Multi-scene signal data-oriented multi-block mixed expert radio frequency fingerprint identification method

Abstract

The invention discloses a multi-scene signal data-oriented multi-block mixed expert radio frequency fingerprint identification method, and belongs to the field of software radio. According to the implementation method, a plurality of expert modules are adopted, and each expert module is designed to be used for extracting feature representations under different conditions so as to cope with signal changes under various complex scenes. A lightweight and efficient route selection module is adopted to dynamically distribute and input signals to the most matched expert module according to the dimensionalities such as statistical characteristics, frequency domain structures or time sequence characteristics of the input signals, or output results are weighted and fused among a plurality of experts, so that the pertinence and flexibility of feature extraction and recognition are improved. A cooperative mechanism between expert modules is optimized by adopting a joint training strategy, so that an individual recognition model can maintain the independent expression ability of experts, and the convergence speed and stability of the whole model are improved by sharing bottom-layer expression and high-layer decision-making. The method is suitable for the field of software radio, and high-precision radiation source individual identification is realized.

Description

Technical Field

[0001] The present invention relates to a radio frequency fingerprint recognition method of multiple-block hybrid experts for multi-scenario signal data, and belongs to the field of software radio. Background Art

[0002] As a key technology in electronic reconnaissance and RF signal analysis, the core goal of radiator individual identification is to uniquely identify the radiator device by capturing subtle features such as transient ripple, frequency deviation characteristics, and modulation fingerprints contained in the signal. RF fingerprint recognition technology achieves device identity authentication and security protection by extracting unique features generated by the wireless device hardware circuit during signal transmission (such as power amplifier nonlinearity, carrier frequency deviation, etc.). Traditional technical systems often use feature engineering methods based on manual experience and rely on manual feature extraction strategies such as time-frequency domain analysis. Although such methods are feasible in restricted scenarios, their feature characterization capabilities face significant degradation in complex signal environments in multiple scenarios (such as dynamic channel conditions, multi-device interference, and cross-protocol signal differences), and have the inherent defect of insufficient adaptability to multi-scenario signals. Therefore, improving the robustness of the recognition system in complex signal environments and its adaptability to complex signal environments in multiple scenarios have become the focus of technological evolution in this field.

[0003] In recent years, deep learning methods have achieved automated feature extraction and mapping from raw signals to device identification by building end-to-end learning mechanisms, providing a new paradigm for addressing the limitations of traditional methods. Existing research mainly focuses on achieving high-precision classification within a predefined device library. Some improved solutions, based on constellation diagrams, perform well in steady-state signals based on heat trace features and power spectrum features, but have difficulty coping with dynamic channel interference or cross-protocol signal differences. In addition, deep learning methods based on device-level simulation rely on accurate hardware damage modeling and cannot be generalized to actual variable environments. At the same time, while domain adversarial networks can alleviate the impact of channel differences, they do not consider the collaborative optimization of multiple tasks (such as simultaneous identification of protocols and devices) and multiple features (time-frequency domain, spatial domain). These deep learning methods still have difficulty coping with the multi-scenario collaboration requirements of complex signal environments in real-world scenarios, and they exhibit significant limitations in complex adversarial scenarios.

[0004] The core challenge facing current technological development stems from the lack of generalizable cognitive capabilities in complex, multi-scenario signal environments. Existing methods are mostly based on single-modal features (such as time or frequency domains), whereas multi-scenario signals require the fusion of time-frequency-space domain features. Current research focuses on improving the generalization and robustness of wireless signal processing models across diverse application scenarios. With the increasing complexity of tasks such as intelligent wireless communications, unmanned system perception, and emitter identification, traditional approaches that rely on fixed features and static models suffer from poor adaptability and rapid performance degradation in diverse environments, including those characterized by variable channel conditions, heterogeneous devices, and increased interference. On the one hand, research has attempted to enhance model robustness to environmental changes by introducing techniques such as data augmentation, domain adaptation, and domain generalization. For example, methods such as adversarial training, maximum mean difference (MMD), and CORAL achieve cross-domain adaptation through feature alignment or discriminator calibration after training in the source domain. However, these methods still suffer from insufficient adaptation accuracy and unstable generalization when faced with more complex scenarios such as channel drift, non-stationary emission characteristics, and mixed interference from multiple sources. On the other hand, some research has begun to focus on the application of methods such as small-sample learning, incremental learning, and meta-learning to signal recognition tasks, attempting to simulate models that can quickly adjust parameters or strategies when encountering unknown environments or devices. However, due to the strong nonlinearity and coupling of wireless signals in the frequency, time, and IQ domains, the transferability of these methods still faces certain technical bottlenecks. Furthermore, current testing in real-world scenarios across devices, regions, and time periods is relatively insufficient. As a result, while models may perform well on a specific dataset, they struggle to adapt to changes in complex real-world deployment scenarios. Therefore, multi-scenario signal adaptability remains a core challenge in the field of intelligent signal processing, and in-depth research is urgently needed on the stability of feature extraction, the generalization capability of discrimination strategies, and the adaptability of model structures. Summary of the Invention

[0005] In response to the key problems of existing radiation source individual identification algorithms in multi-scene complex signal environments, such as insufficient recognition sensitivity, limited generalization accuracy in multi-scene complex signal environments, and the inability of existing identification models to handle multi-scene complex signal environments, the purpose of the present invention is to provide a radio frequency fingerprint identification method based on multi-block hybrid experts for multi-scene signal data. By constructing a model framework based on multi-block hybrid experts and utilizing an expert routing selection mechanism to process multi-scene signal data separately, while strengthening the stability of the representation of some category features, the model's ability to identify individual radiation sources in multi-scenes is significantly improved, thereby realizing the dynamic evolution and accuracy leap of radiation source individual identification in complex signal environments.

[0006] The object of the present invention is achieved through the following technical solutions:

[0007] In order to solve the key problems of the existing technology in complex and changeable scenes, such as the decline in recognition accuracy, insufficient generalization ability and poor model adaptability, the present invention discloses a radio frequency fingerprint recognition method for multi-block hybrid experts for multi-scene signal data, designs a hybrid expert model architecture composed of multiple functionally heterogeneous expert sub-networks, and combines the gating mechanism (Gating Network) to realize dynamic expert selection and fusion of different input signals, thereby constructing an intelligent recognition framework with scene perception and adaptive learning capabilities. Specifically, the present invention introduces multiple expert modules in the model structure, each expert is designed to extract feature representations under different conditions to cope with signal changes under various complex scenarios (such as multipath fading, frequency deviation disturbance, interference superposition, etc.). In order to achieve scene recognition and expert matching of input signals, the system introduces a lightweight but efficient dynamic routing selection module based on path entropy minimization. According to the statistical characteristics, frequency domain structure or time series characteristics of the input signal, the input is dynamically allocated to the most matching expert module, or the output results are weighted and fused among multiple experts, thereby improving the pertinence and flexibility of feature extraction and recognition. In addition, the collaborative mechanism between the ice-melting expert modules is optimized through a joint training strategy, so that the model can maintain the independent expression ability of the experts while improving the convergence speed and stability of the overall model by sharing the underlying representation and high-level decision-making. Compared with the traditional single model structure, the method of the present invention has a stronger model capacity regulation capability, and can adaptively adjust the reasoning path according to the complexity of the task and the diversity of the input, effectively alleviating the problem of overfitting or underfitting. The present invention significantly improves the modeling and recognition accuracy of radio frequency fingerprint features under different scenario conditions (such as different receivers, frequency bands, channel states, equipment backgrounds, etc.), realizes rapid adaptation and stable recognition of diversified signals, and has good generalization ability. The present invention is applicable to multiple application fields such as communication supervision, spectrum situational awareness, radio security and confrontation, and is particularly suitable for the precise identification and continuous tracking of individual radiation sources in dynamic deployment, multi-source heterogeneity and information-constrained environments.

[0008] The present invention discloses a multi-block hybrid expert radio frequency fingerprint recognition method for multi-scenario signal data, comprising the following steps:

[0009] Step 1: Collect the transmitter signal and down-convert it into an IQ signal, obtain the multi-scenario characteristics of the IQ signal, and construct a training set and a test set;

[0010] The RF signal data emitted by the transmitter is collected and multiplied by the carrier with a phase difference of 90° to obtain the radiation source signal data. The in-phase and orthogonal IQ signals are obtained through a low-pass filter. The continuous M pairs of IQ signal points are the basic input units of the network, and their dimension is 2*M. The basic input units are divided into training sets and test sets.

[0011] Step 2: Based on the multi-scenario characteristics obtained in step 1, implement expert cluster grouping and dynamically allocate expert capacity;

[0012] First, based on the multi-scenario characteristics of IQ signals, an expert clustering strategy is designed. The total number of experts, G, is divided into K mutually exclusive subsets, K≥2, where K represents the type of multi-scenario data in the dataset.

[0013] Second, dynamically allocate expert capacity:

[0014] C j =Softmax(W·AvgPool(x))·B total (1)

[0015] Among them, C j represents the actual computing capacity allocation of the jth expert, x represents the feature representation of the current input sample, AvgPool(x) represents the average pooling of the input feature x, W represents the learnable weight matrix, Softmax(·) represents the normalization of all expert scores, and B total For the total computation budget, dynamic expert activation is achieved under resource constraints.

[0016] Step 3: Based on the expert cluster grouping achieved in Step 2, a hierarchical neural network based on multi-block hybrid experts is constructed; the hierarchical neural network optimizes expert selection through a routing optimization function, adjusts data flow paths through a dynamic routing algorithm, gradually activates expert modules through an ice-melting expert thawing strategy, and provides a unified feature representation through a common feature extractor;

[0017] By stacking multiple hybrid expert architecture modules, a hierarchical neural network based on multi-block hybrid experts is constructed, and a network architecture containing L basic hybrid expert modules is constructed; each basic hybrid expert module contains:

[0018] Shared feature extraction layer: a residual structure consisting of 1×1 convolution, 3×3 grouped convolution, and 1×1 convolution;

[0019] Independent expert layer: Based on the expert cluster grouping implemented in step 2, it contains G parallel experts, each of which is a 3×3 convolution, BN layer, and ReLU sequence with independent parameters;

[0020] Dynamic gating network: A learnable routing controller consisting of a global pooling layer, a fully connected layer, and a softmax layer.

[0021] Step 4: Based on the hierarchical neural network of multi-block hybrid experts, a routing optimization function based on path entropy minimization is designed;

[0022] The routing optimization function based on path entropy minimization described in step 4 for:

[0023]

[0024] in, represents the expectation of all input samples, p j is the probability of assigning the sample to expert j, G is the expert gradient matrix, ||G i -G j || F represents the Frobenius norm distance between two experts, and λ represents the trade-off coefficient.

[0025] Step 5: Design a dynamic routing algorithm based on a multi-block hybrid expert neural network;

[0026] The dynamic routing algorithm described in step 5 is:

[0027] ① Warm-up phase training: Hard routing constraint mechanism for the 1st to 10th training cycles

[0028] First, for the input data on day m, only the experts in the corresponding subset are activated;

[0029] Secondly, the gradient updates of non-relevant expert parameters are frozen through the gradient masking technique;

[0030] Finally, the gated network is allowed to learn the weight distribution within the full expert range to achieve collaborative training of the gated network;

[0031] ②Joint training phase: Soft routing release strategy for the 11th to 20th training cycles

[0032] First, we remove the expert gradient constraints, opening up the joint optimization of all expert parameters;

[0033] Secondly, a routing probability hybrid mechanism is introduced to retain the original routing strategy with probability u and enable full expert routing with probability (1-u). Finally, L2 regularization constraints are imposed on expert parameters to achieve elastic parameter constraints.

[0034] Step 6: Design an ice-melting expert thawing strategy based on a multi-block hybrid expert neural network;

[0035] The ice-melting expert thawing strategy described in step 6 is:

[0036] Strict isolation period: In the 1st to 5th training cycles, only the shared feature layer and routing network are trained, the expert parameters are frozen, and the set of network parameters θ allowed to be updated in the current training phase trainable for:

[0037] θ trainable ={θ shared ,θ route}(t<T1) (3)

[0038] Among them, θ shared represents the parameters of the shared feature extraction layer, θ route Represents the parameters of the routing network, t represents the current training round, and T1 represents the end round of the "strict isolation period";

[0039] Gradual thawing period: 6th to 10th training cycles are activated in batches according to the expert's contribution

[0040] First, the data is allowed to activate non-corresponding day experts with probability d. The activation condition is:

[0041]

[0042] Where N represents the total number of samples, represents the probability that sample i is assigned to the jth expert, represents the indicator function, and γ represents the threshold value used;

[0043] Secondly, according to formula (4), the distribution range of the gate weight is dynamically adjusted;

[0044] Complete joint phase: training cycles 11-20;

[0045] The annealing strategy is used to gradually reduce the routing constraint strength.

[0046] Step 7: Based on the multi-block hybrid expert neural network, a differentiable quantum encoder is constructed as a common feature extractor;

[0047] The common feature extractor in step 7 is a differentiable quantized encoder z shared , expressed as:

[0048]

[0049] Among them, B represents the quantized bit width, b represents the index of the current bit, and W b Represents the learnable mapping matrix corresponding to the b-th bit;

[0050] Secondly, for the shared features obtained by the common feature extractor, a cross-expert gradient gating strategy is designed to dynamically adjust the contribution ratio of the shared features to the expert input:

[0051]

[0052] Among them, g j represents the feature fusion weight of the j-th expert, σ(·) represents the nonlinear activation function, U represents the learnable linear transformation matrix, represents the private feature representation of the j-th expert;

[0053] Step 8. Based on the multi-block hybrid expert-based neural network obtained in step 3, design a manifold adaptive residual connection to achieve dynamic dimension matching. Use the training set obtained in step 1 to train the multi-block hybrid expert-based neural network, and use the trained neural network to implement radio frequency fingerprint recognition.

[0054] Based on the multi-block hybrid expert-based neural network obtained in step three, steps four, five, six, and seven are integrated. The training set obtained in step one is used to perform forward propagation and backpropagation processes in sequence until the network parameters converge. The neural network parameters that can be used for radio frequency fingerprint recognition are obtained to implement the training process. The test set is used to implement the testing process to obtain the optimal neural network.

[0055] The obtained neural network for RF fingerprint recognition is applied to signal processing scenarios, which can identify the RF fingerprint category of RF signals, realize energy-efficient RF fingerprint recognition tasks, and provide individual information of the radiation source for subsequent processing of RF signals.

[0056] Beneficial effects:

[0057] 1. This invention discloses a multi-block hybrid expert RF fingerprinting method for multi-scenario signal data. It constructs a high-quality, deep-learning-capable representation of IQ input signals, improving the expressiveness of signal features. By down-converting the original RF signal into an IQ signal and using M consecutive pairs of IQ samples as the basic input unit, this method effectively preserves the key time and frequency domain information in the transmitter's RF fingerprint. This low-dimensional but high-information signal representation provides a good data foundation for subsequent deep model training, enabling neural networks to more accurately learn and extract device-specific characteristics of signals.

[0058] 2. This paper discloses a multi-block hybrid expert RF fingerprinting method for multi-scenario signal data. By leveraging expert grouping strategies and a dynamic routing mechanism, it effectively addresses the distributional differences in RF fingerprint features across different receivers, acquisition days, and distances. This process significantly improves the model's convergence speed and generalization capabilities for multi-scenario RF fingerprinting tasks, providing a stable and reliable feature extractor for RF signal recognition.

[0059] 3. This invention discloses a multi-block hybrid expert RF fingerprinting method for multi-scenario signal data. It utilizes elastic parameter constraints (L2 regularization) in the joint training phase to maintain the stability of the original expert knowledge when the model adapts to new scenarios. It also utilizes hard routing constraints in the warm-up phase to reduce redundant parameter updates. Combined with a shared feature layer design, it promotes cross-environmental data knowledge transfer and enhances the model's robustness and generalization capabilities. This approach strengthens the model's focus on key features, reduces sensitivity to irrelevant disturbances, and improves the model's stability and recognition capabilities in multi-environmental data recognition.

[0060] 4. The present invention discloses a multi-block hybrid expert radio frequency fingerprint recognition method for multi-scenario signal data. Through hierarchical expert module design and dynamic expert routing mechanism, the model can flexibly add or replace expert modules to adapt to new signal scenarios without retraining the entire network structure. It has good scalability and transfer learning capabilities and is suitable for the ever-expanding wireless signal sources and scenario conditions in actual deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of the multi-block hybrid expert radio frequency fingerprint recognition method for multi-scenario signal data disclosed in the present invention;

[0062] Figure 2 This is the overall architecture diagram of the dynamic routing hybrid expert network disclosed in this embodiment;

[0063] Figure 3 This is a schematic diagram of the basic module structure disclosed in this embodiment;

[0064] Figure 4 This is a flow chart of the progressive training strategy disclosed in this embodiment;

[0065] Figure 5 This is a timing diagram of the working principle of the gate control network disclosed in this embodiment. DETAILED DESCRIPTION

[0066] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. The technical problems solved by the technical solution of the present invention and the beneficial effects thereof are also described. It should be noted that the described embodiments are only intended to facilitate understanding of the present invention and do not serve to limit the present invention in any way.

[0067] like Figure 1 As shown, the radio frequency fingerprint recognition method of multi-block hybrid experts for multi-scenario signal data disclosed in this embodiment is specifically implemented in the following steps:

[0068] Step 1: Use two sub-datasets (ManyTx and SingleDay) from the open-source WiSig dataset and the ORACLE dataset to implement individual radiator identification. The WiSig dataset is a large-scale WiFi dataset containing 10 million data packets transmitted from 174 off-the-shelf WiFi transmitters and received by 41 USRP receivers over four acquisitions within a month. The dataset includes both raw data capture and a small, convenient pre-processed subset. The ORACLE dataset is the most commonly used dataset in the field of RF fingerprinting. The dataset sizes and data partitioning used are shown in the following table:

[0069] name Multiple scene conditions Multi-scenario data Number of transmitters Number of receivers Collection days ManyTx Collection time Days 2, 3, and 4 150 18 4 SingleDay Collection equipment Receivers 0-6 28 10 1 ORACLE Transmission distance 52, 56, 62 ft 16 1 -

[0070] The ManySig and ManyRx preprocessing subsets of the WiSig dataset have a relatively small number of transmitters, making it difficult to demonstrate the advantages of the present invention. Therefore, this embodiment only uses the ManyTx and SingleDay subsets, which have a larger number of transmitters. The ManyTx and SingleDay subsets, as well as the ORACLE dataset, are multi-dimensional lists. In addition to the IQ sequence, they also include dimensions such as transmitter, receiver, acquisition day, and whether the data is balanced. Because RF fingerprinting must eliminate the influence of factors such as receiver noise, channel noise, and channel variation, these dimensions, including receiver, number of signal segments, acquisition day, and whether the data is balanced, are integrated before training. Each IQ sequence segment is 2×256 in size, with two channels representing real and imaginary information, respectively. The ManyTx, SingleDay, and ORACLE datasets use different multi-scenario data conditions. The ManyTx dataset uses acquisition days 2, 3, and 4, respectively, while the SingleDay dataset uses receivers 0-6 from different acquisition devices. The ORACLE dataset uses transmission distances of 52, 56, and 62 ft.

[0071] The dataset is divided according to the above table, and the dataset is split according to the ratio of 8:2, 80% for training set and the remaining 20% ​​for test set.

[0072] Step 2: Based on the multi-scenario characteristics obtained in step 1, implement expert cluster grouping and dynamically allocate expert capacity;

[0073] First, if Figure 1 、 Figure 2 As shown in the figure, according to the multi-scenario characteristics of IQ signals, an expert cluster grouping strategy is designed; the total number of experts G is divided into K mutually exclusive subsets K≥2, where K represents the type of multi-scenario data in the dataset;

[0074] Second, dynamically allocate expert capacity:

[0075] C j =Softmax(W·AvgPool(x))·B total (7)

[0076] Among them, C j represents the actual computing capacity allocation of the jth expert, x represents the feature representation of the current input sample, AvgPool(x) represents the average pooling of the input feature x, W represents the learnable weight matrix, Softmax(·) represents the normalization of all expert scores, and B total For the total computation budget, dynamic expert activation is achieved under resource constraints.

[0077] In this embodiment, for the ManyTx dataset, K is set to 3, corresponding to three days of data, and the expert grouping g is configured to be [11, 11, 10]. For the SingleDay dataset, K is set to 7, corresponding to seven receivers, and the expert grouping g is configured to be [5, 5, 5, 5, 5, 2]. For the ORACLE dataset, K is set to 3, corresponding to three different transmission distances, and the expert grouping g is configured to be [11, 11, 10].

[0078] Step 3: Based on the expert cluster grouping achieved in Step 2, a hierarchical neural network based on multi-block hybrid experts is constructed; the hierarchical neural network optimizes expert selection through a routing optimization function, adjusts data flow paths through a dynamic routing algorithm, gradually activates expert modules through an ice-melting expert thawing strategy, and provides a unified feature representation through a common feature extractor;

[0079] like Figure 3 As shown in FIG, a hierarchical neural network based on multi-block hybrid experts is constructed by stacking multi-block hybrid expert architecture modules, and a network architecture containing L basic hybrid expert modules is constructed; each basic hybrid expert module contains:

[0080] Shared feature extraction layer: a residual structure consisting of 1×1 convolution, 3×3 grouped convolution, and 1×1 convolution;

[0081] Independent expert layer: Based on the expert cluster grouping implemented in step 2, it contains G parallel experts, each of which is a 3×3 convolution, BN layer, and ReLU sequence with independent parameters;

[0082] Dynamic gating network: A learnable routing controller consisting of a global pooling layer, a fully connected layer, and a softmax layer.

[0083] Step 4: Figure 4 As shown, based on the hierarchical neural network of multi-block hybrid experts, a routing optimization function based on path entropy minimization is designed;

[0084] The routing optimization function based on path entropy minimization described in step 4 for:

[0085]

[0086] in, represents the expectation of all input samples, p j is the probability of assigning the sample to expert j, G is the expert gradient matrix, ||G i -G j || F represents the Frobenius norm distance between two experts, and λ represents the trade-off coefficient.

[0087] Step 5: Design a dynamic routing algorithm based on a multi-block hybrid expert neural network;

[0088] The dynamic routing algorithm described in step 5 is:

[0089] ① Warm-up phase training: Hard routing constraint mechanism for the 1st to 10th training cycles

[0090] First, for the input data on day m, only the experts in the corresponding subset are activated;

[0091] Secondly, the gradient updates of non-relevant expert parameters are frozen through the gradient masking technique;

[0092] Finally, the gated network is allowed to learn the weight distribution within the full expert range to achieve collaborative training of the gated network;

[0093] ②Joint training phase: Soft routing release strategy for the 11th to 20th training cycles

[0094] First, we remove the expert gradient constraints, opening up the joint optimization of all expert parameters;

[0095] Secondly, a routing probability hybrid mechanism is introduced to retain the original routing strategy with probability u and enable full expert routing with probability (1-u). Finally, L2 regularization constraints are imposed on expert parameters to achieve elastic parameter constraints.

[0096] Step 6: Design an ice-melting expert thawing strategy based on a multi-block hybrid expert neural network;

[0097] The ice-melting expert thawing strategy described in step 6 is:

[0098] Strict isolation period: In the 1st to 5th training cycles, only the shared feature layer and routing network are trained, the expert parameters are frozen, and the set of network parameters θ allowed to be updated in the current training phase trainable for:

[0099] θ trainable ={θ shared ,θ route}(t<T1) (9)

[0100] Among them, θ shared represents the parameters of the shared feature extraction layer, θ route Represents the parameters of the routing network, t represents the current training round, and T1 represents the end round of the "strict isolation period";

[0101] Gradual thawing period: 6th to 10th training cycles are activated in batches according to the expert's contribution

[0102] First, the data is allowed to activate non-corresponding day experts with probability d. The activation condition is:

[0103]

[0104] Where N represents the total number of samples, represents the probability that sample i is assigned to the jth expert, represents the indicator function, and γ represents the threshold value used;

[0105] Secondly, if Figure 5 As shown, the gating weight distribution range is dynamically adjusted;

[0106] Complete joint phase: training cycles 11-20;

[0107] Adopt annealing strategy to gradually reduce the strength of routing constraints;

[0108] Step 7: Based on the multi-block hybrid expert neural network, a differentiable quantum encoder is designed as a common feature extractor;

[0109] The common feature extractor in step 7 is a differentiable quantized encoder z shared , expressed as:

[0110]

[0111] Among them, B represents the quantized bit width, b represents the index of the current bit, and W b Represents the learnable mapping matrix corresponding to the b-th bit;

[0112] Secondly, for the shared features obtained by the common feature extractor, a cross-expert gradient gating strategy is designed to dynamically adjust the contribution ratio of the shared features to the expert input:

[0113]

[0114] Among them, g j represents the feature fusion weight of the j-th expert, σ(·) represents the nonlinear activation function, U represents the learnable linear transformation matrix, represents the private feature representation of the j-th expert;

[0115] Step 8: Based on the multi-block hybrid expert-based neural network obtained in step 3, steps 4, 5, 6, and 7 are integrated, the network is trained using the training set, and tested using the test set, thereby realizing radio frequency fingerprint recognition.

[0116] Based on the multi-block hybrid expert-based neural network obtained in step three, steps four, five, six, and seven are integrated. The training set obtained in step one is used to perform forward propagation and backpropagation processes in sequence until the network parameters converge. The neural network parameters that can be used for radio frequency fingerprint recognition are obtained to implement the training process. The test set is used to implement the testing process to obtain the optimal neural network.

[0117] The obtained neural network for RF fingerprint recognition is applied to signal processing scenarios, which can identify the RF fingerprint category of RF signals, realize energy-efficient RF fingerprint recognition tasks, and provide individual information of the radiation source for subsequent processing of RF signals.

[0118] Training a multi-block hybrid expert-based neural network involves sequentially performing forward and backward propagation based on the basic input units obtained in step 1 until the network parameters converge. This results in neural network parameters that can be used to identify individual radiation sources, completing the training process. This example uses the Adam optimizer with a learning rate of 0.01, a batch size of 256, 20 iterations, and a cross-entropy loss function. The RF fingerprint recognition neural network model is trained using gradient descent.

[0119] The present invention is compared with the ResNeXt-50 network architecture and the hybrid expert network architecture. The ResNeXt-50 network architecture, the hybrid expert network architecture and the present invention use the same network parameters for training and complete the training process through cross entropy loss.

[0120] Table 1 Effects of the present invention and other comparative methods

[0121]

[0122] As shown in the table above, the effect of the present invention is greatly improved on the test set.

[0123] In summary, the radio frequency fingerprint recognition method of multi-block hybrid experts for multi-scenario signal data disclosed in this embodiment first introduces multiple expert modules in the model structure, and each expert is designed to extract feature representations under different conditions to cope with signal changes under various complex scenarios (such as multipath fading, frequency deviation disturbance, interference superposition, etc.). Then, in order to realize scene recognition and expert matching of the input signal, the system introduces a lightweight but efficient routing selection module, which dynamically allocates the input to the most matching expert module according to the statistical characteristics, frequency domain structure or time series characteristics of the input signal, or weightedly fuses its output results among multiple experts, thereby improving the pertinence and flexibility of feature extraction and recognition. Finally, the collaboration mechanism between expert modules is optimized through a joint training strategy, so that the model can improve the convergence speed and stability of the overall model by sharing the underlying representation and high-level decision-making while maintaining the independent expression ability of the experts. Compared with the traditional single model structure, the method of the present invention has a stronger model capacity regulation capability, can adaptively adjust the reasoning path according to the complexity of the task and the diversity of the input, and effectively alleviate the problem of overfitting or underfitting. The core innovation of this invention lies in significantly improving the modeling and recognition accuracy of RF fingerprint features under different scenario conditions (such as different receivers, frequency bands, channel states, device backgrounds, etc.), achieving rapid adaptation and stable recognition of diversified signals, and having good generalization ability and practical promotion value.

[0124] The above specific description further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-block hybrid expert radio frequency fingerprint recognition method for multi-scenario signal data, characterized by: The following steps are included: Step 1: Collect the transmitter signal and down-convert it into IQ signal to construct the training set and test set; Step 2: Based on the characteristics of the data collected in step 1, implement expert cluster grouping and dynamically allocate expert capacity; Step 3: Based on the expert cluster grouping achieved in step 2, a neural network based on multi-block hybrid experts is constructed to build a hierarchical expert network; Step 4: Based on the multi-block hybrid expert-based neural network obtained in step 3, a routing optimization function based on path entropy minimization is constructed; Step 5: Based on the multi-block hybrid expert-based neural network obtained in step 3, a dynamic routing algorithm is constructed; Step 6: Based on the multi-block hybrid expert-based neural network obtained in step 3, an ice-melting expert thawing strategy is constructed; Step 7: Based on the multi-block hybrid expert-based neural network obtained in step 3, a quantized shared feature layer is constructed as a common feature extractor; Step 8. Based on the multi-block hybrid expert-based neural network obtained in step 3, design a manifold adaptive residual connection to achieve dynamic dimension matching, use the training set obtained in step 1 to train the multi-block hybrid expert-based neural network, and use the trained neural network to achieve radio frequency fingerprint recognition.

2. The multi-block hybrid expert RF fingerprint recognition method for multi-scene signal data according to claim 1, characterized in that: The implementation method of step one is: The RF signal data emitted by the transmitter is collected and multiplied by the carrier with a phase difference of 90° to obtain the radiation source signal data. The in-phase and orthogonal IQ signals are obtained through a low-pass filter. The continuous M pairs of IQ signal points are the basic input units of the network, and their dimension is 2*M. The basic input units are divided into training sets and test sets.

3. The multi-block hybrid expert RF fingerprint recognition method for multi-scene signal data according to claim 2, characterized in that: The implementation method of step 2 is: Based on the multi-scenario characteristics of the data, an expert cluster grouping strategy is constructed; the total number of experts, G, is divided into K mutually exclusive subsets K≥2, where K represents the type of multi-scenario data in the dataset, including different receivers, different collection days, and different distances between transmitters and receivers; Design dynamic expert capacity allocation: C j =Softmax(W·AvgPool(x))·B total (1) Among them, C j represents the actual computing capacity allocation of the jth expert, x represents the feature representation of the current input sample, AvgPool(x) represents the average pooling of the input feature x, W represents the learnable weight matrix, Softmax(·) represents the normalization of all expert scores, and B total For the total computation budget, dynamic expert activation is achieved under resource constraints.

4. The multi-block hybrid expert RF fingerprint recognition method for multi-scene signal data according to claim 3, characterized in that: The implementation method of step three is: By stacking multiple hybrid expert architecture modules, a neural network based on multi-block hybrid experts is constructed. A network architecture consisting of L basic hybrid expert modules is constructed, and a hierarchical expert network is designed. Each basic module contains: Shared feature extraction layer: a residual structure consisting of 1×1 convolution, 3×3 grouped convolution, and 1×1 convolution; Independent expert layer: Based on the expert cluster grouping implemented in step 2, it contains G parallel experts, each of which is a 3×3 convolution, BN layer, and ReLU sequence with independent parameters; Dynamic gating network: a learnable routing controller consisting of a global pooling layer, a fully connected layer, and a softmax layer; A neural network based on multi-block hybrid experts is obtained.

5. The multi-block hybrid expert RF fingerprint recognition method for multi-scene signal data according to claim 4, characterized in that: The implementation method of step four is: Based on the multi-block hybrid expert-based neural network obtained in step 3, the routing optimization function based on path entropy minimization is designed through routing distribution entropy control and expert similarity constraints. in, represents the expectation of all input samples, p j is the probability of assigning the sample to expert j, G is the expert gradient matrix, ||G i -G j || F represents the Frobenius norm distance between two experts, and λ represents the trade-off coefficient.

6. The multi-block hybrid expert RF fingerprint recognition method for multi-scene signal data according to claim 5, characterized in that: The implementation method of step five is: Based on the multi-block hybrid expert-based neural network obtained in step 3, a dynamic routing algorithm is constructed: ① Warm-up phase training: hard routing constraint mechanism for the 1st to 10th training cycles; First, for the input data on day m, only the experts in the corresponding subset are activated; Secondly, the gradient updates of non-relevant expert parameters are frozen through the gradient masking technique; Finally, the gated network is allowed to learn the weight distribution within the full expert range to achieve collaborative training of the gated network; ②Joint training phase: Soft routing release strategy for the 11th to 20th training cycles First, we remove the expert gradient constraints, opening up the joint optimization of all expert parameters; Secondly, a routing probability hybrid mechanism is introduced to retain the original routing strategy with probability u and enable full expert routing with probability (1-u). Finally, L2 regularization constraints are imposed on expert parameters to achieve elastic parameter constraints.

7. The multi-block hybrid expert RF fingerprint recognition method for multi-scene signal data according to claim 6, characterized in that: The implementation method of step six is: Based on the multi-block hybrid expert neural network obtained in step 3, an ice-melting expert thawing strategy is designed: Strict isolation period: In the 1st to 5th training cycles, only the shared feature layer and routing network are trained, the expert parameters are frozen, and the set of network parameters θ allowed to be updated in the current training phase trainable for: i trainable ={θ shared ,i route }(t <T1) (3) Among them, θ shared represents the parameters of the shared feature extraction layer, θ route Represents the parameters of the routing network, t represents the current training round, and T1 represents the end round of the "strict isolation period"; Gradual thawing period: 6th to 10th training cycles are activated in batches according to the expert's contribution With probability d, data is allowed to activate non-corresponding day experts. The activation conditions are: Where N represents the total number of samples, represents the probability that sample i is assigned to the jth expert, represents the indicator function, and γ represents the threshold value used; According to formula (4), the gating weight distribution range is dynamically adjusted; Complete joint phase: training cycles 11-20; The annealing strategy is used to gradually reduce the routing constraint strength.

8. The multi-block hybrid expert radio frequency fingerprint recognition method for multi-scene signal data according to claim 7, characterized in that: The implementation method of step seven is: Based on the multi-block hybrid expert neural network obtained in step 3, a shared convolutional layer is cascaded before each basic hybrid expert module. All experts share the underlying feature extractor and design a differentiable quantum encoder z shared for: Among them, B represents the quantized bit width, b represents the index of the current bit, and W b Represents the learnable mapping matrix corresponding to the b-th bit; Secondly, we design cross-expert gradient gating to dynamically adjust the contribution ratio of shared features to expert input: Among them, g j represents the feature fusion weight of the j-th expert, σ(·) represents the nonlinear activation function, U represents the learnable linear transformation matrix, represents the private feature representation of the j-th expert.

9. The multi-block hybrid expert radio frequency fingerprint recognition method for multi-scene signal data according to claim 8, characterized in that: The implementation method of step eight is: Based on the multi-block hybrid expert neural network obtained in step 3, a curvature-aware residual scaling is designed to adaptively adjust the residual strength according to the local curvature of the feature manifold. The residual signal adjustment coefficient α of the j-th expert path is j for: Where z represents the feature representation, Curv(z) represents the local curvature estimate of z, and κ represents the preset curvature sensitivity coefficient; Using the training set obtained in step 1, forward propagation and back propagation processes are performed in sequence until the network parameters converge, thereby obtaining the neural network parameters that can be used for radio frequency fingerprint recognition, thus completing the training process; The obtained neural network for RF fingerprint recognition is applied to signal processing scenarios, which can identify the RF fingerprint category of RF signals, realize energy-efficient RF fingerprint recognition tasks, and provide individual information of the radiation source for subsequent processing of RF signals.

Images