Open set signal identification method and system based on hybrid features and channel awareness

Abstract

The present application relates to the technical field of radio signal identification, and particularly relates to an open set signal identification method and system based on mixed features and channel perception, comprising the following steps: obtaining a baseband complex signal sequence, performing channel estimation to obtain a channel state parameter; extracting a physical layer invariant feature from the baseband complex signal sequence to form a physical feature vector, inputting the baseband complex signal sequence into a pre-trained deep embedding network to extract a deep embedding feature vector; inputting the channel state parameter into a gated fusion network to generate a fusion weight of the physical feature and the deep embedding feature, weighting and fusing the physical feature vector and the deep embedding feature vector according to the fusion weight to obtain a mixed feature vector; calculating the matching degree of the mixed feature vector with each category in a pre-constructed known signal space, and dynamically adjusting the category decision threshold of the signal according to the channel state parameter to output an identification result.

Description

Technical Field

[0001] This invention relates to the field of radio signal identification technology, and more specifically to an open set signal identification method and system based on hybrid features and channel awareness. Background Technology

[0002] Radio spectrum monitoring is a crucial component of spectrum management, and one of its core tasks is the automatic identification and classification of radio signals within the monitored frequency bands. With the rapid development of wireless communication technologies, new communication protocols and modulation methods are constantly emerging, such as LoRa, NB-IoT, SigFox, and 5G NR, making the spectrum environment increasingly complex. In spectrum monitoring practice, monitoring stations frequently collect unknown signal types not included in the training data. These unknown signals may originate from new communication devices, illegal transmitters, or temporarily deployed radio systems.

[0003] Traditional signal recognition methods are typically based on the closed-set assumption, meaning the class of a test sample must be within the training class set. These methods include traditional approaches combining feature extraction and classifiers, as well as methods based on deep neural networks. Traditional feature extraction methods extract physical features such as higher-order cumulants and cyclostationary features from the signal, then use classifiers like support vector machines (SVMs) and K-nearest neighbors (KNNs) for recognition. Deep learning-based methods, on the other hand, directly utilize convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to perform end-to-end feature learning and classification on the signal's I / Q time-domain data or time-frequency plots.

[0004] However, the aforementioned methods, when encountering unknown signals not present in the training set, will forcibly classify the unknown signals into a known category, leading to serious false alarms. To address this problem, open set recognition techniques have been proposed, aiming to detect and reject samples of unknown categories during the recognition process. Existing open set recognition methods are mainly designed for the computer vision domain, such as OpenMax, generative model-based rejection methods, and distance-based open set methods. These methods have achieved good results in image recognition tasks, but directly transferring them to the field of radio signal recognition has the following shortcomings: First, radio signals are affected by the channel during transmission, including multipath fading, Doppler shift, and noise interference, which causes the characteristic distribution of the signal to change drastically with the channel conditions. However, open set methods in the CV field do not consider this characteristic, and their performance degrades significantly when the channel conditions change.

[0005] Second, radio signals have unique physical characteristics, such as cyclostationarity and higher-order statistical properties. These characteristics are relatively invariant to channel changes, but existing open-set methods have not fully explored and utilized these physical features to construct more robust known signal boundaries.

[0006] Third, the category decision threshold of existing methods is usually fixed and cannot be adaptively adjusted according to the current channel quality, resulting in an excessively high false rejection rate when the channel conditions are poor, or an excessively high false negative rate when the channel conditions are good. Summary of the Invention

[0007] The present invention aims to solve the problem of how to accurately classify known types of radio signals while reliably detecting unknown signal types not included in the training data in the context of spectrum monitoring, and to overcome the impact of channel changes on open set recognition performance.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following solution: The open set signal identification method based on hybrid features and channel awareness includes the following steps: S1: Obtain the baseband complex signal sequence, perform channel estimation, and obtain the channel state parameters; S2: Extract physical layer invariant features from the baseband complex signal sequence to form a physical feature vector. Input the baseband complex signal sequence into a pre-trained deep embedding network to extract the deep embedding feature vector. S3: Input the channel state parameters into the gated fusion network to generate fusion weights for physical features and deep embedded features. Then, perform weighted fusion of the physical feature vector and the deep embedded feature vector according to the fusion weights to obtain the hybrid feature vector. S4: Calculate the matching degree between the hybrid feature vector and each category in the pre-constructed known signal space, and dynamically adjust the category decision threshold of the signal according to the channel state parameters. If the matching degree of all categories of the signal is lower than the category decision threshold, it is determined to be an unknown signal; otherwise, the signal is identified as the known signal category with the highest matching degree, and the identification result is output.

[0009] This method achieves highly robust open set recognition of unknown signals under dynamic channel conditions through the above steps: In S3, instead of using fixed feature splicing or fusion based solely on feature internal attention, the channel state parameters are directly input into the gated fusion network to dynamically generate fusion weights for physically invariant features and deeply embedded features. This design, based on external conditions controlling feature fusion weights according to channel state, adaptively increases the weights of more robust physical features when channel quality deteriorates (e.g., reduced signal-to-noise ratio or increased fading), and increases the weights of more discriminative deeply embedded features when channel conditions are favorable. This overcomes the shortcomings of existing feature fusion methods that cannot adapt to drastic changes in the external channel environment, ensuring good separability and stability of the hybrid feature vector under different signal-to-noise ratios. In S4, after calculating the matching degree between the sample and the known signal space, the category decision threshold is dynamically adjusted based on the channel state parameters obtained in S1. Because this method changes the traditional fixed threshold or threshold setting method based solely on feature space statistical distribution (e.g., extreme value theory) in the open set recognition decision stage, and instead introduces channel quality in real time as the basis for threshold contraction or expansion, the decision boundary can dynamically scale with channel conditions. This technique directly overcomes the risk of false rejection of known class samples due to a fixed threshold at low signal-to-noise ratios (SNR), and the risk of missed detection of unknown class samples due to a fixed threshold at high SNR. It achieves a dynamic balance between sensitivity and specificity in detecting unknown signals in complex wireless transmission environments. In S2, by extracting physically invariant features (naturally robust to channel fading) and pre-trained deep embedding features (highly discriminative of subtle signal differences), and through the adaptive fusion and dynamic thresholding in S3 and S4, this method can reliably guarantee accurate classification of known classes and effectively reject interference signals of unknown classes even in complex non-cooperative communication scenarios such as signal received power fluctuations, multipath fading, and dynamic changes in SNR. This significantly improves the adaptability and reliability of the open set identification method for radio signals in real electromagnetic environments.

[0010] Furthermore, in S2, the physical layer invariant features include cyclic spectral density features, higher-order cumulant features, and time-frequency ridge features.

[0011] Furthermore, the method for extracting the cyclic spectral density features is as follows: calculate the cyclic spectral density function of the baseband complex signal sequence, and extract the peak position, peak amplitude, and profile entropy on the cyclic frequency-spectral frequency plane as cyclic spectral density features.

[0012] Furthermore, higher-order cumulants include second-order cumulants C. 20 C 21 Fourth-order cumulant C 40 C 41 C 42 Sixth-order cumulant C60 C 63 and the eighth-order cumulant C 80 At least one combination of the following.

[0013] Furthermore, in S3, the gated fusion network takes the channel state parameters as input, passes through at least one fully connected layer and a Softmax activation function layer, and outputs physical feature weights ω and deep embedding features β, where ω+β=1 and ω,β∈[0,1]; The weighted fusion calculation method is as follows: f mix = ω·f phy + β·f deep ; Where f phy f is the physical feature vector. deep For deep embedding feature vectors, f mix It is a mixed feature vector.

[0014] Furthermore, the channel state parameters include at least the signal-to-noise ratio estimate and the channel fading factor. The gated fusion network increases the weight of physical features when the signal-to-noise ratio is lower than a preset threshold, and increases the weight of deep embedding features when the signal-to-noise ratio is higher than a preset threshold.

[0015] Furthermore, the category decision threshold T is dynamically adjusted based on the channel state parameters: T = T0· f(γ, δ); Where T0 is the baseline threshold and f(γ, δ) is the channel adaptive adjustment function. When the signal-to-noise ratio γ decreases or the channel fading factor δ increases, f(γ, δ) is decreased to reduce the category decision threshold. When the signal-to-noise ratio γ increases or the channel fading factor δ decreases, f(γ, δ) is increased to increase the category decision threshold.

[0016] Furthermore, it also includes S5: performing feature clustering and storage on samples determined to be unknown signals. When the number of clustered samples of a certain type of unknown signal exceeds a preset threshold, an incremental update process is triggered to add the unknown signal of that type to the known signal space.

[0017] Furthermore, the incremental update process includes: manually labeling and confirming the clustering of unknown signals that trigger the update; after confirmation, extracting the mixed features of the signal and constructing its single-class classification boundary, and adding it to the known signal space; keeping the original known class boundaries unchanged during the addition process.

[0018] Another aspect of the present invention provides an open set signal identification system based on hybrid features and channel awareness, comprising: The channel estimation module is used to acquire the baseband complex signal sequence, perform channel estimation, and obtain channel state parameters; The physical feature and deep embedding feature extraction module is used to extract physical layer invariant features from the baseband complex signal sequence to form a physical feature vector. The baseband complex signal sequence is input into a pre-trained deep embedding network to extract the deep embedding feature vector. The gated fusion module is used to input channel state parameters into the gated fusion network, generate fusion weights for physical features and deep embedded features, and perform weighted fusion of physical feature vectors and deep embedded feature vectors according to the fusion weights to obtain a hybrid feature vector. The open set identification module is used to calculate the matching degree between the hybrid feature vector and each category in the pre-constructed known signal space, and dynamically adjust the category decision threshold of the signal according to the channel state parameters. If the matching degree of all categories of the signal is lower than the category decision threshold, it is determined to be an unknown signal; otherwise, the signal is identified as the known signal category with the highest matching degree, and the identification result is output. The incremental learning module, connected to the open set recognition module, is used to perform feature clustering and storage on samples identified as unknown signals. When the number of clustered samples of a certain type of unknown signal exceeds a preset threshold, the incremental update process is triggered to add the new signal category to the known signal space.

[0019] The beneficial effects of this invention are: First, by constructing a hybrid feature space that combines physical features and deep embedded features, the channel robustness of signal physical features and the high discriminative power of deep features are fully utilized, thereby improving the open set recognition performance while ensuring the classification accuracy of known signals.

[0020] Second, a channel-aware gating fusion mechanism is introduced, which adaptively adjusts the fusion weights of physical features and deep features according to the current channel state. When the channel conditions are poor, the weight of physical features is increased to maintain robustness, and when the channel conditions are good, the weight of deep features is increased to improve discriminability.

[0021] Third, a channel-aware adaptive category decision threshold is adopted, which dynamically adjusts the threshold size according to the signal-to-noise ratio and the channel fading factor. This solves the problem of poor adaptability of fixed thresholds under different channel conditions and achieves a balance between sensitivity and specificity in the detection of unknown signals.

[0022] Fourth, through the incremental learning mechanism, it can automatically accumulate and identify newly emerging unknown signal types, enabling the system to continuously learn and adapt to new signal environments, thus avoiding the cost of frequent retraining. Attached Figure Description

[0023] Figure 1 This is a flowchart of the open set signal recognition method based on hybrid features and channel awareness of the present invention; Figure 2 This is a schematic diagram of the hybrid feature extraction and gating fusion of the present invention; Figure 3 This is a diagram showing the ablation experiment category identification results in an embodiment of the present invention; Figure 4 This is a graph showing the comparison of detection rates under different channel models in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0026] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0027] Furthermore, for clarity and brevity, descriptions of well-known structures, functions, and configurations may have been omitted. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of this disclosure.

[0028] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0029] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0030] The present invention will now be described in detail with reference to the accompanying drawings and embodiments: Example 1 This embodiment provides an open set signal identification method based on hybrid features and channel awareness, such as... Figure 1 As shown, it includes the following steps: S1: Obtain the baseband complex signal sequence, perform channel estimation, and obtain the channel state parameters: Signal acquisition is performed in the target frequency band using a spectrum monitoring receiver. The receiver amplifies the received RF signal with low noise, down-converts it, filters it, and performs analog-to-digital conversion to obtain a digital intermediate frequency (IF) signal. The IF signal is then digitally down-converted to obtain a baseband complex signal sequence r(n) = I(n) + J·Q(n), where n is the sampling point index and J is the imaginary unit. The baseband complex signal sequence is then frame-synchronized and normalized to obtain a normalized baseband complex signal sequence x(n) of length N, where n = 0, 1, ..., N-1. The signal length N ranges from 256 to 4096, determined based on the signal bandwidth and sampling rate.

[0031] Channel estimation is performed on the normalized baseband complex signal sequence to obtain channel state parameters. These parameters include at least the signal-to-noise ratio (SNR) estimate γ and the channel fading factor δ. The SNR estimation employs the second-moment-fourth-moment M2M4 method. M2 = E[|x| 2 M4 = E[|x|] 4 ]; σ w 2 = (M2- √(2M2 2 - M4)) / 2; γ = (M2- σ w 2 ) / σ w 2 ; Where E[·] represents the expectation operation, σ w 2 This is the noise variance estimate.

[0032] The channel fading factor δ assesses the severity of multipath fading by calculating the normalized variance of the signal envelope. δ = Var(|x|) / E[|x|] 2 ; S2: Extract physical layer invariant features from the baseband complex signal sequence to form a physical feature vector. Input the baseband complex signal sequence into a deep embedding network pre-trained by contrastive learning to extract the deep embedding feature vector. Specifically, physical layer features that are invariant to channel variations are extracted from the normalized baseband complex signal sequence to form a physical feature vector f. phy .

[0033] Physical layer invariant characteristics include the following three categories: (1) Cyclic spectral density characteristics: Calculate the cyclic spectral density function S of the baseband complex signal sequence. x α(f), where α is the cyclic frequency and f is the spectral frequency. The cyclic spectral density function reflects the cyclic stationary characteristics of the signal; signals with different modulation methods exhibit different spectral peak patterns on the cyclic frequency-spectral frequency plane. The peak positions (α) of the cyclic spectral density function are extracted. i ,f i Peak amplitude |S x α (f)| and the normalized value, along with the cyclic spectral density profile entropy, are used as cyclic spectral density features to form the feature vector f. csd .

[0034] (2) Characteristics of higher-order cumulants: Calculate the second-order cumulant C of the baseband complex signal sequence. 20 C 21 Fourth-order cumulant C 40 C 41 C 42 Sixth-order cumulant C 60 C 63 and the eighth-order cumulant C 80 These cumulants reflect the probability distribution shape of the signal, and different modulation schemes have different theoretical values ​​for cumulants. Higher-order cumulant characteristics exhibit good robustness against Gaussian noise and channel fading. These cumulants and their combined ratios (such as |C|) are then considered. 40 | / |C 42 | 2 |C 60 | / |C 42 | 3 (etc.) constitute the feature vector f hoc .

[0035] (3) Time-frequency ridge features: The baseband complex signal sequence is subjected to short-time Fourier transform to obtain the time-frequency spectrum. The energy ridges in the time-frequency spectrum are extracted, which are the trajectories formed by the frequency points with the highest energy at each moment. The instantaneous frequency mean, variance, and rate of change statistics of the time-frequency ridges are calculated as the time-frequency ridge features, forming the feature sub-vector f. tf .

[0036] By concatenating the above three types of feature vectors, we obtain the physical feature vector: f phy = [f csd , f hoc , f tf ] .

[0037] The normalized baseband complex signal sequence x(n) is input into a deep embedding network pre-trained by contrastive learning to extract the deep embedding feature vector f. deep .

[0038] The deep embedding network is pre-trained using a contrastive learning approach based on the SimCLR framework. The network consists of an encoder g(·) and a projector h(·), with the encoder employing either a one-dimensional residual network ResNet-1D or a temporal convolutional network TCN as its backbone. During pre-training, two different augmented views are generated for the same signal (e.g., adding noise of varying intensities, random cropping, time shifting, etc.). The feature consistency between the different views of the same signal is maximized through a contrastive learning loss function, which is: L = -log(exp(sim(z i ,z j ) / τ) / Σ k=1,k≠i,j 2N exp(sim(z i ,z k ) / τ)); Among them, z i = h(g(x i )), z i and z j For the output embedding vectors of different augmented views of the same signal, z k For the batch excluding z i The output embedding vectors of all other augmented samples (including z) j (and all negative samples), i is the anchor sample index, j is the positive sample index that is homologous to i, k traverses all sample indices that are different from i in the 2N augmented samples, sim(·) is the cosine similarity function, τ is the temperature coefficient and its value ranges from 0.1 to 0.5, N is the batch size, and 2N is the total number of augmented samples.

[0039] After pre-training, the projection head h(·) is removed, and the output of the encoder g(·) is used as the deep embedding feature vector f. deep : f deep = g(x).

[0040] S3: Input the channel state parameters into the gated fusion network to generate fusion weights for physical features and deep embedded features. Then, weight the physical feature vector and the deep embedded feature vector according to these weights to obtain a hybrid feature vector, such as... Figure 2 As shown; The channel state parameters [γ, δ] obtained from S1 are input into the gated fusion network to generate physical feature weights ω and deep embedding feature weights β.

[0041] The gated fusion network employs a two-layer fully connected network. The first layer contains 32 neurons with the ReLU activation function, and the second layer contains 2 neurons followed by the Softmax activation function, outputting two weight values ​​ω and β that satisfy ω + β = 1 and ω, β ∈ [0, 1]. [ω, β] =Softmax(W2·ReLU(W1·[γ, δ]+ b1) + b2); Where W1 and W2 are the weight matrices of the first and second fully connected layers, respectively, and b1 and b2 are the bias vectors of the first and second fully connected layers, respectively.

[0042] The training objective of the gated fusion network is to optimize the combined performance of the fused features in both known signal classification and unknown signal detection. During training, the gated fusion network learns to ensure that when the signal-to-noise ratio γ is below a preset threshold γ... th (In this embodiment, the value range is 0~5dB) When the physical feature weight ω is increased, its robustness to channel changes is utilized; when the signal-to-noise ratio γ is higher than the preset threshold γ th At that time, the weight β of the deep embedding feature is increased to take advantage of its high discriminative power for subtle differences in the signal.

[0043] The physical feature vector f phy and deep embedding feature vector f deep Weighted fusion is performed according to the weights to obtain the mixed feature vector: f mix = ω · f phy + β · f deep .

[0044] S4: Calculate the matching degree between the hybrid feature vector and each category in the pre-constructed known signal space, and dynamically adjust the category decision threshold of the signal according to the channel state parameters. If the matching degree of all categories of the signal is lower than the category decision threshold, it is determined to be an unknown signal; otherwise, the signal is identified as the known signal category with the highest matching degree. First, a known signal space is constructed. During the training phase, for each class of known signals, its mixed feature vector set is extracted. For each signal class c, the Support Vector Data Description (SVDD) method is used to construct the single-class classification boundary. SVDD finds a hypersphere with a minimum volume to enclose the training samples of that class, with the hypersphere having a center a. c and radius R c Description. For the set of mixed feature vectors {f} of the c-th type of signal. mix Given the following optimization problem: min R c 2 + C·Σ i ξ i ; st ||f mix ,i - a c || 2 ≤ R c 2 + ξi , ξ i ≥ 0; Where C is the penalty parameter, ξ i Let a be the slack variable for the i-th sample, where a c and R c Here, denoted as the center and radius of the SVDD hypersphere of the c-th signal, respectively.

[0045] After solving, the hypersphere center a of the c-th type signal is obtained. c and radius R c This constitutes the known spatial boundary of this type of signal. The boundary parameters {(a} of all class C signals are defined. c , R c The combination of ) | c=1,2,...,C} constitutes the known signal space.

[0046] Then, open set identification is performed: During the testing phase, the mixed feature vector f is extracted from the test signal to be identified according to S1-S3. mix Calculate f mix The matching degree d with each category in the known signal space c .

[0047] The matching degree uses a normalized matching degree, which is calculated as follows: d c = exp(-dist(f mix B c ) / σ); where dist(f) mix B c Let be the distance from the mixed feature vector to the spatial boundary of the known signal of class c. For the SVDD boundary, this distance is: dist(f mix B c ) = ||f mix - a c || 2 - R c 2 ; σ is a scaling parameter that controls the rate at which the matching degree decays with distance, and its value is determined through validation set optimization.

[0048] Match degree d c The range of values ​​for d is (0, +∞). c A higher value indicates a higher degree of matching with category c; in particular, when f mix When located inside the boundary of class c, dist<0, d c >1, when f mix When d is located on the boundary c =1, when f mixWhen located outside the boundary, dist>0, d c <1.

[0049] The category decision threshold T is dynamically adjusted according to channel state parameters: T = T0· f(γ, δ); Where T0 is the reference threshold, and f(γ, δ) is the channel adaptive adjustment function. In this embodiment, the channel adaptive adjustment function is designed as: f(γ, δ) = (1 + tanh(k γ · (γ - γ ref [[ID=15>]]) / 2 · (1 - k δ · tanh(δ / δ ref )); Where tanh(·) is the hyperbolic tangent activation function, γ ref is the reference signal-to-noise ratio (with a value range of 5~15dB in this embodiment), k γ is the signal-to-noise ratio adjustment coefficient (with a value range of 0.2~0.5 in this embodiment), δ ref is the reference channel fading factor (with a value range of 0.5~2.0 in this embodiment), k δ is the fading adjustment coefficient and 0<k δ ≤1.

[0050] The characteristics of this adjustment function are: when the signal-to-noise ratio γ decreases, f(γ, δ) decreases accordingly to lower the category decision threshold and reduce the false rejection rate of unknown signals; when the signal-to-noise ratio γ increases, f(γ, δ) increases accordingly to raise the category decision threshold and reduce the missed detection rate of unknown signals. When the channel fading factor δ increases (the channel deteriorates), f(γ, δ) decreases appropriately to adapt to the feature distribution change caused by fading.

[0051] Compare each calculated matching degree d c with the category decision threshold T: If max{d1, d2, ..., d C}<T, the signal is determined as an unknown signal; Otherwise, the signal is identified as the known signal category with the highest matching degree: c * = arg max d c .

[0052] Feature clustering and storage are performed on the samples determined as unknown signals. Specifically, using the mixed feature vector f mix as features, the density-based DBSCAN clustering algorithm is adopted to perform online clustering on the unknown signal samples, and the cosine distance is used as the clustering distance metric.

[0053] S5: Perform feature clustering and storage on samples identified as unknown signals. When the number of clustered samples of a certain type of unknown signal exceeds a preset threshold N... th (In this embodiment, the value range is 50~200) When the value is in the range of 50~200, the incremental update process is triggered to add this type of unknown signal to the known signal space. The incremental update process includes the following sub-steps: S5.1, perform manual annotation on the clustering of unknown signals that trigger updates, confirm their signal types and assign category labels; S5.2 Extract the mixed feature vector of known signal samples of this category, and use the same SVDD method as above to construct the single-class classification boundary of this type of signal; S5.3, add the newly created single-class classification boundary parameters to the known signal space, and expand the number of classes C = C + 1; S5.4, in the process of adding a new category, keep the boundary parameters of the original known categories unchanged to avoid the forgetting of existing knowledge caused by incremental updates.

[0054] Through the aforementioned incremental learning mechanism, the system can continuously accumulate and identify newly emerging signal types during operation, enabling the open set recognition capability to continuously evolve and adapt to the ever-changing spectrum environment.

[0055] Example 2 This embodiment theoretically analyzes the inherent mechanism by which the physical layer invariant characteristics are invariant to channel changes, providing theoretical support for the technical solution of this invention.

[0056] Channel invariance of cyclic spectral density features: The cyclic spectral density (CSD) function is a core tool for analyzing cyclically stationary signals. For the narrowband communication signal s(t) in this embodiment, its cyclic autocorrelation function is defined as: R s α (υ) = lim Φ→∞ (1 / Φ) ∫ -Φ / 2 Φ / 2 s(t+υ / 2) s * (t-υ / 2) e -j2παt dt; Where α is the cycle frequency, υ is the time delay parameter, Φ is the integration time, and s*(·) represents the complex conjugate of s(·). For R s α (υ) Perform a Fourier transform to obtain the cyclic spectral density function S s α (f).

[0057] The received signal after passing through the multipath channel h(t) is x(t) = h(t) * s(t) + w(t), and its time-varying cyclic autocorrelation function is: R x α(υ) = R h 0 (υ) · R s α (υ) + R w α (υ); Where R h 0 (υ) is the autocorrelation function of the channel (at the cyclic frequency α=0), R w α (υ) represents the cyclic autocorrelation of the noise.

[0058] For zero-mean cyclic stationary noise (such as AWGN), R w α (υ) ≡ 0 holds for all α ≠ 0. Therefore, at non-zero cyclic frequencies: R x α(υ) = R h 0 (υ) · R s α (υ), α ≠ 0; This means that: (1) the cyclic frequency position of the signal (corresponding to symbol rate, carrier frequency, chip rate, etc.) remains unchanged before and after the channel. The channel only scales the peak of the cyclic spectrum in amplitude, without changing the frequency position of the peak. (2) The cyclic spectrum of Gaussian noise at α≠0 is theoretically zero, therefore the cyclic spectral density feature is inherently immune to additive Gaussian noise; (3) For multipath fading channels, although the superposition of different paths will cause frequency-selective fading, the cyclic frequency position remains unchanged. This property is the theoretical basis for the use of cyclic spectral density characteristics as the "anchor point" of open set boundary in this invention.

[0059] Channel invariance of higher-order cumulant features: For a zero-mean complex random process X, its p-th order cumulant C p,q It has the following important properties: Property 1 (Gaussian noise suppression): For a complex Gaussian random variable, all cumulants of order greater than 2 are always zero. This means that the characteristics of higher-order cumulants (fourth order and above) are theoretically completely unaffected by additive Gaussian noise.

[0060] Property 2 (Invariance to Linear Transformation): If Y = aX + b, where a is a complex constant and b is a random variable independent of X, then C p,q (Y) = a p · (a * ) q · C p,q (X) (for p>2). This indicates that in a flat fading channel (multiplicative complex constant a), the ratio between cumulants of different orders (e.g., |C|) 40 | / |C 42 | 2 It is channel independent.

[0061] Property 3 (Channel Invariance of Normalized Cumulus): Define the normalized cumulant λ p,q = C p,q / (C 21 ) p / 2 Then λ p,q In a flat fading channel, it is independent of the channel gain and depends only on the probability distribution shape of the modulation scheme.

[0062] The above properties indicate that the normalized ratios of higher-order cumulants are physical characteristics that are strictly invariant to channel variations. This invention incorporates these normalized ratios as part of a physical feature vector, providing a stable constraint unaffected by the channel in the construction of open set boundaries.

[0063] Channel invariance of time-frequency ridge features: For narrowband communication signals, they can be represented in complex envelope form as s(t) = A(t)exp(jφ(t)). After passing through a frequency-flat, slowly fading channel, the received signal is x(t) = h(t)s(t) + w(t) ≈ |h|e {jθA(t)exp(jφ(t))} + w(t), the instantaneous frequency of its time-frequency ridge is: finst(t) = (1 / 2π) · d[φ(t) + θ] / dt = (1 / 2π) · dφ(t) / dt; Since the channel phase shift θ is constant (slow fading assumption), the instantaneous frequency f inst (t) is independent of the channel. Therefore, under slow fading channel conditions, the shape and trajectory of the time-frequency ridge are entirely determined by the modulation method of the signal itself.

[0064] The three physical features mentioned above provide invariance to channel variations from different dimensions: the cyclic spectral density feature is invariant in the cyclic frequency dimension, the higher-order cumulants are invariant in the statistical distribution dimension (through normalized ratios), and the time-frequency ridge is invariant in the instantaneous frequency dimension. This invention combines these three complementary invariant physical features to construct a physical feature vector, providing multi-dimensional and multi-level channel robust constraints for open set boundaries.

[0065] Example 3: This embodiment comprehensively verifies the independent contribution and channel robustness of each innovative component in the technical solution of this invention through ablation experiments and comparative experiments under different channel models.

[0066] Simulation settings: Known signal categories include five modulation schemes: BPSK, QPSK, 8PSK, 16QAM, and 64QAM; unknown signal categories include three modulation schemes: 32QAM, GMSK, and OFDM. Evaluation metrics include the unknown signal detection rate (Det). unknown The F1-score is determined by the overall open set.

[0067] To quantify the independent contribution of each innovative component, the following ablation variants are defined: Variant A (baseline): Only deep embedding features are used + fixed threshold (no gated fusion, no channel awareness); Variant B (+Physical Features): Fixed splicing of physical features and deep features + fixed threshold (no gating, no channel awareness); Variant C (+Gated Fusion): Physical feature + deep feature gated fusion + fixed threshold (no adaptive threshold); Variant D (+Adaptive Threshold): Physical feature + deep feature gating fusion + channel adaptive threshold (complete solution of this invention).

[0068] like Figure 3 As shown, the ablation experiment results (SNR=5dB, Rayleigh multipath channel): Variant A (baseline): Unknown signal detection rate 52.3%, known signal accuracy 89.7%, F1-score 0.631; Variant B (+physical characteristics): Unknown signal detection rate 61.8%, known signal accuracy 92.1%, F1-score 0.708 (+12.2%). Variant C (+Gated Fusion): Unknown signal detection rate 72.5%, known signal accuracy 94.3%, F1-score 0.796 (+26.1%). Variant D (Complete Solution): Unknown signal detection rate 83.1%, known signal accuracy 94.8%, F1-score 0.871 (+38.0%). The results of the ablation experiment showed that: (1) Variant A→B: Adding physical features alone improves the detection rate of unknown signals by 9.5%, which verifies the contribution of physical features to the channel robustness of open set boundaries; (2) Variant B→C: After adding gating fusion, the detection rate of unknown signals is increased by 10.7%, which verifies the effectiveness of channel-aware dynamic fusion; (3) Variant C→D: After adding the adaptive threshold, the detection rate of unknown signals increased by 10.6%, which verified the effectiveness of the channel perception threshold adjustment.

[0069] To verify the generalization capability of this invention under different channel conditions, comparative experiments were conducted under four channel models (SNR=10dB), such as... Figure 4 As shown: AWGN channel: This invention 95.7%, OpenMax 82.3%, distance method 86.5%, this invention's advantage +13.4%; Rayleigh fading (K=0): This invention 88.2%, OpenMax 65.7%, distance method 71.2%, this invention's advantage +22.5%; Ricean fading (K=5dB): This invention 92.1%, OpenMax 74.8%, distance method 78.9%, this invention's advantage +17.3%; Ricean fading (K=10dB): This invention 94.3%, OpenMax 79.6%, distance method 83.1%, this invention's advantage +14.7%; Nakagami-m (m=1): This invention has 87.5% superiority, OpenMax has 63.2%, distance methods have 69.4%, and this invention has an advantage of +24.3%. Nakagami-m (m=3): This invention 91.8%, OpenMax 72.5%, distance method 77.6%, this invention's advantage +19.3%.

[0070] Experimental results show that, under all channel models, the unknown signal detection rate of the present invention is significantly better than that of the comparative methods, with the lowest advantage being 13.4% (AWGN) and the highest reaching 24.3% (Nakagami-m, m=1). As the channel degradation increases, the performance gap between the present invention and the comparative methods widens, indicating that the channel adaptive mechanism of the present invention is particularly effective under adverse channel conditions.

[0071] Further analysis of the sensitivity of this invention to channel model parameters: Under Ricean channel conditions, as the K-factor changes from -5dB to 15dB, the unknown signal detection rate of this invention increases from 81.3% to 96.1%, a change of 14.8%; while the OpenMax method increases from 51.6% to 83.2%, a change of 31.6%. This invention is far less sensitive to changes in channel parameters than the comparative methods, indicating that the channel-aware mechanism of this invention effectively suppresses the disturbance of open set recognition performance caused by channel changes.

[0072] In the Nakagami-m channel, as the m-factor changes from 0.5 to 5.0, the detection rate of this invention increases from 79.8% to 93.5%, a change of 13.7%; the OpenMax method shows a change of 35.2%. The results consistently demonstrate that this invention has significant channel robustness advantages.

[0073] The above theoretical analysis and experimental verification together show that, through hybrid modeling of physical and deep features and a channel-aware adaptive mechanism, this invention significantly improves the robustness and channel adaptability of unknown signal detection while maintaining the accuracy of known signal classification.

[0074] Example 4 This embodiment provides an open-set radio signal identification system based on hybrid feature space and channel awareness, including the following modules: The channel estimation module is used to acquire radio signals in a specified frequency band through an antenna and RF front-end, perform low-noise amplification, down-conversion, filtering, analog-to-digital conversion, and digital down-conversion processing, and output a normalized baseband complex signal sequence; it receives the baseband complex signal sequence and estimates channel state parameters, including at least the signal-to-noise ratio (SNR) estimate and the channel fading factor. This module employs the M2M4 SNR estimation method and the envelope normalized variance fading factor estimation method, and includes an antenna, an RF front-end chip, an analog-to-digital converter, and a digital signal processor.

[0075] The physical feature and deep embedding feature extraction module, connected to the signal acquisition and preprocessing module, is used to extract physically invariant features from the baseband complex signal sequence, including cyclic spectral density features, higher-order cumulant features, and time-frequency ridge features, outputting a physical feature vector. It also inputs the baseband complex signal sequence into a deep embedding network pre-trained through contrastive learning, outputting a deep embedding feature vector. This module includes an encoder network based on ResNet-1D or TCN, whose parameters are determined through contrastive learning pre-training during the training phase.

[0076] The gated fusion module, connected to the channel estimation module, physical feature extraction module, and deep embedding feature extraction module, respectively, is used to input channel state parameters into the gated fusion network to generate feature fusion weights, and then weight and fuse the physical feature vector and the deep embedding feature vector according to the weights to output a hybrid feature vector. This module contains a two-layer fully connected network as the gated network.

[0077] The open set identification module, connected to the gated fusion module, is used to calculate the matching degree between the hybrid feature vector and each category of the known signal space. It dynamically adjusts the category decision threshold based on channel state parameters and outputs the identification result. The known signal space is stored in the module's internal memory in the form of SVDD hypersphere parameters.

[0078] The incremental learning module, connected to the open set recognition module, is used to perform feature clustering and storage on samples identified as unknown signals. When the number of clustered samples of a certain type of unknown signal exceeds a preset threshold, the incremental update process is triggered to add the new signal category to the known signal space.

[0079] The above modules can be implemented using hardware platforms such as FPGA, DSP or embedded processor, or they can be implemented in software on a general computing platform.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An open set signal recognition method based on hybrid features and channel awareness, characterized in that, Includes the following steps: S1: Obtain the baseband complex signal sequence, perform channel estimation, and obtain the channel state parameters; S2: Extract physical layer invariant features from the baseband complex signal sequence to form a physical feature vector. Input the baseband complex signal sequence into a pre-trained deep embedding network to extract the deep embedding feature vector. S3: Input the channel state parameters into the gated fusion network to generate fusion weights for physical features and deep embedded features. Then, perform weighted fusion of the physical feature vector and the deep embedded feature vector according to the fusion weights to obtain the hybrid feature vector. S4: Calculate the matching degree between the hybrid feature vector and each category in the pre-constructed known signal space, and dynamically adjust the category decision threshold of the signal according to the channel state parameters. If the matching degree of all categories of the signal is lower than the category decision threshold, it is determined to be an unknown signal; otherwise, the signal is identified as the known signal category with the highest matching degree, and the identification result is output.

2. The open set signal identification method based on hybrid features and channel awareness according to claim 1, characterized in that, In S2, physical layer invariant features are extracted from the baseband complex signal sequence. Specifically, physical layer features that are invariant to channel changes are extracted from the normalized baseband complex signal sequence. These physical layer invariant features include cyclic spectral density features, higher-order cumulant features, and time-frequency ridge features.

3. The open set signal identification method based on hybrid features and channel awareness according to claim 2, characterized in that, The method for extracting the cyclic spectral density features is as follows: calculate the cyclic spectral density function of the baseband complex signal sequence, and extract the peak position, peak amplitude, and profile entropy on the cyclic frequency-spectral frequency plane as cyclic spectral density features.

4. The open set signal identification method based on hybrid features and channel awareness according to claim 2, characterized in that, Higher-order cumulants include second-order cumulants C. 20 C 21 Fourth-order cumulant C 40 C 41 C 42 Sixth-order cumulant C 60 C 63 and the eighth-order cumulant C 80 At least one of them.

5. The open set signal identification method based on hybrid features and channel awareness according to claim 1, characterized in that, In S3, the gated fusion network takes the channel state parameters as input, passes through at least one fully connected layer and a Softmax activation function layer, and outputs physical feature weights ω and deep embedding features β, where ω+β=1 and ω,β∈[0,1]; The weighted fusion calculation method is as follows: f mix = ω·f phy + β·f deep ; Where f phy f is the physical feature vector. deep For deep embedding feature vectors, f mix It is a mixed feature vector.

6. The open set signal identification method based on hybrid features and channel awareness according to claim 5, characterized in that, The channel state parameters include at least the signal-to-noise ratio (SNR) estimate and the channel fading factor. The gated fusion network increases the weight of physical features when the SNR estimate is lower than a preset threshold, and increases the weight of deep embedding features when the SNR estimate is higher than a preset threshold.

7. The open set signal identification method based on hybrid features and channel awareness according to claim 1, characterized in that, The category decision threshold T is dynamically adjusted based on the channel state parameters: T = T0 · f(γ, δ); Where T0 is the baseline threshold and f(γ, δ) is the channel adaptive adjustment function. When the signal-to-noise ratio γ decreases or the channel fading factor δ increases, f(γ, δ) is decreased to reduce the category decision threshold. When the signal-to-noise ratio γ increases or the channel fading factor δ decreases, f(γ, δ) is increased to increase the category decision threshold.

8. The open set signal identification method based on hybrid features and channel awareness according to claim 1, characterized in that, It also includes S5: performing feature clustering and storage on samples determined to be unknown signals. When the number of clustered samples of a certain type of unknown signal exceeds a preset threshold, an incremental update process is triggered to add the unknown signal of that type to the known signal space.

9. The open set signal identification method based on hybrid features and channel awareness according to claim 8, characterized in that, The incremental update process includes: manually labeling and confirming the clustering of unknown signals that trigger the update; after confirmation, extracting the mixed features of the signal and constructing its single-class classification boundary, and adding it to the known signal space; keeping the original known class boundaries unchanged during the addition process.

10. An open set signal recognition system based on hybrid features and channel awareness, characterized in that, include: The channel estimation module is used to acquire the baseband complex signal sequence, perform channel estimation, and obtain channel state parameters; The physical feature and deep embedding feature extraction module is used to extract physical layer invariant features from the baseband complex signal sequence to form a physical feature vector. The baseband complex signal sequence is input into a pre-trained deep embedding network to extract the deep embedding feature vector. The gated fusion module is used to input channel state parameters into the gated fusion network, generate fusion weights for physical features and deep embedded features, and perform weighted fusion of physical feature vectors and deep embedded feature vectors according to the fusion weights to obtain a hybrid feature vector. The open set identification module is used to calculate the matching degree between the hybrid feature vector and each category in the pre-constructed known signal space, and dynamically adjust the category decision threshold of the signal according to the channel state parameters. If the matching degree of all categories of the signal is lower than the category decision threshold, it is determined to be an unknown signal; otherwise, the signal is identified as the known signal category with the highest matching degree, and the identification result is output. The incremental learning module, connected to the open set recognition module, is used to perform feature clustering and storage on samples identified as unknown signals. When the number of clustered samples of a certain type of unknown signal exceeds a preset threshold, the incremental update process is triggered to add the new signal category to the known signal space.

Images