Method for recognizing spatial target behavior based on time-frequency characteristics of radio signals

Abstract

The application provides a space target behavior recognition method based on radio signal time-frequency characteristics, obtains a transmission signal of a detected target, and performs time-frequency conversion on the transmission signal to obtain a time-frequency graph containing space target behavior; a light convolutional network is used to extract features of the time-frequency graph, the advantages of CNN lightness, high efficiency, a self-attention mechanism of a transformer and a global receptive field are combined to obtain local features and global features of the time-frequency graph, the local features and the global features are fused to obtain fused features, and the fused features are classified to obtain a prediction label of space target behavior in the time-frequency graph. The application can recognize different behaviors of a target faster while ensuring high precision, has the advantages of low calculation amount and low memory requirement, and can also meet the requirements of performance and accuracy. Good recognition precision and recognition speed are still maintained under various non-ideal conditions such as spectrum offset, phase offset and amplitude offset.

Description

Technical Field

[0001] This invention belongs to the field of space target recognition technology, specifically relating to a method for recognizing space target behavior based on the time-frequency characteristics of radio signals. Background Technology

[0002] For critical spacecraft, reconnaissance satellites, and other space threat targets, analyzing the composition and anomalies of their key components provides crucial information support for precise tracking and attack. Every orbital change or close-range reconnaissance maneuver by a space target is accompanied by an attitude change. Therefore, it is urgent to comprehensively and rapidly perceive the attitude and anomalies of on-orbit space targets to effectively detect and predict their on-orbit behavior, thereby safeguarding my country's space assets and critical national infrastructure. Currently, broadband radar and radio reconnaissance equipment are primarily used to monitor space targets. However, due to signal repetition frequency limitations, it is difficult to obtain well-focused imaging results after the target undergoes rapid rotation. Radio equipment can detect signals from radar, satellite, and other payloads on space targets, and analyze their waveforms and spectra to obtain behavioral characteristics of the space targets, which is of great significance for protecting the normal on-orbit operation of my country's satellites.

[0003] Current radar target recognition technology utilizes radar and other detection equipment to analyze and judge the radar echoes of targets. However, with the increasing number of space targets—tens of thousands—monitoring them all using high-resolution imaging technology would be prohibitively expensive. Target imaging is one of the key characteristics of radar for detecting and identifying space targets; however, factors such as illumination, target characteristics, and environmental random noise can cause image degradation, affecting the efficiency of space target identification. There is currently limited research on the behavior recognition of space targets based on time-frequency characteristics; most existing target recognition technologies focus on the behavior recognition of conical targets. Many scholars have conducted extensive research on the behavior recognition of ballistic conical targets, which can now be broadly categorized into three main types:

[0004] The first category involves constructing and extracting manually statistical features. This method first analyzes the target's scattering characteristics, then extracts regions containing micro-motion information from the time-frequency distribution of the target's received signal based on energy intensity. These regions include statistical features such as time-frequency average entropy, spectral entropy, and peak-to-peak ratio of the autocorrelation function. A support vector machine is then used to classify the micro-motions of the cone target, including spin, precession, and nutation. The second category involves manually extracting the Doppler spectrum peak spacing after setting a micro-Doppler threshold. This method first establishes mathematical models for different forms of micro-Doppler, summarizes the peak distribution forms of the cone target's scattering source micro-Doppler spectrum under different micro-motion forms, and points out classification methods using micro-Doppler thresholds to identify spin and using whether the micro-Doppler spectrum peaks are evenly spaced to identify nutation and precession. While the above classification methods have achieved good results, constructing statistical features and transforming signal models require certain prior knowledge, and their universality needs further discussion. Furthermore, manual feature extraction inevitably introduces human error, significantly impacting classification accuracy. Moreover, effective micro-Doppler information cannot be extracted when the signal-to-noise ratio is low, leading to failure. Therefore, it is necessary to introduce intelligent classification methods. The third category involves using deep learning for identification. In recent years, deep learning, as an intelligent classification method, has been applied to image processing, human recognition, and radar target recognition, such as the detection and classification of micro-moving targets at sea, ship classification, and aircraft target recognition. For example, a Bidirectional Long Short-Term Memory (BiLSTM) network can be used to process the range cells of high-resolution range imaging samples, combined with a Softmax classifier and a voting mechanism to achieve target recognition. Alternatively, by analyzing the micro-movement patterns and time-frequency characteristics of the target, a Recurrent Neural Network (RNN) or a One-Dimensional Convolutional Neural Network (CNN) can be used to process the time-frequency map to achieve target classification.

[0005] In summary, existing technologies are only suitable for simple applications of typical networks. Under non-ideal channel conditions, the network's recognition performance is affected, leading to a decrease in recognition rate and thus instability and poor robustness. While complex networks, such as Alexnet-BiLSTM, offer some improvement in recognition performance, the computational complexity increases, hindering the rapid and effective identification of spatial targets. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a spatial target behavior recognition method based on the time-frequency characteristics of radio signals. The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides a spatial target behavior recognition method based on the time-frequency characteristics of radio signals, comprising:

[0008] S100: Acquire the transmission signal of the detected target, and perform time-frequency transformation on the transmission signal to obtain a time-frequency map containing the behavior of the space target;

[0009] S200: Input the time-frequency map from S100 into a trained lightweight convolutional network to extract features from the time-frequency map. The local feature extraction module projects these features into a high-dimensional space to obtain local features of the time-frequency map. These local features are then divided into multiple patches, and the relationships between each patch are encoded to obtain a corresponding global patch representation. The global patch representations are then folded to obtain a global time-frequency map representation. This global representation is then convolved point-by-point and merged with the features extracted from the time-frequency map by the lightweight convolutional network to obtain fused features. The fused features are then classified to obtain predicted labels for spatial target behavior in the time-frequency map.

[0010] In a second aspect, the present invention provides a spatial target behavior recognition device based on the time-frequency characteristics of radio signals, comprising:

[0011] The acquisition module is configured to acquire the transmission signal of the detected target and perform a time-frequency transformation on the transmission signal to obtain a time-frequency map containing the behavior of the space target;

[0012] The recognition module is configured to input the time-frequency map acquired by the acquisition module into a trained lightweight convolutional network to extract features from the time-frequency map using the lightweight convolutional network. A local feature extraction module then projects these features into a high-dimensional space to obtain local features of the time-frequency map. These local features are then divided into multiple patches, and the relationships between each patch are encoded to obtain a corresponding global patch representation. The global patch representation is then folded to obtain a global time-frequency map representation. This global representation is then convolved point-by-point and merged with the features extracted by the lightweight convolutional network to obtain a fused feature. Finally, the fused feature is classified to obtain predicted labels for spatial target behaviors in the time-frequency map.

[0013] Thirdly, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0014] Memory, used to store computer programs;

[0015] When a processor executes a program stored in memory, it implements the steps of the method described in the first aspect.

[0016] Beneficial effects:

[0017] This invention provides a method for recognizing spatial target behavior based on the time-frequency characteristics of radio signals. It involves acquiring the transmitted signal of a detected target and performing a time-frequency transformation on the signal to obtain a time-frequency map containing the target's behavior. A lightweight convolutional network is used to extract features from the time-frequency map, and a local feature extraction module projects these features into a high-dimensional space to obtain local features. These local features are then divided into multiple patches, and the relationships between each patch are encoded to obtain a corresponding global representation. The global representation is then folded to obtain a global representation of the time-frequency map. This global representation is then convolved point-by-point and merged with the features extracted by the lightweight convolutional network to obtain a fused feature. The fused feature is then classified to obtain predicted labels for the spatial target behavior in the time-frequency map. This invention achieves high accuracy while recognizing different target behaviors more quickly, maintaining good recognition accuracy and speed even under various non-ideal conditions such as spectral shift, phase shift, and amplitude shift.

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 A flowchart of a spatial target behavior recognition method based on the time-frequency characteristics of radio signals provided in an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram illustrating the process of training the lightweight convolutional network of this invention.

[0021] Figure 3 This is a schematic diagram of the lightweight convolutional network structure for training according to the present invention;

[0022] Figure 4 This is a comparison chart of the accuracy of the algorithm proposed in this invention with other different methods on a simulation dataset;

[0023] Figure 5 This is a schematic diagram of the confusion matrix of the algorithm proposed in this invention at a signal-to-noise ratio of 0dB. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0025] refer to Figure 1 This invention provides a spatial target behavior recognition method based on the time-frequency characteristics of radio signals, comprising:

[0026] S100: Acquire the transmission signal of the detected target, and perform time-frequency transformation on the transmission signal to obtain a time-frequency map containing the behavior of the space target;

[0027] S200: Input the time-frequency map from S100 into a trained lightweight convolutional network to extract features from the time-frequency map. The local feature extraction module projects these features into a high-dimensional space to obtain local features of the time-frequency map. These local features are then divided into multiple patches, and the relationships between each patch are encoded to obtain a corresponding global patch representation. The global patch representations are then folded to obtain a global time-frequency map representation. This global representation is then convolved point-by-point and merged with the features extracted from the time-frequency map by the lightweight convolutional network to obtain fused features. The fused features are then classified to obtain predicted labels for spatial target behavior in the time-frequency map.

[0028] The prediction labels of this invention indicate the type to which the behavior of a spatial target belongs.

[0029] Since micro-Doppler modulation information can be clearly reflected in the time-frequency graph, the lightweight convolutional network of this invention combines inverse residual connections (structure) and a self-attention mechanism. This allows for the effective combination of global and local features with a small number of parameters, ensuring reduced computational complexity and inference speed while improving accuracy. Therefore, this invention maintains good recognition performance even under various non-ideal conditions such as spectral shift, phase shift, and amplitude shift.

[0030] refer to Figure 2 The training process of the lightweight convolutional network of the present invention includes:

[0031] Step 1: Perform time-frequency analysis on the received signal to construct a dataset, and divide the dataset into a training set and a test set. The dataset includes multiple samples, each of which is a time-frequency map containing different spatial target behaviors, and each time-frequency map carries a real label.

[0032] The construction process in step 1 is as follows:

[0033] Step 11: Acquire multiple target transmission signals with BPSK modulation.

[0034] Let the modulation scheme of the transmitted signal be BPSK, where A(t) is the power and τ is the time delay, then the expression for the received signal is:

[0035]

[0036] Step 12: Use time-frequency analysis to perform time-frequency transformation on the transmitted signal to obtain multiple time-frequency maps containing the behavior of space targets;

[0037] Under various non-ideal conditions, time-frequency analysis is performed on the transmitted signal. The Short-Time Fourier Transform (SFT) is one of the most commonly used time-frequency analysis methods. It calculates the Fourier transform of the signal within each time window using a sliding time window, thus obtaining the two-dimensional time-frequency distribution. Generally, the horizontal direction of the spectrum represents time, and the vertical direction represents frequency. The discrete form of the SFT is:

[0038]

[0039] Where z(k) is the discrete form of the signal, T and F are the time and frequency sampling intervals, m and n are the time and frequency sampling intervals, and ω(k) is the window function.

[0040] Step 13: Combine the multiple time-frequency images containing spatial target behavior into a dataset;

[0041] Step 14: Divide the dataset into a training set and a test set in a 7:3 ratio.

[0042] The dataset used in the experiment was obtained by performing time-frequency transformation on the received signal. A total of 6000 time-frequency images of spatial target behavior were included, covering five categories: three-axis stabilization, three-axis slow rotation, three-axis fast rotation, precession, and orbital lifting. Each behavior category contained 1200 time-frequency images, distributed across five signal-to-noise ratios (-10dB, -5dB, 0dB, 5dB, 10dB), with each image measuring 64×64 pixels. Different rotation speeds were used for fast rotation, slow rotation, and precession. The dataset was divided into training and test sets in a 7:3 ratio.

[0043] Step 2: In each iteration, a set of time-frequency maps from the training set are input into a pre-built lightweight convolutional network to obtain the predicted label for each time-frequency map.

[0044] refer to Figure 3 ,exist Figure 3 In the code, `conv` represents a convolutional module, and the number 3 after the module name indicates the kernel size. `Inverted Residual Block` is the inverted residual module; an `Inverted Residual Block` without an arrow indicates a stride of 1, while an `Inverted Residual Block` with an arrow indicates a stride of 2. `MobileViT Block` is the MobileViT module. (The text repeats itself here.) Figure 3 As shown, the lightweight convolutional network of the present invention includes a convolutional module with a kernel of 3, an inverse residual module with a stride of 1, an inverse residual module with a stride of 2, a MobileViT module, a convolutional module with a kernel of 1, and a classification module.

[0045] The lightweight convolutional network consists of modules connected sequentially in the following order: a convolutional module, a reverse residual module with a stride of 1, a reverse residual module with a stride of 2, a reverse residual module with a stride of 1, a reverse residual module with a stride of 1, a reverse residual module with a stride of 2, a MobileViT module, a reverse residual module with a stride of 2, a MobileViT module, a reverse residual module with a stride of 2, a MobileViT module, a convolutional module with a kernel of 1, and a classification module.

[0046] The convolutional module with a kernel of 3 includes multiple input channels and multiple output channels. The convolutional module with a kernel of 3 includes a convolutional layer, a batch normalization layer, and a SiLu layer connected in sequence. The inverse residual module with a stride of 2 and the inverse residual module with a stride of 1 both include a convolutional layer, a batch normalization layer, and a SiLu layer connected in sequence, and a linear activation function is connected after the SiLu layer through pointwise convolution. The inverse residual module with a stride of 1 uses residual connections between the convolutional layer, the batch normalization layer, the SiLu layer, and the activation function connected in sequence.

[0047] In the specific experiment, the convolutional module with a kernel of 3 was set with a stride of 1, 3 input channels, 16 output channels, and padding of 1.

[0048] The inverse residual module avoids the gradient vanishing problem that occurs as the network deepens. High-dimensional information is lost less after passing through the SiLu activation function, thus increasing feature representation capability. The specific structure involves first performing pointwise convolution, followed by batch normalization layers and SiLu layers, then using depthwise separable convolution (DSC) to extract features, followed by batch normalization layers and SiLu layers, and finally connecting to a linear activation function using pointwise convolution. Residual connections are used for the portion of the depthwise separable convolution with a stride of 1, and their expression is:

[0049] H(x)=F(x)+x (3)

[0050] Where F(x) is the data after passing through the inverted residual module, x is the data before passing through the inverted residual module, and H(x) is the data obtained after fusing the two.

[0051] Continue to refer to Figure 3 Each MobileViT module of the present invention includes a local representation module, a global representation module, and a fusion module;

[0052] The local representation module includes an n×n and a 1×1 convolutional block for extracting features from the input data and projecting the features into a high-dimensional space. The global representation module divides the data output by the local feature module into multiple patches, and then encodes the relationships between each patch to obtain the corresponding patch local representation. The patch local representation is then folded back to obtain the global representation. The global features are input to the fusion module. The fusion module performs pointwise convolution on the global representation and then merges it with the extracted features from the input data to obtain the fused features.

[0053] In the global representation module, the data is first transformed, then input into the Transformer, and finally transformed back to its original form. Specifically, the data X∈R H×W×C Data X is obtained after passing through the local representation module. L ∈R H×W×d , where d>C, C is the number of channels, and H and W are the height and width of the feature map.

[0054] Then, the data is divided into multiple patches, with each patch having a length and width of h and w, respectively, and the total number of patches is [number missing].

[0055]

[0056] Where P = wh. The final result after expansion is the data X. U ∈R P×N×d .

[0057] Then, for each p∈{1,...,P}, the relationships between patches are encoded using a Transformer to obtain X. G ∈R P ×N×d The expression is:

[0058] X G (p) = Transformer(X) U (p)) 1≤p≤P (5)

[0059] Then X G ∈R P×N×d Restore to X F ∈R H×W×d Because of X U (p) uses convolutional encoding to obtain local information from an n×n region, while X G (p) Encodes global information for the p patches at the p-th position, therefore X G Each pixel in X can encode information from all pixels in X, so the overall receptive field of the network is H×W.

[0060] The fusion module refers to the module that integrates the global representation X.F ∈R H×W×d The output X obtained after pointwise convolution Fu ∈R H×W×d With the original input information X∈R H×W×C The features of these connections are then merged and then another n×n convolutional layer is used to fuse them.

[0061] After stacking the global representation module, local representation module, and fusion module, the final fused feature X is obtained after passing through pointwise convolution, average pooling layers, and linear activation layers. L This invention will fuse feature X L The input is given to the classifier, and the prediction formula is:

[0062] p(y|x)=softmax(C(X L (6)

[0063] Step 3: Calculate the cross-entropy loss using the predicted and true labels from the time-frequency graph in Step 2.

[0064] Based on the predicted probability, the predicted label of the sample can be obtained as follows: The loss function used is cross-entropy loss, and its expression is:

[0065]

[0066] Where N is the total number of categories, y i p represents the true class of the sample as class i. i This represents the probability that the model predicts the class to be i.

[0067] Step 4: Determine the direction of the decrease in cross-entropy loss using stochastic gradient descent, and update the parameters of the lightweight convolutional network according to this direction;

[0068] Step 5: Repeat steps 2 to 4 until the maximum number of iterations is reached to obtain a trained lightweight convolutional network.

[0069] This step involves setting the maximum number of iterations for the network model to Epoch and the number of training categories to C; training the network model; calculating the loss function for the current iteration based on the recognition results and their corresponding labels; and then modifying the network model's parameters using stochastic gradient descent to obtain the updated network model. This training process is then repeated until the maximum number of iterations is reached.

[0070] This invention can also be used to test trained lightweight convolutional networks, as follows:

[0071] Step 6: Take out all the samples in the test set in sequence and input them into the trained lightweight convolutional network to perform classification prediction and obtain the predicted label of the input sample;

[0072] Step 7: Compare the predicted labels and real labels of the input samples from Step 6 to obtain the classification and recognition results;

[0073] Step 8: Statistically analyze the recognition results of all input samples to obtain the classification accuracy of the trained lightweight convolutional network.

[0074] The classification module includes a convolutional layer and a classifier. The convolutional layer convolves the fused features and then inputs them into the classifier. The classifier classifies the convolved fused features to obtain the predicted labels of spatial target behavior in the time-frequency map.

[0075] In a second aspect, the present invention provides a spatial target behavior recognition device based on the time-frequency characteristics of radio signals, comprising:

[0076] The acquisition module is configured to acquire the transmission signal of the detected target and perform a time-frequency transformation on the transmission signal to obtain a time-frequency map containing the behavior of the space target;

[0077] The recognition module is configured to input the time-frequency map acquired by the acquisition module into a trained lightweight convolutional network to extract features from the time-frequency map using the lightweight convolutional network. A local feature extraction module then projects these features into a high-dimensional space to obtain local features of the time-frequency map. These local features are then divided into multiple patches, and the relationships between each patch are encoded to obtain a corresponding global patch representation. The global patch representation is then folded to obtain a global time-frequency map representation. This global representation is then convolved point-by-point and merged with the features extracted by the lightweight convolutional network to obtain a fused feature. Finally, the fused feature is classified to obtain predicted labels for spatial target behaviors in the time-frequency map.

[0078] Thirdly, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0079] Memory, used to store computer programs;

[0080] When a processor executes a program stored in memory, it implements the steps of the method described in the first aspect.

[0081] The effects of this invention can be further illustrated by the following simulation experiments.

[0082] Using the data processed in step 1 as the experimental dataset, 70% of the samples were randomly selected to form the training set, and the remaining 30% of the samples were used as the test set.

[0083] To verify the spatial target recognition performance of the method of this invention, Table 1 lists the detailed experimental parameters and provides verification results. The confusion matrix of the experimental results at 0 dB is shown below. Figure 2 As shown, the present invention has good recognition performance.

[0084] Table 1 Network simulation parameter settings

[0085] Number of space target categories 5 Time-frequency graph size 64×64 Learning rate 1e-4 Training rounds 100 Optimizer Adam loss function Cross-entropy loss Signal-to-noise ratio [-10, -5, 0, 5, 10] dB

[0086] To evaluate the effectiveness of the proposed spatial target behavior recognition method, this invention compares the classification performance of AlexNet and VggNet transfer learning methods for classifying micro-motions of cone targets. Experiments were conducted using the aforementioned dataset on the proposed method, and the results are as follows: Figure 3 As shown, VggNet's accuracy is somewhat lower. AlexNet and MobileViT both achieved recognition accuracy of over 90% at signal-to-noise ratios above 0dB. However, as the signal-to-noise ratio decreases, AlexNet's recognition accuracy shows a significant downward trend. Compared to recognition methods based on AlexNet and VggNet, the method proposed in this invention achieves the best recognition performance in both high and low signal-to-noise ratios.

[0087] It is worth noting that the terms "first" and "second" in this invention are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0088] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0089] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for spatial target behavior recognition based on the time-frequency characteristics of radio signals, characterized in that, include: S100: Acquire the transmission signal of the detected target, and perform time-frequency transformation on the transmission signal to obtain a time-frequency map containing the behavior of the space target; S200: Input the time-frequency map from S100 into the trained lightweight convolutional network to extract features from the time-frequency map. Then, use a local feature extraction module to project these features into a high-dimensional space to obtain local features of the time-frequency map. Divide the local features of the time-frequency map into multiple patches, and encode the relationships between each patch to obtain a corresponding global patch representation. Fold the global patch representations to obtain a global representation of the time-frequency map. Convolve the global representation of the time-frequency map point by point, and then merge it with the features of the time-frequency map extracted by the lightweight convolutional network to obtain a fused feature. The fused features are classified to obtain predicted labels for spatial target behavior in the time-frequency map; the lightweight convolutional network includes a convolutional module with 3 kernels, an inverse residual module with a stride of 1, an inverse residual module with a stride of 2, a MobileViT module, a convolutional module with a kernel of 1, and a classification module; The lightweight convolutional network consists of modules connected sequentially in the following order: a convolutional module, an inverse residual module with a stride of 1, an inverse residual module with a stride of 2, an inverse residual module with a stride of 1, an inverse residual module with a stride of 1, an inverse residual module with a stride of 2, a MobileViT module, an inverse residual module with a stride of 2, a MobileViT module, an inverse residual module with a stride of 2, a MobileViT module, a convolutional module with a kernel of 1, and a classification module. Each MobileViT module includes a local representation module, a global representation module, and a fusion module; The local representation module includes a and The convolutional blocks are used to extract features from the input data and project the features into a high-dimensional space; the global representation module divides the data output by the local feature module into multiple patches, and then encodes the relationship between each patch to obtain the corresponding patch local representation; the patch local representation is folded back to obtain the global representation; the global features are input to the fusion module; the fusion module convolves the global representation point by point, and then combines it with the features of the extracted input data to obtain the fused features.

2. The spatial target behavior recognition method based on the time-frequency characteristics of radio signals according to claim 1, characterized in that, The training process of the lightweight convolutional network includes: Step 1: Perform time-frequency analysis on the received signal to construct a dataset, and divide the dataset into a training set and a test set. The dataset includes multiple samples, each of which is a time-frequency map containing different spatial target behaviors, and each time-frequency map carries a real label. Step 2: In each iteration, a set of time-frequency maps from the training set are input into a pre-built lightweight convolutional network to obtain the predicted label for each time-frequency map. Step 3: Calculate the cross-entropy loss using the predicted and true labels from the time-frequency graph in Step 2. Step 4: Determine the direction of the decrease in cross-entropy loss using stochastic gradient descent, and update the parameters of the lightweight convolutional network according to this direction; Step 5: Repeat steps 2 to 4 until the maximum number of iterations is reached to obtain a trained lightweight convolutional network.

3. The spatial target behavior recognition method based on the time-frequency characteristics of radio signals according to claim 2, characterized in that, Following step 5, the spatial target behavior recognition method based on the time-frequency characteristics of radio signals further includes: Step 6: Take out all the samples in the test set in sequence and input them into the trained lightweight convolutional network to perform classification prediction and obtain the predicted label of the input sample; Step 7: Compare the predicted labels and real labels of the input samples from Step 6 to obtain the classification and recognition results; Step 8: Statistically analyze the recognition results of all input samples to obtain the classification accuracy of the trained lightweight convolutional network.

4. The spatial target behavior recognition method based on the time-frequency characteristics of radio signals according to claim 2, characterized in that, Step 1 includes: Step 11: Acquire multiple target transmission signals with BPSK modulation. Step 12: Use time-frequency analysis to perform time-frequency transformation on the transmitted signal to obtain multiple time-frequency maps containing the behavior of space targets; Step 13: Combine the multiple time-frequency images containing spatial target behavior into a dataset; Step 14: Divide the dataset into a training set and a test set in a 7:3 ratio.

5. The spatial target behavior recognition method based on the time-frequency characteristics of radio signals according to claim 1, characterized in that, The convolutional module with a kernel of 3 includes multiple input channels and multiple output channels. The convolutional module with a kernel of 3 includes a convolutional layer, a batch normalization layer, and a SiLu layer connected in sequence. The inverse residual module with a stride of 2 and the inverse residual module with a stride of 1 both include a convolutional layer, a batch normalization layer, and a SiLu layer connected in sequence, and a linear activation function is connected after the SiLu layer through pointwise convolution. The inverse residual module with a stride of 1 uses residual connections between the convolutional layer, the batch normalization layer, the SiLu layer, and the activation function connected in sequence.

6. The spatial target behavior recognition method based on the time-frequency characteristics of radio signals according to claim 1, characterized in that, The classification module includes a convolutional layer and a classifier. The convolutional layer convolves the fused features and then inputs them into the classifier. The classifier classifies the convolved fused features to obtain the predicted labels of spatial target behavior in the time-frequency map.

7. A spatial target behavior recognition device based on the time-frequency characteristics of radio signals, characterized in that, The space target behavior recognition device, which implements the space target behavior recognition method based on the time-frequency characteristics of radio signals according to any one of claims 1 to 6, comprises: The acquisition module is configured to acquire the transmission signal of the detected target and perform a time-frequency transformation on the transmission signal to obtain a time-frequency map containing the behavior of the space target; The recognition module is configured to input the time-frequency map acquired by the acquisition module into a trained lightweight convolutional network to extract features from the time-frequency map using the lightweight convolutional network. A local feature extraction module then projects these features into a high-dimensional space to obtain local features of the time-frequency map. These local features are then divided into multiple patches, and the relationships between each patch are encoded to obtain a corresponding global patch representation. The global patch representation is then folded to obtain a global time-frequency map representation. This global representation is then convolved point-by-point and merged with the features extracted by the lightweight convolutional network to obtain a fused feature. Finally, the fused feature is classified to obtain predicted labels for spatial target behaviors in the time-frequency map.

8. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-6.

Images