Time-frequency domain joint multi-dimensional migration optimization radio frequency signal identification method, device and equipment

Abstract

The present application relates to a time-frequency domain joint multi-dimensional migration optimization radio frequency signal identification method, device and equipment, innovatively realize the collaborative optimization of time-frequency characteristics through time sequence constraint and frequency domain feature reconstruction, overcome the limitation of traditional method processing time domain or frequency domain characteristics alone. Among them, the time sequence constraint effectively maintains the dynamic characteristics of the signal through the modeling of time sequence correlation, the frequency domain feature reconstruction adopts multi-scale analysis to enhance the anti-interference ability of the key frequency band, and the dynamic joint optimization adaptively adjusts the migration weight of the time-frequency feature, realizes the automatic balance of the distribution difference between the source domain and the target domain. The collaborative work of the three core modules makes the above method show stronger generalization performance in the cross-device and cross-environment radio frequency signal identification task.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency signal processing technology, and relates to a time-frequency domain joint multidimensional migration optimization method, apparatus and equipment for radio frequency signal identification. Background Technology

[0002] With the rapid development of the Industrial Internet, radio frequency (RF) signal identification technology is ushering in significant development opportunities. The widespread application of devices such as Wi-Fi, RFID, and millimeter-wave radar generates massive amounts of high-dimensional RF data, characterized by strong time-frequency coupling and rapid dynamic changes. In practical applications across devices and scenarios, traditional signal processing methods struggle to effectively address data distribution shifts caused by device heterogeneity, environmental noise, and multipath effects. This has made intelligent signal processing technology based on transfer learning a research hotspot, as it leverages existing knowledge to solve signal identification problems in new scenarios, potentially overcoming the limitations of traditional methods.

[0003] However, current radio frequency signal transfer learning technology still faces many challenges. First, signals generated by different radio frequency devices exhibit significant differences in the time and frequency domains, making it difficult for existing methods to effectively model the complex mapping relationships of these high-dimensional features. Second, dynamically changing channel conditions and sudden interference render traditional static transfer models insufficiently adaptable. Furthermore, unsupervised domain adaptation methods are susceptible to interference from low-quality samples and lack reliable mechanisms for assessing transfer feasibility. In addition, the high computational complexity of joint time-frequency analysis contradicts the real-time requirements of industrial applications, and the lack of a standardized evaluation system often leads to problems such as large fluctuations in recognition accuracy and the need for manual parameter tuning in practical deployments, severely hindering the practical application of intelligent signal processing technology based on transfer learning. Summary of the Invention

[0004] To address the problems existing in the aforementioned traditional technologies, this invention proposes a time-frequency domain joint multidimensional transfer optimization method for radio frequency signal recognition and a time-frequency domain joint multidimensional transfer optimization radio frequency signal recognition device. These methods can demonstrate stronger generalization performance in radio frequency signal recognition tasks across devices and environments, and promote the practical application of intelligent signal processing technology based on transfer learning.

[0005] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0006] On the one hand, a time-frequency domain joint multidimensional migration optimization method for radio frequency signal identification is provided, including the following steps:

[0007] Acquire the radio frequency signal to be identified received by the radio frequency signal identification system;

[0008] The radio frequency signal to be identified is input into the trained improved XCIT network for feature extraction, and the output is the time domain features and frequency features of the radio frequency signal to be identified. The input layer of the improved XCIT network is extended to a dual-channel parallel convolutional structure and is equipped with a cross-channel feature fusion module. The trained improved XCIT network is obtained by using pre-collected unlabeled signal data and training according to the designed overall loss function, which includes time feature aggregation loss, time feature reconstruction loss and frequency feature reconstruction loss.

[0009] The radio frequency signal type of the radio frequency signal to be identified is determined by utilizing the time-domain and frequency characteristics of the radio frequency signal to be identified.

[0010] In one embodiment, the training process for the improved XCIT network includes the following steps:

[0011] Tagless signal data collection is performed using a general-purpose software radio peripheral; key parameters are configured in the general-purpose software radio peripheral using open-source software radio software, including setting the carrier frequency, selecting the modulation method, determining the signal bandwidth and transmit power.

[0012] After converting the unlabeled signal data into IQ dual-channel complex form, it is input into the improved XCIT network for feature extraction to obtain time-domain features and frequency features.

[0013] The temporal features are input into a multilayer perceptron-based temporal decoder for decoding, and then the temporal feature reconstruction loss is calculated. The temporal decoder consists of a linear projection layer, a GELU activation function, and a symmetric linear dimensionality reduction layer connected in sequence.

[0014] The frequency features are input into a multilayer perceptron-based frequency decoder for decoding to obtain reconstructed frequency data. The frequency feature reconstruction loss is calculated after performing an inverse Fourier transform on the reconstructed frequency data. The frequency decoder adopts a multilayer perceptron structure that is symmetrical to the time decoder.

[0015] The temporal feature aggregation loss of time-domain features is calculated using the ranking loss function; the ranking loss function is a loss function based on the ranking constraint mechanism of the time dimension.

[0016] After aggregating the losses based on temporal feature reconstruction loss, frequency feature reconstruction loss, and temporal feature aggregation loss, the AdamW optimization algorithm is used to update the model.

[0017] When the training iterations reach convergence or the set stopping condition, the optimized model parameters are saved to obtain the trained improved XCIT network.

[0018] In one embodiment, during the collection of tagless signal data using a general-purpose software radio peripheral, the power amplifier nonlinearity is compensated by digital predistortion.

[0019] In one embodiment, the overall loss function is:

[0020]

[0021] in, and These are the weights, and the time feature reconstruction loss is:

[0022]

[0023] The frequency feature reconstruction loss is:

[0024]

[0025] The time feature aggregation loss is:

[0026]

[0027] Among them, the target variable t Set to -1, spacing margin margin Setting it to 1, max indicates taking the maximum value. D intra The average distance between sequences. D inter This represents the average distance within the sequence.

[0028] On the other hand, a time-frequency domain joint multidimensional migration optimization radio frequency signal identification device is also provided, comprising:

[0029] The signal acquisition module is used to acquire the radio frequency signal to be identified received by the radio frequency signal identification system;

[0030] The feature extraction module is used to input the radio frequency signal to be identified into the trained improved XCIT network for feature extraction and output the time domain features and frequency features of the radio frequency signal to be identified. The input layer of the improved XCIT network is extended to a dual-channel parallel convolutional structure and is equipped with a cross-channel feature fusion module. The trained improved XCIT network is obtained by using pre-collected unlabeled signal data and training according to the designed overall loss function, which includes time feature aggregation loss, time feature reconstruction loss and frequency feature reconstruction loss.

[0031] The identification output module is used to determine the radio frequency signal type of the radio frequency signal to be identified by utilizing the time domain characteristics and frequency characteristics of the radio frequency signal to be identified.

[0032] In another aspect, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned time-frequency domain joint multidimensional migration optimization radio frequency signal identification method.

[0033] One of the above technical solutions has the following advantages and beneficial effects:

[0034] The aforementioned time-frequency domain joint multidimensional transfer optimization method, device, and equipment for radio frequency signal identification innovatively achieves synergistic optimization of time-frequency features through time-ordering constraints and frequency-domain feature reconstruction, overcoming the limitations of traditional methods that process time-domain or frequency-domain features separately. Specifically, the time-ordering constraints effectively preserve the dynamic characteristics of the signal through time-series correlation modeling; the frequency-domain feature reconstruction employs multi-scale analysis to enhance the anti-interference capability of key frequency bands; and the dynamic joint optimization adaptively adjusts the transfer weights of time-frequency features, achieving automatic balancing of the differences in the distribution between the source and target domains. This collaborative work of the three core modules enables the method to exhibit stronger generalization performance in radio frequency signal identification tasks across devices and environments. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating a time-frequency domain joint multidimensional migration optimization method for radio frequency signal identification in one embodiment.

[0037] Figure 2 This is a schematic diagram of the flowchart design framework for a time-frequency domain joint multidimensional migration optimization method for radio frequency signal identification in one embodiment;

[0038] Figure 3 This is a schematic diagram of the multidimensional transfer optimization framework of a time-frequency domain joint multidimensional transfer optimization radio frequency signal recognition method in one embodiment;

[0039] Figure 4 This is a block diagram of a time-frequency domain joint multidimensional migration optimization radio frequency signal identification device in one embodiment. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0041] It should be noted that, in this document, the reference to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination of one or more of the associated listed items, and all possible combinations, including such combinations.

[0042] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0043] To address the limitations hindering the practical application of transfer learning-based intelligent signal processing technologies, researchers have explored various approaches. In feature representation learning, a dynamic feature adaptation mechanism based on joint time-frequency constraints has been proposed, enhancing model transferability by decoupling time-domain correlation and frequency-domain physical characteristics. Regarding model architecture, lightweight network designs and model compression techniques have been developed to balance computational efficiency and transfer performance. In algorithm optimization, a progressive transfer strategy combining meta-learning and online learning has been explored to achieve dynamic adaptive adjustment of parameters. Furthermore, attempts have been made to construct robust evaluation metrics that integrate prior signal knowledge, providing more reliable guidance for the transfer process. These explorations offer potential pathways to overcome existing technological bottlenecks.

[0044] This invention addresses the feature distribution shift problem caused by device heterogeneity, environmental noise, and multipath effects in cross-scenario applications. It proposes a multi-dimensional transfer optimization scheme based on joint time-frequency domain constraints. By constructing a collaborative optimization mechanism of temporal correlation constraints and frequency domain feature reconstruction, it achieves cross-domain alignment of signal feature distribution. This overcomes the feature distribution shift problem caused by device heterogeneity and dynamic environmental changes in existing RF signal recognition methods in cross-scenario applications, as well as the shortcomings of traditional transfer learning methods in joint time-frequency feature optimization. It significantly improves the model's generalization ability under unlabeled RF data and is suitable for the real-time signal recognition needs of high-dynamic, low signal-to-noise ratio scenarios in the Industrial Internet.

[0045] In one embodiment, such as Figure 1 As shown, a time-frequency domain joint multidimensional migration optimization method for radio frequency signal identification is provided, which may include the following steps S12 to S16:

[0046] S12, acquire the radio frequency signal to be identified received by the radio frequency signal identification system;

[0047] S14, the radio frequency signal to be identified is input into the trained improved XCIT network for feature extraction, and the time domain features and frequency features of the radio frequency signal to be identified are output; wherein, the input layer of the improved XCIT network is extended to a dual-channel parallel convolutional structure and is equipped with a cross-channel feature fusion module. The trained improved XCIT network is obtained by using pre-collected unlabeled signal data and training according to the designed overall loss function, which includes time feature aggregation loss, time feature reconstruction loss and frequency feature reconstruction loss;

[0048] S16. The radio frequency signal type of the radio frequency signal to be identified is determined by utilizing the time domain characteristics and frequency characteristics of the radio frequency signal to be identified.

[0049] Understandable, such as Figure 2 As shown, through three core improvements—time ordering constraints, frequency domain feature reconstruction, and dynamic joint optimization—an end-to-end adaptive optimization framework was constructed, which significantly improved the model's generalization ability and recognition accuracy across devices and environments, providing a more reliable solution for intelligent radio frequency signal processing.

[0050] The aforementioned time-frequency domain joint multidimensional migration optimization method for RFID signal recognition significantly improves the reliability and adaptability of RFID systems in complex industrial IoT environments through the collaborative work of three core modules: time sequencing constraints, frequency domain feature reconstruction, and dynamic joint optimization. In the time dimension, the innovative time-series feature optimization mechanism enables the system to automatically adapt to signal timing differences caused by different devices and environments, such as signal delay variations caused by equipment movement or electromagnetic interference in a factory workshop. In the frequency domain dimension, intelligent frequency domain feature reconstruction technology effectively overcomes the inconsistency in frequency response characteristics caused by hardware differences between devices from different manufacturers, which is particularly important for hybrid networks composed of multi-brand devices. Finally, the dynamic joint optimization mechanism ensures that the system can automatically adjust according to the actual scenario requirements, guaranteeing recognition accuracy in stable environments while rapidly adapting to sudden interference.

[0051] Compared to traditional technologies, the above methods ultimately enable RFID systems to maintain stable performance under conditions such as equipment upgrades and environmental changes, significantly reducing the cost of system maintenance and retraining. This provides reliable technical support for the intelligent transformation of industrial sites and offers more stable and economical innovative solutions to the challenges faced by RFID in actual industrial scenarios such as equipment status monitoring and asset tracking in logistics and warehousing, demonstrating significant market application value.

[0052] In one embodiment, the training process for an improved XCIT network includes the following steps:

[0053] Tagless signal data collection is performed using a general-purpose software radio peripheral; key parameters are configured in the general-purpose software radio peripheral using open-source software radio software, including setting the carrier frequency, selecting the modulation method, determining the signal bandwidth and transmit power.

[0054] After converting the unlabeled signal data into IQ dual-channel complex form, it is input into the improved XCIT network for feature extraction to obtain time-domain features and frequency features.

[0055] The temporal features are input into a multilayer perceptron-based temporal decoder for decoding, and then the temporal feature reconstruction loss is calculated. The temporal decoder consists of a linear projection layer, a GELU activation function, and a symmetric linear dimensionality reduction layer connected in sequence.

[0056] The frequency features are input into a multilayer perceptron-based frequency decoder for decoding to obtain reconstructed frequency data. The frequency feature reconstruction loss is calculated after performing an inverse Fourier transform on the reconstructed frequency data. The frequency decoder adopts a multilayer perceptron structure that is symmetrical to the time decoder.

[0057] The temporal feature aggregation loss of time-domain features is calculated using the ranking loss function; the ranking loss function is a loss function based on the ranking constraint mechanism of the time dimension.

[0058] After aggregating the losses based on temporal feature reconstruction loss, frequency feature reconstruction loss, and temporal feature aggregation loss, the AdamW optimization algorithm is used to update the model.

[0059] When the training iterations reach convergence or the set stopping condition, the optimized model parameters are saved to obtain the trained improved XCIT network.

[0060] Understandably, the first step is to collect unlabeled signal data. Specifically, this is achieved by configuring key parameters in the open-source software-defined radio software GNU Radio. These parameters can include setting the carrier frequency, selecting the modulation scheme (such as AM, FM, or QPSK digital modulation), determining the signal bandwidth, and determining the transmit power. The baseband signal is then combined with the carrier wave by a modulator to convert it into a high-frequency signal suitable for wireless transmission. This can be achieved, for example, by changing the carrier amplitude through amplitude modulation or by using OFDM (Orthogonal Frequency Division Multiplexing) technology to distribute the data across multiple subcarriers to improve interference immunity.

[0061] The modulated high-frequency signal is transmitted via the USRP (Universal Software Radio Peripheral) hardware platform. After the transmission flowchart is written in GNU Radio and executable code is generated, the code is burned into the USRP device. The USRP's FPGA chip handles real-time digital signal processing, including performing digital up-conversion to shift the baseband signal to the target RF frequency, interpolation to improve the sampling rate, and optionally, adding digital predistortion to compensate for power amplifier nonlinearity. Finally, the signal is amplified by a power amplifier, noise is suppressed by a bandpass filter, and finally radiated into space by the antenna.

[0062] Transmitted signals are affected by various channel characteristics during space transmission. Multipath effects cause signal delay spread due to propagation along different paths, relative motion between the transmitter and receiver can cause Doppler shift, and thermal noise and adjacent channel interference in the channel can also affect signal quality. Cognitive radio systems use dynamic spectrum sensing technology to detect idle frequency bands in real time and employ adaptive modulation and coding to adjust transmission parameters according to the current channel conditions.

[0063] After the receiving antenna captures electromagnetic signals in space, they undergo preliminary processing via a low-noise amplifier and bandpass filter. The USRP hardware then comes into play; its FPGA chip performs digital down-conversion to convert the RF signal back to baseband and performs decimation to reduce the data volume. The received signal is demodulated using GNU Radio software to recover the original information, while simultaneously displaying signal waveforms and spectrum characteristics in real time. Key quality indicators such as signal-to-noise ratio and bit error rate are calculated to obtain unlabeled signal data, providing a basis for system decision-making.

[0064] Next comes model feature extraction. Specifically, the real-valued time-domain signal acquired by GNU Radio software (i.e., the unlabeled signal data collected above) is first converted into an IQ dual-channel complex form, fully preserving the amplitude and phase information of the signal. The traditional XCIT (Cross-Covariance Image Transformer) network is modified by expanding the input layer into a dual-channel parallel convolutional structure and designing a cross-channel feature fusion module to obtain an improved XCIT network capable of simultaneously processing both I (In-phase) and Q (Quadrature) signals. The cross-channel feature fusion module is structured as follows: the feature maps of the I and Q channels are mapped to a unified dimension through fully connected layers, and then a cross-attention mechanism is used to calculate the correlation weights of the two channel features, achieving dynamic interaction and feature fusion. The fused feature is added to the original feature through residual connections, and nonlinear processing and regularization are performed using a multilayer perceptron (MLP) and a layer normalization layer to output the final fused feature. Finally, the input signal data is used to extract time-domain features.

[0065] The time-domain signal is subjected to a Fast Fourier Transform to obtain the frequency domain, which is then input into an improved XCIT network for processing. The network employs a bandpass attention mechanism to automatically focus on the effective frequency band and extracts frequency features such as the dominant frequency component and harmonic distribution through multi-resolution convolution.

[0066] Next, the temporal reconstruction loss, frequency domain reconstruction loss, and ranking constraint loss are calculated;

[0067] Specifically, to achieve high-precision reconstruction of the original signal from time-domain features, this step designs a time decoder based on a multilayer perceptron (MLP). This time decoder adopts a classic encoding / decoding architecture of "dimensionality increase-nonlinear transformation-dimensionality reduction": first, the input features are mapped to a high-dimensional latent space through a learnable linear projection layer; then, a nonlinear feature enhancement is performed using the GELU (Gaussian Error Linear Unit) activation function; finally, the time-domain signal is reconstructed through a symmetric linear dimensionality reduction layer. To improve the reconstruction quality, the L1 norm is used as the loss function, which is more robust to outliers than the L2 norm. The mathematical expression of the L1 norm is:

[0068]

[0069] in, N This represents the number of signal sampling batches. x i For the first i Batch timing signal data. For the first i Batch reconstruction of time data. The specific processing procedure is as follows:

[0070] When performing time-series reconstruction based on time-domain features, the reconstructed time data can be obtained through a time decoder. For example... Figure 3 As shown, the time decoder is mainly constructed using a multilayer perceptron, as illustrated in the following formula:

[0071]

[0072] in, h 1 represents the high-dimensional feature representation of the output of the first linear layer. h 2 represents the nonlinear features processed by the activation function. W 1 and W 2 are the weight matrices of the first and second linear layers, respectively, which can be optimized through neural network training. b 1 and b 2 represents the bias vector of the corresponding linear network layer. For temporal activation functions, ReLU activation is typically used to enhance nonlinear expressiveness. The first linear layer expands the feature dimension to obtain a high-dimensional feature representation. This step helps the model capture linear relationships between different time points and provides the necessary feature representation for subsequent nonlinear processing. Next, by introducing activation functions, the model can reveal complex patterns and dependencies in the input data. Finally, the last linear layer is responsible for reducing the feature dimension, remapping the nonlinearly transformed feature space back to the original data dimension, thereby achieving effective data reconstruction.

[0073] Time domain features The aforementioned multilayer perceptron is used to decode the reconstructed temporal data. The temporal feature reconstruction loss is calculated using the mean squared error loss function defined by the following formula:

[0074]

[0075] Considering the complementary nature of the time and frequency domains, a dual-domain joint optimization framework was also designed. The frequency decoder employs an MLP structure symmetrical to the time domain, but an inverse (short-time) Fourier transform is introduced before loss calculation to convert the frequency domain reconstructed features back to the time domain: that is, the complex form of the frequency domain features is first reconstructed through the MLP network. The inverse short-time Fourier transform converts the features to the time domain. Then, the L1 norm is also used as the loss function to calculate the reconstruction loss in the frequency domain.

[0076] The specific process is as follows: The radio frequency signal transmitted from the transmitter mainly consists of a modulation signal. S ( t impulse response of wireless channels H (t The signal is composed of several components, and different frequency modulation methods can have a significant impact on the signal. Modulation signal S ( t It mainly utilizes frequency variations to transmit binary bit streams on different transmission channels. s b ( t The information is frequency shift keying (FSK).

[0077]

[0078] in, A c For carrier amplitude, For frequency modulation factor, This is the time delay. By analyzing the above expression for frequency shift keying, we can see that how to interpret the information carried by a signal can be transformed into how to analyze information in the frequency domain. The information in the time domain is transformed into information in the frequency domain using a Fast Fourier Transform. After acquiring the discrete complex signal, the discrete spectrum of the signal is further calculated using the following formula:

[0079]

[0080] in, For Fast Fourier Transform (FFT) w The frequencies are discrete. To reconstruct the time-domain information from the frequency domain features, a frequency decoder and inverse Fourier transform are used. The frequency decoder is defined similarly to that used in time-series feature reconstruction, as shown below:

[0081]

[0082] in, and These are parameters that can be trained using a neural network. and For the bias of the linear network layer, For frequency domain activation functions, For reconstructed frequency data.

[0083] After obtaining the reconstructed frequency data, in order to ensure that it can be accurately aligned with the original time-series signal data, an inverse Fourier transform is performed on the reconstructed data. The specific operation is shown in the following formula:

[0084]

[0085] in, This is the inverse Fourier transform. The reconstructed frequency data is converted back to the time domain information. Finally, the frequency feature reconstruction loss is calculated using the mean squared error loss function defined by the following formula:

[0086]

[0087] To fully leverage the potential value of temporal information in transfer learning, this embodiment innovatively draws upon the importance ranking mechanism in recommender systems, designing a time-dimensional ranking constraint mechanism. This mechanism introduces structured prior knowledge from the time dimension to guide the model in learning the differences in importance of data at different time points and their inherent correlations. Specifically, the ranking constraint mechanism introduces ranking loss functions for both intra-sequence and inter-sequence processing:

[0088]

[0089] in, t Let `max` be the target variable, indicating that the maximum value should be taken. d 1 represents the distance between positive and negative sample pairs (such as the Euclidean distance between semantically similar feature vectors in a time series or the similarity score predicted by the model). d 2 represents the distance between negative sample pairs (such as the distance between feature vectors with significant semantic differences). Positive and negative sample pairs originate from feature representations extracted from the time domain or fused across channels. margin To establish a margin, the distance difference between positive and negative sample pairs must be at least [a certain value]. margin This is to prevent the model from "lazily" learning subtle differences. The ranking constraint mechanism helps the model identify the feature patterns of key time nodes in the time series and distinguish the information contribution of different time segments, thereby establishing a more robust time feature representation and significantly improving the model's generalization performance in cross-task and cross-dataset scenarios.

[0090] Specifically, the ranking constraint mechanism uses the target variable t Controlling the distance between positive and negative sample pairs d 1 and d 2. When t When =1, the goal is to construct d 2> d A loss of 1 means that the distance between negative sample pairs is greater than the distance between positive sample pairs; when t When =-1, the goal is to construct d 2< d A loss of 1. Introducing a margin of interval. margin Adjust the distance between positive and negative sample pairs in the sorting task to enhance the model's discriminative ability.

[0091] More specifically, the original signal is input into the base model to obtain time-domain features. ,in bs It is the batch size of the experiment. patches It represents the number of segments in each time sample. featureIt is a feature dimension extracted from the PCCP (Predetermined Change Control Plan) framework, which uses time-domain features and distance calculation methods to construct a distance matrix. Finally, the average distance within and between sequences is calculated using the distance matrix, and the ranking loss is calculated using the ranking loss function described above. Here, the distance matrix is ​​the Euclidean distance. The calculated features and specific distance matrix of the sample can be represented as follows:

[0092]

[0093] Then, a mask matrix is ​​constructed to extract the average distance between sequences. D intra and the average distance within the sequence D inter It can be expressed by the following formula:

[0094]

[0095] in, Mask distence Is and M distence Mask matrices of the same shape have all zero values ​​on their diagonal. It is a batch i The Middle k Inner distance of sequences Indicates batch i The Middle m The distance vector between the sequence and all other sequences. They are the first m The minimum and maximum distances between sequences are used to normalize these distance values. The average distance between sequences is calculated. D intra and the average distance within the sequence D inter Inserting the formula into the above formula for the ranking loss function yields the final ranking loss function (also known as the temporal feature aggregation loss):

[0096]

[0097] in, t Set to -1, margin Set to 1.

[0098] This leads to the aggregation of the three loss functions mentioned above. Time-dimensional feature aggregation, by comprehensively considering data from different time points, forces optimization of the correlation between time features, thereby effectively capturing the dynamic changes of time series and providing richer contextual information. Simultaneously, time-dimensional feature reconstruction, by recombining time features, increases the model's understanding of time dependencies, preventing the model's understanding of time features from deviating from the overall distribution, and further improving classification accuracy. In the frequency dimension, the above feature reconstruction focuses on analyzing and optimizing frequency-related features, recovering the original information in the frequency domain through inverse Fourier transform, improving the expression of frequency features, and enhancing the model's adaptability to signal frequency characteristics.

[0099] Finally, to comprehensively optimize the model's transfer performance, the overall loss function takes into account the three optimization objectives mentioned above, adding the temporal feature aggregation loss, temporal feature reconstruction loss, and frequency feature reconstruction loss to form the final overall loss function as shown below:

[0100]

[0101] in, and These are the weights of each loss function.

[0102] Finally, the model is updated.

[0103] It is understandable that the model calculates the loss function using the backpropagation algorithm. The gradient is calculated for each parameter. Specifically, starting from the output layer, the chain rule is applied layer by layer to backpropagate the error signal to each layer, and the partial derivatives of the weights and biases are calculated. The gradient value reflects the direction and magnitude of the change in the loss function when the parameters change slightly.

[0104] The AdamW optimization algorithm is used to calculate parameter updates. AdamW is an improved version of the Adam optimizer, primarily optimizing the calculation method of weight decay (L2 regularization). Building upon Adam, it first calculates the exponential moving average of the gradient (momentum) and the exponential moving average of the squared gradient (adaptive learning rate adjustment), and then corrects for bias. Unlike Adam, AdamW no longer mixes weight decay into gradient calculation; instead, it directly applies it to parameter updates, making regularization more independent of the optimization process. This decoupling approach more stably controls model complexity and avoids training bias caused by the mutual interference between adaptive learning rate and L2 regularization in Adam, thereby improving the model's generalization performance.

[0105] Iterate until convergence or a preset stopping condition is met: The model will repeatedly execute forward propagation, loss calculation, backpropagation, and parameter updates until convergence or a preset stopping condition is met, such as loss convergence, validation set performance reaching the target, reaching the maximum number of training epochs, or manual termination of training. Finally, the model saves the optimized model parameters, thus completing the entire training process.

[0106] It should be understood that, although the above process Figure 1 The steps in the diagram are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed; they can be performed in other orders. Furthermore, the above process... Figure 1 At least some of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0107] In one embodiment, such as Figure 4 As shown, a time-frequency domain joint multidimensional transfer optimization radio frequency signal identification device 100 is also provided, including a signal acquisition module 11, a feature extraction module 13, and an identification output module 15. The signal acquisition module 11 is used to acquire the radio frequency signal to be identified received by the radio frequency signal identification system. The feature extraction module 13 is used to input the radio frequency signal to be identified into a trained improved XCIT network for feature extraction, and output the time-domain features and frequency features of the radio frequency signal to be identified. The input layer of the improved XCIT network is extended to a dual-channel parallel convolutional structure and includes a cross-channel feature fusion module. The trained improved XCIT network is obtained by using pre-collected unlabeled signal data and training according to a designed overall loss function, which includes time feature aggregation loss, time feature reconstruction loss, and frequency feature reconstruction loss. The identification output module 15 is used to determine the radio frequency signal type of the radio frequency signal to be identified using the time-domain features and frequency features of the radio frequency signal to be identified.

[0108] The aforementioned time-frequency domain joint multidimensional transfer optimization radio frequency signal recognition device 100 innovatively achieves synergistic optimization of time-frequency features through time-ordering constraints and frequency-domain feature reconstruction, overcoming the limitations of traditional methods that process time-domain or frequency-domain features separately. Specifically, the time-ordering constraints effectively preserve the dynamic characteristics of the signal through time-series correlation modeling; the frequency-domain feature reconstruction employs multi-scale analysis to enhance the anti-interference capability of key frequency bands; and the dynamic joint optimization adaptively adjusts the transfer weights of time-frequency features, achieving automatic balancing of the differences in the distribution between the source and target domains. This collaborative work of the three core modules enables the method to exhibit stronger generalization performance in radio frequency signal recognition tasks across devices and environments.

[0109] In one embodiment, the training process for improving the XCIT network includes:

[0110] Tagless signal data collection is performed using a general-purpose software radio peripheral; key parameters are configured in the general-purpose software radio peripheral using open-source software radio software, including setting the carrier frequency, selecting the modulation method, determining the signal bandwidth and transmit power.

[0111] After converting the unlabeled signal data into IQ dual-channel complex form, it is input into the improved XCIT network for feature extraction to obtain time-domain features and frequency features.

[0112] The temporal features are input into a multilayer perceptron-based temporal decoder for decoding, and then the temporal feature reconstruction loss is calculated. The temporal decoder consists of a linear projection layer, a GELU activation function, and a symmetric linear dimensionality reduction layer connected in sequence.

[0113] The frequency features are input into a multilayer perceptron-based frequency decoder for decoding to obtain reconstructed frequency data. The frequency feature reconstruction loss is calculated after performing an inverse Fourier transform on the reconstructed frequency data. The frequency decoder adopts a multilayer perceptron structure that is symmetrical to the time decoder.

[0114] The temporal feature aggregation loss of time-domain features is calculated using the ranking loss function; the ranking loss function is a loss function based on the ranking constraint mechanism of the time dimension.

[0115] After aggregating the losses based on temporal feature reconstruction loss, frequency feature reconstruction loss, and temporal feature aggregation loss, the AdamW optimization algorithm is used to update the model.

[0116] When the training iterations reach convergence or the set stopping condition, the optimized model parameters are saved to obtain the trained improved XCIT network.

[0117] In one embodiment, during the collection of tagless signal data using a general-purpose software radio peripheral, digital predistortion is used to compensate for the nonlinearity of the power amplifier.

[0118] In one embodiment, the overall loss function is:

[0119]

[0120] in, and These are the weights, and the time feature reconstruction loss is:

[0121]

[0122] The frequency feature reconstruction loss is:

[0123]

[0124] The time feature aggregation loss is:

[0125]

[0126] Among them, the target variable t Set to -1, spacing margin margin Setting it to 1, max indicates taking the maximum value. D intra The average distance between sequences. D inter This represents the average distance within the sequence.

[0127] It is understood that the explanations of the features in the above-mentioned time-frequency domain joint multidimensional migration optimization radio frequency signal recognition device 100 can be understood by referring to the corresponding explanations in the various embodiments of the above-mentioned time-frequency domain joint multidimensional migration optimization radio frequency signal recognition method.

[0128] Each module in the aforementioned time-frequency domain joint multidimensional migration optimization radio frequency signal identification device 100 can be implemented entirely or partially through software, hardware, or a combination thereof. These components can be embedded in hardware or independently of a device with data processing capabilities, or stored in software within the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of computers already existing in the art.

[0129] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the following processing steps: acquiring a radio frequency signal to be identified received by a radio frequency signal identification system; inputting the radio frequency signal to be identified into a trained improved XCIT network for feature extraction, and outputting time-domain features and frequency features of the radio frequency signal to be identified; wherein the input layer of the improved XCIT network is extended to a dual-channel parallel convolutional structure and is provided with a cross-channel feature fusion module; the trained improved XCIT network is obtained by using pre-collected unlabeled signal data and training according to a designed overall loss function, the overall loss function including time feature aggregation loss, time feature reconstruction loss, and frequency feature reconstruction loss; and using the time-domain features and frequency features of the radio frequency signal to be identified, determining the radio frequency signal type of the radio frequency signal to be identified.

[0130] In one embodiment, when the processor executes the computer program, it can also implement the steps or sub-steps added in the various embodiments of the above-described time-frequency domain joint multidimensional migration optimization radio frequency signal identification method.

[0131] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), memory bus DRAM (RDRAM), and interface DRAM (DRDRAM), etc.

[0132] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0133] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of protection of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and all such modifications and improvements fall within the scope of protection of the present invention.

Claims

1. A time-frequency domain joint multidimensional migration optimization method for radio frequency signal identification, characterized in that, Including the following steps: Acquire the radio frequency signal to be identified received by the radio frequency signal identification system; The radio frequency signal to be identified is input into the trained improved XCIT network for feature extraction, and the output is the time domain features and frequency features of the radio frequency signal to be identified. The input layer of the improved XCIT network is extended to a dual-channel parallel convolutional structure and is equipped with a cross-channel feature fusion module. The trained improved XCIT network is obtained by using pre-collected unlabeled signal data and training according to the designed overall loss function, which includes time feature aggregation loss, time feature reconstruction loss and frequency feature reconstruction loss. The type of radio frequency signal to be identified is determined by utilizing the time-domain and frequency characteristics of the radio frequency signal to be identified. The training process for improving the XCIT network includes the following steps: Tagless signal data collection is performed using a general-purpose software radio peripheral; key parameters are configured in the general-purpose software radio peripheral using open-source software radio software, including setting the carrier frequency, selecting the modulation method, determining the signal bandwidth and transmit power. After converting the unlabeled signal data into IQ dual-channel complex form, it is input into the improved XCIT network for feature extraction to obtain time-domain features and frequency features. The temporal features are input into a multilayer perceptron-based temporal decoder for decoding, and then the temporal feature reconstruction loss is calculated. The temporal decoder consists of a linear projection layer, a GELU activation function, and a symmetric linear dimensionality reduction layer connected in sequence. The frequency features are input into a multilayer perceptron-based frequency decoder for decoding to obtain reconstructed frequency data. The frequency feature reconstruction loss is calculated after performing an inverse Fourier transform on the reconstructed frequency data. The frequency decoder adopts a multilayer perceptron structure that is symmetrical to the time decoder. The temporal feature aggregation loss of time-domain features is calculated using the ranking loss function; the ranking loss function is a loss function based on the ranking constraint mechanism of the time dimension. After aggregating the losses based on temporal feature reconstruction loss, frequency feature reconstruction loss, and temporal feature aggregation loss, the AdamW optimization algorithm is used to update the model. When the training iterations reach convergence or the set stopping condition, the optimized model parameters are saved to obtain the trained improved XCIT network.

2. The time-frequency domain joint multidimensional migration optimization radio frequency signal identification method according to claim 1, characterized in that, In the process of collecting tagless signal data using general software radio peripherals, digital predistortion is used to compensate for the nonlinearity of the power amplifier.

3. The time-frequency domain joint multidimensional migration optimization radio frequency signal identification method according to claim 1 or 2, characterized in that, The overall loss function is: in, and These are the weights, and the time feature reconstruction loss is: The frequency feature reconstruction loss is: The time feature aggregation loss is: Among them, the target variable t Set to -1, spacing margin margin Setting it to 1, max indicates taking the maximum value. D intra The average distance between sequences. D inter This represents the average distance within the sequence.

4. A time-frequency domain joint multidimensional migration optimization radio frequency signal identification device, characterized in that, include: The signal acquisition module is used to acquire the radio frequency signal to be identified received by the radio frequency signal identification system; The feature extraction module is used to input the radio frequency signal to be identified into the trained improved XCIT network for feature extraction and output the time domain features and frequency features of the radio frequency signal to be identified. The input layer of the improved XCIT network is extended to a dual-channel parallel convolutional structure and is equipped with a cross-channel feature fusion module. The trained improved XCIT network is obtained by using pre-collected unlabeled signal data and training according to the designed overall loss function, which includes time feature aggregation loss, time feature reconstruction loss and frequency feature reconstruction loss. The identification output module is used to determine the radio frequency signal type of the radio frequency signal to be identified by utilizing the time domain characteristics and frequency characteristics of the radio frequency signal to be identified; The improved training process for the XCIT network includes: Tagless signal data collection is performed using a general-purpose software radio peripheral; key parameters are configured in the general-purpose software radio peripheral using open-source software radio software, including setting the carrier frequency, selecting the modulation method, determining the signal bandwidth and transmit power. After converting the unlabeled signal data into IQ dual-channel complex form, it is input into the improved XCIT network for feature extraction to obtain time-domain features and frequency features. The temporal features are input into a multilayer perceptron-based temporal decoder for decoding, and then the temporal feature reconstruction loss is calculated. The temporal decoder consists of a linear projection layer, a GELU activation function, and a symmetric linear dimensionality reduction layer connected in sequence. The frequency features are input into a multilayer perceptron-based frequency decoder for decoding to obtain reconstructed frequency data. The frequency feature reconstruction loss is calculated after performing an inverse Fourier transform on the reconstructed frequency data. The frequency decoder adopts a multilayer perceptron structure that is symmetrical to the time decoder. The temporal feature aggregation loss of time-domain features is calculated using the ranking loss function; the ranking loss function is a loss function based on the ranking constraint mechanism of the time dimension. After aggregating the losses based on temporal feature reconstruction loss, frequency feature reconstruction loss, and temporal feature aggregation loss, the AdamW optimization algorithm is used to update the model. When the training iterations reach convergence or the set stopping condition, the optimized model parameters are saved to obtain the trained improved XCIT network.

5. The time-frequency domain joint multidimensional migration optimization radio frequency signal identification device according to claim 4, characterized in that, In the process of collecting tagless signal data using general software radio peripherals, digital predistortion is used to compensate for the nonlinearity of the power amplifier.

6. The time-frequency domain joint multidimensional migration optimization radio frequency signal identification device according to claim 4 or 5, characterized in that, The overall loss function is: in, and These are the weights, and the time feature reconstruction loss is: The frequency feature reconstruction loss is: The time feature aggregation loss is: Among them, the target variable t Set to -1, spacing margin margin Setting it to 1, max indicates taking the maximum value. D intra The average distance between sequences. D inter This represents the average distance within the sequence.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the time-frequency domain joint multidimensional migration optimization radio frequency signal identification method according to any one of claims 1 to 3.

Images