intelligent radiation source identification method based on Swin Transformer and transfer learning described in this embodiment includes the following steps:

S1: Input the original radiation source signal as the input sequence, perform data enhancement through a combined signal preprocessing method, and generate a series of samples;

S2: Convert each sample generated in step S1 into a vector representation and add position encoding;

S3: Input the vector representation obtained in step S2 into the encoder GPT. The encoder consists of multiple identical layers, each layer including a multi-head attention mechanism and a feedforward neural network;

S4: Input the output of the encoder into the decoder. The decoder is also composed of multiple identical layers. Each layer contains a multi-head attention mechanism, a feedforward neural network and an encoder-decoder attention mechanism;

S5: Convert the output of the decoder into target sequence samples.

Example 2

In this embodiment, in view of the problem of how to expand the amount of data so that the model can obtain more information under the current condition that the amount of tag data is small, the present invention uses a combined signal preprocessing method for data enhancement.

In the new scenario, under the condition that there are few labeled signal samples, how to use a variety of different signal preprocessing methods to generate new training data from existing data to expand the original data set. Studying the effectiveness of these signal preprocessing can help improve sample diversity, expand data and data enhancement, improve model generalization capabilities, reduce overfitting, and improve model robustness.

This invention combines traditional signal analysis methods to retain all the information of the orthogonal component (Q path) and in-phase component (I path) of the dual-channel signal in complex form using short-time Fourier transform (STFT), high Order moment high-order spectral parameter features (integrated bispectral features) and wavelet transform (Wavelet transform, WT) are used to diversify the sample sequence. The samples after the diversification operation are used to build networks through one-dimensional and two-dimensional convolution layers. At the same time, the original dual-channel time domain data can better retain the fingerprint characteristics of the signal. The signal after transformation domain or modulation domain transformation includes time domain and frequency domain. Two types of information in the domain can reduce the impact of frequency on signal fingerprint feature learning. The basic architecture of the model is shown in Figure 2.

For original dual-channel time domain data, time domain translation, time domain stretching, time domain compression, and time domain addition of various noises can be used for data enhancement. Building a network through one-dimensional convolutional layers can reduce computational complexity and shorten training time.

For the two-dimensional sample sequence of the signal after transformation in the transformation domain or modulation domain, data enhancement can be performed on the two-dimensional sample sequence by translating, rotating, scaling, flipping, and adding noise. These data enhancement methods can increase the diversity of the data set, thereby improving the generalization ability and accuracy of the model. For example, for a two-dimensional time-frequency diagram that contains both the time axis and the frequency axis, random cropping, flipping, and rotation can make the model better adapt to different frequencies and times. Adding noise can simulate different channel environments and make the model better adapt to Different noise environments.

In order to effectively enhance data, the present invention uses the method of data combination to enhance data:

By randomly setting parameters for cropping, flipping and rotating, each sample is operated in a different way to increase diversity.

If the original sample size is w×h, the crop size is c _w ×c _h , and the number of crops is n, the number of pictures that can be randomly cropped is:

(wc _w +1)×(hc _h +1)

Flip includes vertical rotation and horizontal rotation; rotating in four directions can quadruple the number of images. The number of samples that can be expanded by a sample is:

4×2×(wc _w +1)×(hc _h +1)

Example 3

In this embodiment, in order to obtain changes in data characteristics in a new environment, how to capture the common characteristics and patterns of the new environment and the original environment and improve the reusability of the model; the data characteristics after the present invention is migrated to the new environment Change, using the self-attention mechanism to better utilize prior knowledge to learn radiation source fingerprint features from new data.

The self-attention mechanism is an attention mechanism. It is a network configuration that aims to solve the problem that the input received by the neural network is many vectors of different sizes, and there is a certain relationship between different vectors, but the actual training When the relationship between these inputs cannot be fully exploited, the performance of the model decreases. The self-attention mechanism can solve this problem by calculating the similarity between each vector and other vectors, and then use these similarities as weights to weight and sum, thereby obtaining a weighted sum vector, which is the self-attention mechanism. output. The following is the basic process of the model of this invention:

First, each radiation source signal in the input sequence is data enhanced through a combined signal preprocessing method to generate a series of samples.

Next, each sample in the input sequence is converted into a vector representation and positional encoding is added.

Next, the vector representation is fed into the encoder, which consists of multiple identical layers, each containing a multi-head attention mechanism and a feed-forward neural network.

The output of the encoder is then input into the decoder, which is also composed of multiple identical layers, each containing a multi-head attention mechanism, a feedforward neural network, and an encoder-decoder attention mechanism.

Finally, the output of the decoder is converted into target sequence samples.

The present invention uses a new local attention mechanism called Shifted Window, which can calculate the global receptive field without increasing the computational complexity. This calculation method can make capturing the relationship between new scenes and old scenes more efficient. And it can obtain all the information related to the output sample points in the feature library. The calculation method of the global receptive field is to start from the current layer and calculate downward layer by layer. For each layer, a pixel on the calculated output feature map corresponds to an area on the input image. This area contains all the information related to the pixel on the input image. This approach can increase the global receptive field without increasing computational complexity. The following is the process for calculating self-attention:

First, a score vector is obtained by taking the dot product of the query vector with the key vector. Next, the score vector is divided by an adjustable scaling factor to avoid scores that are too large or too small.

Then, the scaled score vector is input into the softmax function to obtain the attention weights.

Finally, the attention weight and the value vector are weighted and averaged to obtain the final context vector (Attention), which can assign a weight to each input position to better capture important information in the input sequence.

Here is the formula used to calculate self-attention:

Q, K, V represent the query vector, key vector and value vector respectively, and d _k represents the adjustable scaling factor.

Example 4

In this embodiment, the problem of how to provide a faster and less resource-intensive method for updating model parameters for new tasks is addressed. The present invention uses a sliding window method to reduce the computational complexity of updating network parameters and improve the accuracy of the model. Training speed and inference speed.

The local attention mechanism of Shifted Window used in this invention is a hierarchical Transformer, and its representation is calculated using the shifted window method. The shifted window scheme brings higher efficiency by limiting the self-attention calculation to non-overlapping local windows while also allowing cross-window connections.

ViT calculates self-attention from the perspective of the global window, which requires a relatively large amount of calculation. Shifted Window calculates self-attention within the local window, and the windows are evenly divided in a non-overlapping manner. Assuming that each window is divided into M×M size sub-windows, the computational complexity of the global MSA module and the window based on h×w size divided samples is:

Ω(MSA)＝4hwC ² +2(hw) ² C

Ω(W-MSA)＝4hwC ² +2M ² hwC

The method of moving windows allows communication between windows and contextual information. At the same time, relative position coding B is added to Q, K, and V in the original formula for calculating self-attention.

Example 5

In this embodiment, the problem of how to use a trained network to classify radiation sources through a classifier based on model update is addressed; the present invention is for the classification of a layer of neural network that maps the output of the model to a specific category method, using the softmax function in data processing tasks.

Softmax can be used in one-dimensional data processing tasks. The softmax function can map each element in the input data to between 0 and 1, and the sum of all elements is 1, so that each element in the output data can be regarded as a probability value. This method can effectively transform the information in the input data into a probability distribution, thereby improving the performance of the model. The softmax function maps a vector to a probability distribution and is used to convert the output of the neural network into class probabilities in this classification task. The softmax function takes as input the output of a neural network and maps it to a probability distribution, where each element represents the probability of a class. The formula of the softmax function is as follows, x _i is the i-th element in the input vector, and n is the length of the vector:

In two-dimensional data processing tasks, the fully connected layer can connect every pixel in the input data to every pixel in the output data, so that every pixel in the output data can be common to all pixels in the input data. Decide. This method can effectively extract features from the input data, thereby improving the performance of the model. The fully connected layer uses a forward-structured artificial neural network such as a multi-layer perceptron, including an input layer, an output layer and multiple hidden layers. As a directed graph, a multilayer perceptron is composed of multiple node layers, and each layer is fully connected to the next layer. Except for the input node, each node is a neuron (or processing unit) with a nonlinear activation function.

Example 6

The present invention uses a combined signal preprocessing method for data enhancement, and uses short-time Fourier transform and one-dimensional vector and two-dimensional vector enhancement methods of time domain data to increase the sample size of the data set by several orders of magnitude, thereby expanding The amount of data allows the model to get more information. The use of self-attention mechanism makes it better to use prior knowledge to learn radiation source fingerprint features from new data. Even if the data characteristics change in the new environment, the common characteristics and patterns of the new environment and the original environment can be captured in this way, improving the reusability of the model. Using the sliding window method reduces the computational complexity of updating network parameters and improves the training speed and inference speed of the model. The computational complexity is reduced to the original (M/h) ² , that is, the sub-window, which brings higher efficiency and Training speed. A classification method in which the output of the model is mapped to a specific class of one-layer neural network, using the softmax function in a dual-channel time domain one-dimensional data processing task. Fully connected layers are used in two-dimensional data processing tasks in the changing domain (frequency domain, etc.).

Figure 3 shows the performance of the traditional technology residual network resnet50 in the new scene data set training and testing of the 80th generation. The train_acc and valid_acc on the left represent the training set and test set recognition accuracy curves respectively; the train_loss and valid_loss on the right represent the training set and Test set loss function curve. The lack of convergence of the target domain loss function curve shows that the data and tasks of the new scene are very different from the original scene, causing the existing technology residual network resnet50 to be unable to adapt to the new scene.

Figure 4 shows the recognition accuracy confusion matrix of the new scene data set using the traditional technology residual network resnet50. The targets corresponding to labels 0-15 in the table are targets to be identified. The low accuracy of the diagonal line indicates that resnet50 cannot accurately identify targets to be identified in the new scene data set.

Figure 5 shows the performance of using the present invention in 80 generations of training and testing on the same data set. The train_acc and valid_acc on the left represent the recognition accuracy curves of the training set and the test set respectively; the train_loss and valid_loss on the right represent the loss function curves of the training set and the test set. . The convergence of the target domain loss function curve shows that the present invention can adapt to this new scenario where the data and tasks are very different from the original scenario.

Figure 6 shows the recognition accuracy confusion matrix of the new scene data set using the present invention. The targets corresponding to labels 0-15 in the table are targets to be identified. The higher accuracy of the diagonal line shows that the technology of the present invention can more accurately identify targets to be identified in new scenarios.

The preferred embodiments of the invention disclosed above are only intended to help illustrate the invention. The preferred embodiments do not describe all details, nor do they limit the invention to the specific implementations described. Obviously, many modifications and variations are possible in light of the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention. The invention is limited only by the claims and their full scope and equivalents.

Claims (6)

1. An intelligent radiation source identification method based on Swin transducer and transfer learning is characterized by comprising the following steps:

s1: inputting an original radiation source signal as an input sequence, and performing data enhancement in a combined signal preprocessing mode to generate a series of samples;

s2: converting each sample generated in the step S1 into a vector representation, and adding a position code;

s3: inputting the vector representation obtained in the step S2 into an encoding GPT device, wherein the encoding GPT device is composed of a plurality of identical layers, and each layer comprises a multi-head attention mechanism and a feedforward neural network;

s4: inputting the output of the encoder into a decoder, the decoder also being composed of a plurality of identical layers, each layer containing a multi-headed attention mechanism, a feed-forward neural network and an encoder-decoder attention mechanism;

s5: the output of the decoder is converted into target sequence samples.

2. The intelligent radiation source identification method based on Swin transducer and transfer learning as claimed in claim 1, wherein: the signal preprocessing mode comprises the following steps: all information retaining the two-channel signal quadrature component (Q-way) and in-phase component (I-way) in complex form is used to diversify the input sequence using Short-time fourier transform (Short-time Fouriertransform, STFT), high-order moment Gao Jiepu parametric features (integral bispectral features) and wavelet transform (Wavelet transform, WT).

3. The intelligent radiation source identification method based on Swin transducer and transfer learning as claimed in claim 2, wherein: the diversification operation specifically includes:

for original dual channel time domain data, data enhancements include, but are not limited to, time domain translation, stretching, compression, and noise addition; the network is built through the one-dimensional convolution layer, so that the operation complexity is reduced, and the training time is shortened;

for a signal two-dimensional sample sequence after transformation in a transformation domain or a modulation domain, data enhancement includes, but is not limited to, translation, rotation, scaling, flipping and noise addition of the two-dimensional sample sequence; the diversity of the data set is increased, and the generalization capability and the accuracy of the model are improved;

the original sample size is set as w×h, and the clipping size is set as c _w ×c _h If the clipping number is n, the number of pictures which can be randomly clipped is as follows:

(w-c _w +1)×(h-c _h +1)

the overturning comprises two modes of vertical rotation and horizontal rotation; the number of images can be extended by four times by rotating in four directions; the number of samples that a sample can expand is:

4×2×(w-c _w +1)×(h-c _h +1)。

4. the intelligent radiation source identification method based on Swin transducer and transfer learning as claimed in claim 1, wherein: the attention mechanism is a Shifted Window local attention mechanism, and specifically includes: firstly, carrying out dot product on the query vector and the key vector to obtain a fraction vector; dividing the score vector by an adjustable scaling factor to avoid excessive or insufficient scores; then inputting the scaled score vector into a softmax function to obtain an attention weight; finally, the Attention weight and the value vector are weighted and averaged to obtain a final context vector (Attention), and the vector can be distributed with a weight for each input position so as to better capture important information in an input sequence;

the formula for calculating self-attention is:

q, K, V represent a query vector, a key vector, and a value vector, d, respectively _k Representing an adjustable scaling factor.

5. The intelligent radiation source identification method based on Swin transducer and transfer learning as claimed in claim 4, wherein: the Shifted Window local attention mechanism is a layered transform, and the representation is calculated by a shift Window method; the shift window scheme has the effect of improving efficiency by limiting self-intent calculation to non-overlapping local windows and allowing cross-window connection;

shifted Window calculates self-attention in local windows, which are evenly divided in a non-overlapping manner; let each window be divided into sub-windows of m×m size, the computational complexity of the global MSA module and the windows dividing samples based on h×w size be:

Ω(MSA)＝4hwC ² +2(hw) ² C

Ω(W-MSA)＝4hwC ² +2M ² hwC

the window is moved, so that communication can be carried out between the windows, and the window has contextual information; meanwhile, the relative position codes B are added when Q, K and V in the formula of self-attention are originally calculated;

6. the intelligent radiation source identification method based on Swin transducer and transfer learning as claimed in claim 4, wherein: the softmax function maps a vector to a probability distribution, converting the output of the neural network into class probabilities; the softmax function is given by the formula x _i Is the i-th element in the input vector, n is the length of the vector:

in a two-dimensional data processing task, the fully connected layer may connect each pixel in the input data with each pixel in the output data, such that each pixel in the output data may be determined jointly by all pixels in the input data.