Intelligent interference detection method and system based on data enhancement dual-granularity diffusion model

Abstract

The invention discloses an intelligent interference detection method based on a data enhancement dual-granularity diffusion model, and the method comprises the steps: obtaining a to-be-detected wireless interference signal, and carrying out the time-frequency conversion processing of the to-be-detected wireless interference signal, so as to obtain a corresponding time-frequency characteristic pattern; inputting the time-frequency characteristic graph of the wireless interference signal to be detected into a pre-trained dual-granularity condition guidance DCG model to obtain condition guidance information containing signal multi-scale characteristics; and inputting the condition guidance information into a pre-trained U-Net network to carry out reverse denoising sampling processing, and selecting a category corresponding to a maximum value from a plurality of obtained normalization processing results as a final detection result. According to the method, the technical problem that dirty data misleads model training due to scarcity of real interference samples and easy damage of a signal physical structure by a conventional enhancement means in an existing detection method based on a discriminant deep neural network and a reconstruction method based on a generative model can be solved.

Description

Technical Field

[0001] This invention belongs to the fields of wireless communication signal processing, electronic countermeasures and artificial intelligence technology, and more specifically, relates to an intelligent interference detection method and system based on a data-enhanced dual-granularity diffusion model. Background Technology

[0002] In recent years, with the rapid development of wireless communication technology and the explosive growth of IoT devices, electromagnetic spectrum resources have become increasingly scarce and congested. In fields such as civilian communications, cognitive radio, and military electronic warfare, various illegal or hostile wireless interference signals are emerging in large numbers, seriously affecting the communication quality and link security of legitimate users. Therefore, in order to ensure spectrum security and take timely anti-interference measures, there is an urgent need for an intelligent interference detection and identification technology that can quickly and accurately identify the types of interference signals.

[0003] Existing intelligent interference detection methods can be mainly divided into three categories: The first category is detection methods based on expert-designed features, which extract statistical features such as higher-order cumulants and cyclic spectral moments of signals and combine them with preset thresholds or simple classifiers to achieve identification; the second category is detection methods based on discriminative deep neural networks, which use convolutional neural networks (CNNs) or residual networks (ResNets) to mine deep features of signals from massive amounts of data and achieve classification by training fixed decision boundaries; the third category is reconstruction methods based on generative models, such as the Denoising Diffusion Probabilistic Model (DDPM), which extracts deep features with reconstruction capabilities by simulating the reverse process of gradually recovering data from Gaussian noise.

[0004] However, the aforementioned existing intelligent interference detection methods all have some significant drawbacks: (1) Both the detection methods based on discriminative deep neural networks and the reconstruction methods based on generative models mentioned above face the problem of scarce real interference samples and the tendency of conventional enhancement methods to destroy the physical structure of the signal. Specifically, in non-cooperative communication scenarios, it is extremely difficult to capture and label a large number of real interference signals, resulting in a serious "small sample" problem; while directly applying geometric transformation enhancement methods in the field of image processing will destroy the strict temporal causality and phase orthogonality of radio frequency signals, generating "dirty data" that misleads model training; (2) Most of the above-mentioned detection methods based on discriminative deep neural networks adopt a single-granularity feature extraction method, which is difficult to take into account both the macroscopic contour and micro-transient features of the signal. Most existing discriminative models use global feature classifiers. When faced with interference signals accompanied by complex background noise and containing transient local subtle features, they often lack detailed characterization of local regions of interest, resulting in insufficient ability to distinguish complex modulation methods.

[0005] (3) The existing detection methods based on discriminative deep neural networks are essentially looking for fixed hard decision boundaries. In low signal-to-noise ratio (SNR) environments, relying only on a limited number of strong signal samples for training, the model cannot learn to approximate the optimal decision boundary of the maximum likelihood estimate, resulting in a serious decrease in the recognition accuracy of weak signals. (4) The diffusion models used in existing generative model-based reconstruction methods are mostly applied to image generation. When directly transferred to the field of signal detection, due to the lack of strong conditional constraints on the features of interference signals, the model is prone to distribution shift during the reverse mapping process, resulting in large computational overhead and insufficient classification robustness. Summary of the Invention

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides an intelligent interference detection method and system based on a data-augmented dual-granularity diffusion model. Its purpose is to solve the technical problems of existing detection methods based on discriminative deep neural networks and reconstruction methods based on generative models, which suffer from the scarcity of real interference samples and the tendency of conventional enhancement methods to damage the physical structure of the signal, leading to "dirty data" misleading model training. It also addresses the technical problems of existing detection methods based on discriminative deep neural networks, which, due to the use of a single-granularity feature extraction method, cannot simultaneously consider the macroscopic contours and microscopic transient features of the signal, resulting in low discrimination against complex modulated signals. Furthermore, it addresses the technical problems of existing discriminative models exhibiting blurred decision boundaries and poor robustness in low signal-to-noise ratio environments, and the technical problems of existing diffusion models lacking targeted guidance mechanisms, resulting in weak generalization and high computational overhead.

[0007] To achieve the above objectives, according to one aspect of the present invention, an intelligent interference detection method based on a data-augmented dual-granularity diffusion model is provided, comprising the following steps: Step 1: Acquire the wireless interference signal to be detected, and perform time-frequency transformation processing on the wireless interference signal to be detected to obtain the corresponding time-frequency feature map; Step 2: Analyze the time-frequency characteristic map of the wireless interference signal to be detected obtained in Step 1. Input a pre-trained dual-granularity conditionally guided DCG model to obtain conditionally guided information containing multi-scale features of the signal. ; Step 3: Input the conditional guidance information obtained in Step 2 into the pre-trained U-Net network for inverse denoising and sampling processing, and select the category corresponding to the maximum value from the multiple normalized processing results as the final detection result.

[0008] Preferably, step one includes the following sub-steps: (A1) Use radio frequency receiving equipment to collect the wireless interference signal to be detected in the environment, and perform down-conversion, filtering and analog-to-digital conversion processing on the wireless interference signal to be detected in sequence to obtain the in-phase orthogonal IQ complex sequence data of the wireless interference signal to be detected. (A2) Perform short-time Fourier transform (STFT) processing on the IQ complex sequence data of the wireless interference signal to be detected obtained in step (A1) to obtain the time-frequency feature map corresponding to the wireless interference signal to be detected. It is a two-dimensional time-frequency characteristic map.

[0009] Preferably, step two includes the following sub-steps: (B1) Input the time-frequency feature map of the wireless interference signal to be detected obtained in step one into the global flow branch of the DCG model to extract the deep features using the backbone network, and perform global average pooling on the deep features to obtain a global prior vector reflecting the overall energy distribution of the wireless interference signal to be detected. ; (B2) Input the time-frequency feature map of the wireless interference signal to be detected obtained in step one into the local flow branch of the DCG model to crop multiple regions of interest (ROI) patches from the time-frequency feature map, and use a local encoder to extract features from all ROI patches. Then, use gated attention weights to fuse all features from all ROI patches to obtain a local prior vector. ; (B3) The global prior vector obtained in step (B1) The local prior vector obtained in step (B2) The system concatenates all feature vectors from the time-frequency feature map obtained in step one to obtain conditional guidance information containing both global and local prior vectors. .

[0010] Preferably, step three includes the following sub-steps: (C1) Set the iteration counter Preset acceleration steps ; (C2) Judgment If the number of iterations is less than the preset threshold, proceed to step (C6); otherwise, proceed to step (C3). (C3) will be the first noise variables in the next iteration Iteration counter and the conditional guidance information obtained in step two. Input the U-Net network to obtain the predicted noise Among them, noise variables The initial values ​​are derived from a standard normal distribution; (C4) Predicted noise obtained from step (C3) Get the noise variables from the previous time step : ; in, and These are predefined noise scheduling parameters. The cumulative product, It is a random variable that follows a standard normal distribution. It is the variance term; (C5) Set the iteration counter , and return to step (C2); (C6) The prediction vector at time step 0 obtained after the iteration is completed. Softmax normalization is performed, and the category corresponding to the maximum value among the multiple normalization results is selected as the final detection result.

[0011] Preferably, the DCG model and the U-Net network are trained through the following steps: (1) Construct an initial simulation dataset consisting of multiple wireless interference signals, and use a data augmentation strategy based on channel impairment modeling to superimpose noise components simulating channel impairment characteristics onto the initial simulation dataset to obtain a simulation dataset. Divide the simulation dataset into a training set, a validation set, and a test set in a ratio of 8:2:1. The samples in the training set and the validation set all contain the real interference category label corresponding to the sample. ; (2) Use the DCG model to perform feature extraction and condition construction on the training set obtained in step (1) to obtain condition-guided information. ; (3) Using the training set obtained in step (1) and the conditional guidance information obtained in step (2), perform end-to-end training on the DCG model and the U-Net network to obtain the trained DCG model and U-Net network.

[0012] Preferably, step (1) includes the following sub-steps: (1-1) Use simulation software to generate multiple wireless interference signals containing various modulation types. All wireless interference signals form the initial simulation dataset. (1-2) Construct standard Gaussian white noise sequences containing different signal-to-noise ratio levels ,in This represents the first sequence in the standard Gaussian white noise sequence library. A standard Gaussian white noise sequence, This indicates the size of the standard Gaussian white noise sequence library; (1-3) For each wireless interference signal in the initial simulation dataset obtained in step (1-1) In this regard, the standard Gaussian white noise sequence obtained from step (1-2) Randomly select a noise sample ,in This indicates the sequence number of the selected noise sample in the standard Gaussian white noise sequence library; (1-4) The noise samples obtained in step (1-3) Perform a random permutation operation to obtain the permuted noise samples. ; (1-5) The noise samples obtained after the permutation operation in step (1-4) Superimposed on each wireless interference signal Above, to obtain the corresponding enhanced samples. : ; (1-6) All the augmented samples obtained The data is then aggregated to obtain the simulation dataset. (1-7) Divide the simulation dataset obtained in step (1-6) into a ratio of 8:2:1 to obtain a training set, a validation set, and a test set. The samples in the training set and the validation set all contain the corresponding real interference category labels. .

[0013] Preferably, step (2) includes the following sub-steps: (2-1) Obtain each sample and its corresponding true class label from the training set obtained in step (1). The sample is then converted into a two-dimensional time-frequency graph using a short-time Fourier transform (STFT). (2-2) Input the two-dimensional time-frequency graph obtained in step (2-1) into the global flow branch of the DCG model. It first uses the backbone network to extract deep features, and then inputs the deep features into the DCG model. Convolution is performed to generate a global saliency map, and finally, global average pooling is applied to this global saliency map to obtain the global prior vector. ; (2-3) Input the two-dimensional time-frequency map obtained in step (2-1) into the local flow branch of the DCG model, locate multiple regions of interest (ROI) patches based on the global saliency map, extract features from all ROI patches using a local encoder, obtain the weight of each feature through a gated attention mechanism, and then perform weighted fusion of all feature weights to obtain the local prior vector. ; (2-4) Input the two-dimensional time-frequency graph obtained in step (2-1) into the backbone network to obtain the time-frequency feature vector. ; (2-5) The global prior vector obtained in step (2-2) The local prior vector obtained in step (2-3) and the time-frequency feature vector obtained in step (2-4) Perform splicing to obtain conditional guidance information. .

[0014] Preferably, in step (2-3), each feature is calculated using the following formula, i.e., the first... Weights of each feature: ; in, , They represent the first The and the first Features of an ROI patch ∈[1, the total number of ROI patches obtained], These are learnable weight parameters.

[0015] Preferably, step (3) specifically includes the following sub-steps: (3-1) Label the true interference category corresponding to each sample in the training set obtained in step (1). Input the forward diffusion module in the U-Net network and at random time steps Label the real interference category Add Gaussian noise To obtain random time steps Corresponding noise variables : ; in It is the product of preset noise scheduling parameters; (3-2) The noise variable obtained in step (4-1) Random time step and the conditional guidance information obtained in step (2) The input is processed by the U-Net network for feature fusion and inverse prediction to obtain prediction noise. ; (3-3) The Gaussian noise obtained from step (3-1) The predicted noise obtained in step (3-2) Mean square error is calculated to obtain the noise estimation loss. ; (3-4) Based on the global prior vector obtained in step (2-2) The local prior vector obtained in step (2-3) Obtain the maximum mean difference MMD regularization loss : ; in, and They represent the first and second training sets, respectively. Seed granularity (i.e., global granularity) The global prior vector corresponding to the training sample and the first training sample Seed size (i.e., local grain size) under the first Local prior vectors corresponding to each training sample; and These represent the obtained global prior vectors, respectively. The total number and local prior vectors The total number; This represents the kernel mapping function, used to map feature vectors to a high-dimensional reproducing kernel Hilbert space RKHS; This represents norm operations in RKHS; (3-5) The noise estimation loss obtained in step (3-3) The MMD regularization loss obtained in step (3-4) Perform a weighted summation to obtain the total loss value. ; (3-6) For each sample in the training set obtained in step (1-7), use the gradient descent algorithm based on the total loss value obtained in step (3-5). All learnable weight parameters in the DCG module and U-Net network are synchronously iteratively updated until the DCG module and U-Net network reach a preset number of iterations (500 times in this embodiment of the invention), and the optimal parameters of the DCG module and U-Net network at this time are obtained, thereby obtaining the initially trained DCG module and U-Net network. (3-7) Use the test set obtained in step (1-7) to evaluate the performance of the DCG module and U-Net network initially trained in step (3-6) until the predicted class matches the true class label. The detection accuracy is optimized until the DCG module and U-Net network are finally trained.

[0016] According to another aspect of the present invention, an intelligent interference detection system based on a data-enhanced dual-granularity diffusion model is provided, comprising: The first module is used to acquire the wireless interference signal to be detected and to perform time-frequency transformation processing on the wireless interference signal to be detected in order to obtain the corresponding time-frequency feature map. The second module is used to process the time-frequency characteristic map of the wireless interference signal to be detected obtained from the first module. Input a pre-trained dual-granularity conditionally guided DCG model to obtain conditionally guided information containing multi-scale features of the signal. ; The third module is used to input the conditional guidance information obtained from the second module into the pre-trained U-Net network for inverse denoising and sampling processing, and select the category corresponding to the maximum value from the multiple normalized processing results as the final detection result.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) Since the present invention adopts steps (1-1) to (1-7) in the training phase, it uses a data augmentation strategy based on channel impairment modeling to simulate real channel impairment by randomly arranged noise without destroying the temporal causality and phase orthogonality of radio frequency signals. This effectively expands the training data covering the distribution of strong and weak signals, and can solve the technical problems of existing detection methods based on discriminative deep neural networks and reconstruction methods based on generative models, which are caused by the scarcity of real interference samples and the easy destruction of the physical structure of the signal by conventional augmentation methods, resulting in "dirty data" misleading the model training. (2) Since the present invention adopts steps (2-1) to (2-5) in the training phase, it constructs a dual-granularity condition guidance (DCG) module to extract global priors that reflect the overall energy distribution and local priors that focus on detailed features in parallel, thereby realizing the deep mining of multi-scale features of the signal. It can solve the technical problems of existing detection methods based on discriminative deep neural networks, which cannot take into account both macroscopic contours and microscopic transient features of the signal and have low discrimination for complex modulation signals due to the use of a single-granularity feature extraction method. (3) Since the present invention adopts step (1) in the training stage and step (3) in the inference stage, it pushes the decision boundary of the classifier to the optimal position of the maximum likelihood estimation by enhancing data through the physical layer, and extracts deep features by using the generative reconstruction mechanism. It can solve the technical problems of the existing discriminative model having fuzzy decision boundary and poor recognition robustness in low signal-to-noise ratio environment. (4) Since the present invention adopts steps (3-2) to (3-4) in the training phase, it introduces a dual-granularity strong conditional guidance vector and uses the maximum mean difference MMD regularization loss to constrain the consistency of feature distribution, which effectively reduces the search space from Gaussian noise to class distribution, thereby significantly improving the inference efficiency and accuracy of the diffusion model in signal classification tasks. It can solve the technical problems of poor generalization and high computational cost caused by the lack of a targeted guidance mechanism in the existing diffusion model.

[0018] (5) The implementation of the present invention is simple, based on the mature U-Net architecture and short-time Fourier transform (STFT), and the engineering deployment is easy; the present invention has wide applicability and can be flexibly extended to spectrum situational awareness scenarios in various complex electromagnetic environments. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the intelligent interference detection method based on the data-enhanced dual-granularity diffusion model of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the data augmentation strategy based on channel impairment modeling of the present invention; Figure 3 This is a schematic diagram of the inference and detection process for wireless interference signals in this invention; Figure 4 This is a schematic diagram of the structure of the DCG model and U-Net network in this invention; Figure 5 This is a schematic diagram of the training process of the DCG model and U-Net network in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] In the description of this invention, it should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, the technical features involved in the different embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] The basic idea of ​​this invention is to improve the detection effect of wireless interference signals from four aspects. First, a data augmentation strategy based on channel impairment modeling is used to construct an initial distribution field of interference signals covering the entire signal-to-noise ratio distribution by superimposing randomly arranged Gaussian white noise. This expands the training manifold without destroying the temporal causality and phase orthogonality of the radio frequency signal, providing a high-fidelity training base for the subsequent diffusion model. Second, a dual-granularity conditional guidance module is constructed. The dual granularity includes a global granularity reflecting the overall energy distribution of the signal and a local granularity reflecting the subtle jump features of the signal. This enables parallel global and local flow branches to extract global prior vectors reflecting the energy distribution of the interference signal and local prior vectors focusing on the subtle features of the region of interest. A gated attention mechanism is used to achieve efficient fusion of multi-scale features, thereby achieving synchronous deep characterization of the macroscopic contour and microscopic transient features of the interference signal. Then, the fused strong conditional guidance vector is introduced into a U-Net-based system. The proposed denoising network establishes an inverse mapping relationship from Gaussian noise distribution to the distribution of real interference signal categories. This leverages a generative reconstruction mechanism to overcome the fuzzy decision boundary of traditional discriminative models, pushing the classifier's decision boundary closer to the optimal position of maximum likelihood estimation. Finally, during the model training phase, MMD regularization loss is introduced to constrain the consistency of global and local feature distributions in the Reproducing Kernel Hilbert Space (RKHS). This consistency is incorporated as a crucial component of the joint optimization objective, enabling the model to achieve deep alignment of multi-granularity feature distributions while learning the denoising process. This significantly improves the robustness of interference detection methods in extremely low signal-to-noise ratio and complex electromagnetic environments.

[0023] like Figure 1 As shown, this invention provides an intelligent interference detection method based on a data-augmented dual-granularity diffusion model, comprising the following steps: (1) Acquire the wireless interference signal to be detected and perform time-frequency transformation processing on the wireless interference signal to be detected to obtain the corresponding time-frequency feature map; Specifically, as shown in Figure 2, this step includes the following sub-steps: (1-1) The radio frequency receiving equipment is used to collect the wireless interference signal to be detected in the environment. The wireless interference signal to be detected is then subjected to down-conversion, filtering and analog-to-digital conversion to obtain the in-phase quadrature (IQ) complex sequence data of the wireless interference signal to be detected. (1-2) Perform a Short-Time Fourier Transform (STFT) on the IQ complex sequence data of the wireless interference signal to be detected obtained in step (1-1) to obtain the time-frequency feature map corresponding to the wireless interference signal to be detected. (It is a two-dimensional time-frequency feature map, for example, with a size of 224×224); The advantage of step (1-2) is that it uses time-frequency transformation to convert a one-dimensional time-series signal into a two-dimensional map with spatial texture features, which lays the foundation for subsequent extraction of interference signal features.

[0024] (2) The time-frequency characteristic map of the wireless interference signal to be detected obtained in step (1) Input a pre-trained dual-granularity condition guidance (DCG) model to obtain condition guidance information that includes multi-scale features of the signal. ; Specifically, this step includes the following sub-steps: (2-1) Input the time-frequency feature map of the wireless interference signal to be detected obtained in step (1) into the global flow branch of the DCG model to extract the deep features using the backbone network (such as ResNet-18), and perform global average pooling on the deep features to obtain the global prior vector reflecting the overall energy distribution of the wireless interference signal to be detected. ; (2-2) Input the time-frequency feature map of the wireless interference signal to be detected obtained in step (1) into the local flow branch of the DCG model to cut out multiple regions of interest (ROI) patches from the time-frequency feature map, and use the local encoder to extract the features in all ROI patches, and use gated attention weights to fuse all features in all ROI patches to obtain the local prior vector. (It focuses on microscopic details); (2-3) The global prior vector obtained in step (2-1) The local prior vector obtained in step (2-2) The feature vectors obtained in step (1) are concatenated to obtain conditional guidance information containing both global and local prior vectors. ; The advantage of step (2) is that by constructing parallel flow branches, the macroscopic envelope contour and microscopic transient pulse characteristics of the interference signal are synchronously represented, which effectively improves the model's ability to distinguish subtle complex modulation signals.

[0025] (3) Input the condition guidance information obtained in step (2) into the pre-trained U-Net network for inverse denoising and sampling processing, and select the category corresponding to the maximum value from the multiple normalization processing results as the final detection result.

[0026] Specifically, such as Figure 3 As shown, this step includes the following sub-steps: (3-1) Set the iteration counter Preset acceleration steps (In this embodiment, =100); (3-2) Judgment If the number of iterations is less than the preset threshold (its value ranges from -1 to 5, preferably 0), then proceed to step (3-6); otherwise proceed to step (3-3). (3-3) The first noise variables in the next iteration (its initial value) Using a standard normal distribution and an iterative counter and the conditional guidance information obtained in step (2) Input the U-Net network to obtain the predicted noise ; (3-4) Predicted noise obtained from step (3-3) Get the noise variables from the previous time step ; Specifically, this step uses the following inverse denoising sampling formula to obtain the first... Noise variables in the next iteration: ; in, and These are predefined noise scheduling parameters. The value range is from 0.98 to 0.9999, preferably decreasing linearly from 0.9999 to 0.98 with time step. The cumulative product, whose value ranges from 0.0063 to 0.9999, preferably varies with time step. It monotonically decreases from 0.9999 to 0.0063. It is a random variable that follows a standard normal distribution. It is the variance term.

[0027] (3-5) Set the iteration counter And return to step (3-2); (3-6) The prediction vector at time step 0 obtained after the iteration is completed. Softmax normalization is performed, and the category corresponding to the maximum value among the multiple normalization results is selected as the final detection result.

[0028] In this embodiment of the invention, it should be noted that wireless interference signals include, but are not limited to, various modulation types such as single-tone interference, multi-tone interference, linear frequency modulation interference, and noise frequency modulation interference. These signals often exhibit large dynamic changes in signal-to-noise ratio in actual non-cooperative communication scenarios, and it is difficult to obtain a large number of labeled real samples.

[0029] like Figure 5 As shown, the DCG model and U-Net network of this invention are obtained through the following steps: (1) Construct an initial simulation dataset consisting of multiple wireless interference signals, and use a data augmentation strategy based on channel impairment modeling to superimpose noise components simulating channel impairment characteristics onto the initial simulation dataset to obtain a simulation dataset. Divide the simulation dataset into a training set, a validation set, and a test set in a ratio of 8:2:1. The samples in the training set and the validation set all contain the real interference category label corresponding to the sample. ; In order to solve the problem of model overfitting caused by the scarcity of real samples, this invention uses simulation software such as MATLAB or GNURadio to generate multiple wireless interference signals containing various modulation types. All wireless interference signals form the initial simulation dataset. Each wireless interference signal is an IQ complex sequence with a length of 1024, and the initial signal-to-noise ratio is set to 10dB. Considering that a robust model cannot be trained by relying solely on a small number of high signal-to-noise ratio samples (especially in low signal-to-noise ratio environments), this invention designs a data augmentation strategy based on channel impairment modeling, aiming to simulate a real additive white Gaussian noise (AWGN) channel and push the decision boundary toward the optimum.

[0030] Specifically, this step includes the following sub-steps: (1-1) Use simulation software to generate multiple wireless interference signals containing various modulation types. All wireless interference signals form the initial simulation dataset. (1-2) Construct standard Gaussian white noise sequences containing different signal-to-noise ratio levels ,in This represents the first sequence in the standard Gaussian white noise sequence library. A standard Gaussian white noise sequence, This indicates the size of the standard Gaussian white noise sequence library; (1-3) For each wireless interference signal in the initial simulation dataset obtained in step (1-1) In this regard, the standard Gaussian white noise sequence obtained from step (1-2) Randomly select a noise sample ,in This indicates the sequence number of the selected noise sample in the standard Gaussian white noise sequence library; (1-4) The noise samples obtained in step (1-3) Perform a random permutation operation to obtain the permuted noise samples. ; (1-5) The noise samples obtained after the permutation operation in step (1-4) Superimposed on each wireless interference signal Above, to obtain the corresponding enhanced samples. ; Specifically, this step uses the following formula: ; (1-6) All the augmented samples obtained The data is then aggregated to obtain the simulation dataset. (1-7) Divide the simulation dataset obtained in step (1-6) into a ratio of 8:2:1 to obtain a training set, a validation set, and a test set. The samples in the training set and the validation set all contain the corresponding real interference category labels. ; In this way, embodiments of the present invention expand the dataset so that the training data covers a continuous change from strong signals to weak signals, forcing the classifier to learn a decision boundary that approximates the maximum likelihood estimate.

[0031] (2) Use the DCG model to perform feature extraction and condition construction on the training set obtained in step (1) to obtain condition-guided information. ; Specifically, this step includes the following sub-steps: (2-1) Obtain each sample and its corresponding true class label from the training set obtained in step (1). The sample is then transformed into a two-dimensional time-frequency graph using the Short-time Fourier transform (STFT). To extract more discriminative features, this embodiment of the invention converts the enhanced IQ signal using an STFT. A two-dimensional time-frequency plot. Considering that the wireless interference signal to be detected contains both macroscopic modulation textures and microscopic transient features, this embodiment of the invention constructs a DCG model, such as... Figure 4 As shown; (2-2) Input the two-dimensional time-frequency graph obtained in step (2-1) into the global flow branch of the DCG model. It first uses a backbone network (such as ResNet-18) to extract deep features, and then inputs these deep features into the DCG model. Convolution is performed to generate a global saliency map, and finally, global average pooling is applied to this global saliency map to obtain the global prior vector. (It can reflect the overall energy distribution of the signal); (2-3) Input the two-dimensional time-frequency graph obtained in step (2-1) into the local flow branch of the DCG model, and locate and obtain multiple regions of interest (ROI) patches based on the global saliency map (each ROI patch size is...). The algorithm uses a local encoder (a 3-layer CNN network with shared weights) to extract features from all ROI patches, obtains the weights of each feature through a gating attention mechanism, and then weights from all features are weighted and fused to obtain a local prior vector. (It focuses on microscopic transient characteristics); In this step, the following formula is used to calculate each feature (i.e., the first feature). Weights of each feature: ; in, , They represent the first The and the first Features of an ROI patch ∈[1, the total number of ROI patches obtained], These are learnable weight parameters. (2-4) Input the two-dimensional time-frequency graph obtained in step (2-1) into the backbone network to obtain the time-frequency feature vector. ; (2-5) The global prior vector obtained in step (2-2) The local prior vector obtained in step (2-3) and the time-frequency feature vector obtained in step (2-4) Perform splicing to obtain conditional guidance information. (It is used to guide the reverse denoising process of the diffusion model); (3) Using the training set obtained in step (1) and the conditional guidance information obtained in step (2), perform end-to-end training on the DCG model and the U-Net network to obtain the trained DCG model and U-Net network. This step specifically includes the following sub-steps: (3-1) Label the true interference category corresponding to each sample in the training set obtained in step (1). Input the forward diffusion module in the U-Net network and at random time steps Label the real interference category Add Gaussian noise To obtain random time steps Corresponding noise variables ; Specifically, acquiring noise The following formula is used: ; in It is the product of preset noise scheduling parameters; (3-2) The noise variable obtained in step (4-1) Random time step and the conditional guidance information obtained in step (2) The input is processed by the U-Net network for feature fusion and inverse prediction to obtain prediction noise. ; (3-3) The Gaussian noise obtained from step (3-1) The predicted noise obtained in step (3-2) Mean square error is calculated to obtain the noise estimation loss. ; (3-4) Based on the global prior vector obtained in step (2-2) The local prior vector obtained in step (2-3) Obtain the maximum mean discrepancy (MMD) regularization loss. ; Specifically, the MMD regularization loss The calculation formula is as follows: ; in, and They represent the first and second training sets, respectively. Seed granularity (i.e., global granularity) The global prior vector corresponding to the training sample and the first training sample Seed size (i.e., local grain size) under the first Local prior vectors corresponding to each training sample; and These represent the obtained global prior vectors, respectively. The total number and local prior vectors The total number; This represents the kernel mapping function, used to map feature vectors to a high-dimensional reproducing kernel Hilbert space (RKHS). This represents norm operations in RKHS; (3-5) The noise estimation loss obtained in step (3-3) The MMD regularization loss obtained in step (3-4) Perform a weighted summation to obtain the total loss value. ; (3-6) For each sample in the training set obtained in step (1-7), use the gradient descent algorithm based on the total loss value obtained in step (3-5). All learnable weight parameters in the DCG module and U-Net network are synchronously iteratively updated until the DCG module and U-Net network reach a preset number of iterations (500 times in this embodiment of the invention), and the optimal parameters of the DCG module and U-Net network at this time are obtained, thereby obtaining the initially trained DCG module and U-Net network. (3-7) Use the test set obtained in step (1-7) to evaluate the performance of the DCG module and U-Net network initially trained in step (3-6) until the predicted class matches the true class label. The detection accuracy is optimized until the DCG module and U-Net network are finally trained.

[0032] The advantage of this step (3) is that it uses a generative diffusion reconstruction mechanism to replace the traditional discriminative decision boundary. By jointly optimizing the noise estimation loss and the distribution regularization loss, the model can adaptively recover high-purity interference category features from the noise during the inference stage, which greatly improves the robustness of detection.

[0033] This invention constructs an initial simulation dataset of wireless interference signals and expands it using a data augmentation strategy based on channel impairment modeling. The augmented signal is converted into a time-frequency graph, and a dual-granularity conditional guidance module is constructed to extract global and local priors in parallel. A denoising diffusion probability model based on the U-Net architecture is built, establishing an inverse mapping relationship from Gaussian noise distribution to the distribution of real interference signal categories. The model is trained using joint optimization noise estimation loss and maximum mean difference regularization loss, and the final detection result is output through inverse denoising sampling. Compared with existing technologies, this invention effectively solves the problem of scarce real interference samples through physical layer data augmentation, pushing the classifier's decision boundary to approach the optimal position of maximum likelihood estimation. The dual-granularity conditional guidance mechanism takes into account both the macroscopic contour and microscopic transient features of the signal, overcoming the shortcomings of existing models in distinguishing complex modulated signals. The generative denoising mechanism unique to the diffusion model, combined with MMD regularization constraints, significantly improves the detection robustness and accuracy in low signal-to-noise ratio and complex electromagnetic environments.

[0034] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart interference detection method based on a data-augmented dual-granularity diffusion model, characterized in that, Includes the following steps: Step 1: Acquire the wireless interference signal to be detected, and perform time-frequency transformation processing on the wireless interference signal to be detected to obtain the corresponding time-frequency feature map; Step 2: Analyze the time-frequency characteristic map of the wireless interference signal to be detected obtained in Step 1. Input a pre-trained dual-granularity conditionally guided DCG model to obtain conditionally guided information containing multi-scale features of the signal. ; Step 3: Input the conditional guidance information obtained in Step 2 into the pre-trained U-Net network for inverse denoising and sampling processing, and select the category corresponding to the maximum value from the multiple normalized processing results as the final detection result.

2. The intelligent interference detection method based on a data-augmented dual-granularity diffusion model according to claim 1, characterized in that, Step one includes the following sub-steps: (A1) Use radio frequency receiving equipment to collect the wireless interference signal to be detected in the environment, and perform down-conversion, filtering and analog-to-digital conversion processing on the wireless interference signal to be detected in sequence to obtain the in-phase orthogonal IQ complex sequence data of the wireless interference signal to be detected. (A2) Perform short-time Fourier transform (STFT) processing on the IQ complex sequence data of the wireless interference signal to be detected obtained in step (A1) to obtain the time-frequency feature map corresponding to the wireless interference signal to be detected. It is a two-dimensional time-frequency characteristic map.

3. The intelligent interference detection method based on a data-augmented dual-granularity diffusion model according to claim 1 or 2, characterized in that, Step two includes the following sub-steps: (B1) Input the time-frequency feature map of the wireless interference signal to be detected obtained in step one into the global flow branch of the DCG model to extract the deep features using the backbone network, and perform global average pooling on the deep features to obtain a global prior vector reflecting the overall energy distribution of the wireless interference signal to be detected. ; (B2) Input the time-frequency feature map of the wireless interference signal to be detected obtained in step one into the local flow branch of the DCG model to crop multiple regions of interest (ROI) patches from the time-frequency feature map, and use a local encoder to extract features from all ROI patches. Then, use gated attention weights to fuse all features from all ROI patches to obtain a local prior vector. ; (B3) The global prior vector obtained in step (B1) The local prior vector obtained in step (B2) The system concatenates all feature vectors from the time-frequency feature map obtained in step one to obtain conditional guidance information containing both global and local prior vectors. .

4. The intelligent interference detection method based on a data-augmented dual-granularity diffusion model according to any one of claims 1 to 3, characterized in that, Step 3 includes the following sub-steps: (C1) Set the iteration counter Preset acceleration steps ; (C2) Judgment If the number of iterations is less than the preset threshold, proceed to step (C6); otherwise, proceed to step (C3). (C3) will be the first noise variables in the next iteration Iteration counter and the conditional guidance information obtained in step two. Input the U-Net network to obtain the predicted noise Among them, noise variables The initial values ​​are derived from a standard normal distribution; (C4) Predicted noise obtained from step (C3) Get the noise variables from the previous time step : ； in, and These are predefined noise scheduling parameters. The cumulative product, It is a random variable that follows a standard normal distribution. It is the variance term; (C5) Set the iteration counter , and return to step (C2); (C6) The prediction vector at time step 0 obtained after the iteration is completed. Softmax normalization is performed, and the category corresponding to the maximum value among the multiple normalization results is selected as the final detection result.

5. The intelligent interference detection method based on a data-augmented dual-granularity diffusion model according to claim 4, characterized in that, The DCG model and U-Net network were trained using the following steps: (1) Construct an initial simulation dataset consisting of multiple wireless interference signals, and use a data augmentation strategy based on channel impairment modeling to superimpose noise components simulating channel impairment features onto the initial simulation dataset to obtain a simulation dataset. Divide the simulation dataset into a training set, a validation set, and a test set in a ratio of 8:2:

1. The samples in the training set and the validation set all contain the real interference category label corresponding to the sample. ; (2) Use the DCG model to perform feature extraction and condition construction on the training set obtained in step (1) to obtain condition-guided information. ; (3) Using the training set obtained in step (1) and the conditional guidance information obtained in step (2), perform end-to-end training on the DCG model and the U-Net network to obtain the trained DCG model and U-Net network.

6. The intelligent interference detection method based on a data-augmented dual-granularity diffusion model according to claim 5, characterized in that, Step (1) includes the following sub-steps: (1-1) Use simulation software to generate multiple wireless interference signals containing various modulation types. All wireless interference signals form the initial simulation dataset. (1-2) Construct standard Gaussian white noise sequences containing different signal-to-noise ratio levels ,in This represents the first sequence in the standard Gaussian white noise sequence library. A standard Gaussian white noise sequence, This indicates the size of the standard Gaussian white noise sequence library; (1-3) For each wireless interference signal in the initial simulation dataset obtained in step (1-1) In this regard, the standard Gaussian white noise sequence obtained from step (1-2) Randomly select a noise sample ,in This indicates the sequence number of the selected noise sample in the standard Gaussian white noise sequence library; (1-4) The noise samples obtained in step (1-3) Perform a random permutation operation to obtain the permuted noise samples. ; (1-5) The noise samples obtained after the permutation operation in step (1-4) Superimposed on each wireless interference signal Above, to obtain the corresponding augmented sample. : ； (1-6) All the augmented samples obtained The data is then aggregated to obtain the simulation dataset. (1-7) Divide the simulation dataset obtained in step (1-6) into a ratio of 8:2:1 to obtain a training set, a validation set, and a test set. The samples in the training set and the validation set all contain the corresponding real interference category labels. .

7. The intelligent interference detection method based on a data-augmented dual-granularity diffusion model according to claim 6, characterized in that, Step (2) includes the following sub-steps: (2-1) Obtain each sample and its corresponding true class label from the training set obtained in step (1). The sample is then converted into a two-dimensional time-frequency graph using a short-time Fourier transform (STFT). (2-2) Input the two-dimensional time-frequency graph obtained in step (2-1) into the global flow branch of the DCG model. It first uses the backbone network to extract deep features, and then inputs the deep features into the DCG model. Convolution is performed to generate a global saliency map, and finally, global average pooling is applied to this global saliency map to obtain the global prior vector. ; (2-3) Input the two-dimensional time-frequency map obtained in step (2-1) into the local flow branch of the DCG model, locate multiple regions of interest (ROI) patches based on the global saliency map, extract features from all ROI patches using a local encoder, obtain the weight of each feature through a gated attention mechanism, and then perform weighted fusion of all feature weights to obtain the local prior vector. ; (2-4) Input the two-dimensional time-frequency graph obtained in step (2-1) into the backbone network to obtain the time-frequency feature vector. ; (2-5) The global prior vector obtained in step (2-2) The local prior vector obtained in step (2-3) and the time-frequency feature vector obtained in step (2-4) Perform splicing to obtain conditional guidance information. .

8. The intelligent interference detection method based on a data-augmented dual-granularity diffusion model according to claim 7, characterized in that, In steps (2-3), each feature is calculated using the following formula, i.e., the... Weights of each feature: ； in, , They represent the first The and the first Features of an ROI patch ∈[1, the total number of ROI patches obtained], These are learnable weight parameters.

9. The intelligent interference detection method based on a data-augmented dual-granularity diffusion model according to claim 8, characterized in that, Step (3) specifically includes the following sub-steps: (3-1) Label the true interference category corresponding to each sample in the training set obtained in step (1). Input the forward diffusion module in the U-Net network and at random time steps Label the real interference category Add Gaussian noise To obtain random time steps Corresponding noise variables : ； in It is the product of preset noise scheduling parameters; (3-2) The noise variable obtained in step (4-1) Random time step and the conditional guidance information obtained in step (2) The input is processed by the U-Net network for feature fusion and inverse prediction to obtain prediction noise. ; (3-3) The Gaussian noise obtained from step (3-1) The predicted noise obtained in step (3-2) Mean square error is calculated to obtain the noise estimation loss. ; (3-4) Based on the global prior vector obtained in step (2-2) The local prior vector obtained in step (2-3) Obtain the maximum mean difference MMD regularization loss : ； in, and They represent the first and second training sets, respectively. Seed granularity (i.e., global granularity) The global prior vector corresponding to the training sample and the first training sample Seed size (i.e., local grain size) under the first Local prior vectors corresponding to each training sample; and These represent the obtained global prior vectors, respectively. The total number and local prior vectors The total number; This represents the kernel mapping function, used to map feature vectors to a high-dimensional reproducing kernel Hilbert space RKHS; This represents norm operations in RKHS; (3-5) The noise estimation loss obtained in step (3-3) The MMD regularization loss obtained in step (3-4) Perform a weighted summation to obtain the total loss value. ; (3-6) For each sample in the training set obtained in step (1-7), use the gradient descent algorithm based on the total loss value obtained in step (3-5). All learnable weight parameters in the DCG module and U-Net network are synchronously iteratively updated until the DCG module and U-Net network reach a preset number of iterations (500 times in this embodiment of the invention), and the optimal parameters of the DCG module and U-Net network at this time are obtained, thereby obtaining the initially trained DCG module and U-Net network. (3-7) Use the test set obtained in step (1-7) to evaluate the performance of the DCG module and U-Net network initially trained in step (3-6) until the predicted class matches the true class label. The detection accuracy is optimized until the DCG module and U-Net network are finally trained.

10. An intelligent interference detection system based on a data-augmented dual-granularity diffusion model, characterized in that, include: The first module is used to acquire the wireless interference signal to be detected and to perform time-frequency transformation processing on the wireless interference signal to be detected in order to obtain the corresponding time-frequency feature map. The second module is used to process the time-frequency characteristic map of the wireless interference signal to be detected obtained from the first module. Input a pre-trained dual-granularity conditionally guided DCG model to obtain conditionally guided information containing multi-scale features of the signal. ; The third module is used to input the conditional guidance information obtained from the second module into the pre-trained U-Net network for inverse denoising and sampling processing, and select the category corresponding to the maximum value from the multiple normalized processing results as the final detection result.

Images