An interference identification method based on knowledge distillation and fusion time-frequency graph

Abstract

The application discloses a kind of interference identification method based on knowledge distillation and fusion time-frequency graph, it belongs to communication interference identification technical field.The present application solves the problem that existing method is difficult to extract strong robustness representation information, and the parameter quantity and the big problem of computing overhead of identification model.The application uses time-frequency analysis technique to convert one-dimensional time domain signal into two-dimensional time-frequency image, this mode can completely retain the coupling characteristics of signal on time axis and frequency axis, to extract strong robustness representation information, to effectively improve the discrimination precision of model in low SNR environment.The application introduces improved knowledge distillation training framework, so that the teacher model guides the learning of lightweight student model, in actual identification process, the complex mapping capability of deep network can be transferred to the simplified network, greatly reducing the parameter quantity and the computing overhead of model.The application method can be applied to the field of communication interference identification technology.

Description

Technical Field

[0001] This invention belongs to the field of communication interference identification technology, specifically relating to an interference identification method based on knowledge distillation and fusion of time-frequency maps. Background Technology

[0002] With the increasing complexity of the electromagnetic environment, various forms of man-made interference and natural noise intertwine, posing a severe challenge to the signal transmission of wireless communication systems. In complex electromagnetic countermeasures, accurate anti-interference decisions rely on real-time perception of the interference environment. Therefore, interference identification, as a key link in anti-interference communication, aims to quickly and accurately determine the type of interference from complex background signals. Only by achieving high-precision interference identification can the system take targeted anti-interference measures such as frequency avoidance, power control, or code modulation optimization, thereby ensuring the robustness of the communication link.

[0003] Traditional interference identification methods often rely on manually designed expert features. However, when dealing with non-stationary and transient interference signals, they often struggle to extract robust representational information, resulting in low interference identification accuracy for such signals. While mainstream deep convolutional models, such as ResNet, perform exceptionally well in image classification, their large parameter count and computational overhead hinder their deployment in embedded communication devices or mobile terminals. In resource-constrained real-world applications, achieving model lightweighting while maintaining recognition performance is a pressing challenge in the field of communication warfare. Summary of the Invention

[0004] This invention addresses the problems of existing methods struggling to extract robust representational information and having high parameter counts and computational overhead in the identification model. It proposes an interference identification method based on knowledge distillation and fusion of time-frequency graphs.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is: an interference identification method based on knowledge distillation and fusion of time-frequency maps, the method specifically including the following steps:

[0006] Step 1: Generate a dataset containing interference signal sequence sequences of various interference types;

[0007] Step 2: Add noise to the interference signal sequences in the dataset, and then normalize the power of each noise-added interference signal sequence to obtain the processed interference signal sequence dataset.

[0008] Step 3: Process the interference signal sequences in the processed interference signal sequence dataset to obtain the fused time-frequency diagram corresponding to each interference signal sequence. Use the fused time-frequency diagram corresponding to the interference signal sequence and the interference type to form a sample. Then divide all the obtained samples into two parts: training sample set and validation sample set.

[0009] Step 4: Train the teacher model using the training sample set;

[0010] Step 5: Build a student model. Train the student model using knowledge distillation and a trained teacher model to obtain a well-trained student model.

[0011] Step 6: Receive the signal to be identified from the actual communication environment. Normalize the power of the received signal and then process the normalized signal to obtain the fused time-frequency map of the signal to be identified from the interference type. Then input the obtained fused time-frequency map into the trained student model and output the interference type identification result through the trained student model.

[0012] Furthermore, the interference types include single-tone interference, multi-tone interference, linear frequency modulation interference, partial frequency band noise interference, noise frequency modulation interference, periodic rectangular pulse interference, single-tone and noise frequency modulation composite interference, multi-tone and linear frequency modulation composite interference, multi-tone and partial frequency band noise composite interference, and partial frequency band noise and noise frequency modulation composite interference.

[0013] Furthermore, the specific process of step two is as follows:

[0014] Step 21: Add Gaussian white noise to the interference signal sequence to obtain the noisy interference signal sequence;

[0015] Step 22: Normalize the power of the noise-added interference signal sequence.

[0016] Furthermore, the specific process of step two is as follows:

[0017]

[0018] in, Indicates the first digit of the noise-added interference signal sequence. Power at each sampling point This indicates taking the absolute value. Indicates the total number of sampling points. This represents the average power across all sampling points;

[0019] Then the first noise-added interference signal sequence The normalized power of each sampling point is:

[0020]

[0021] in, Indicates the first digit of the noise-added interference signal sequence. Normalized power of each sampling point.

[0022] Furthermore, the specific process of step three is as follows:

[0023] Step 3: 1. Perform synchronous compressed wavelet transform on the interference signal sequence to obtain the first time-frequency diagram; 2. Perform bilinear time-frequency transform based on Born-Jordan distribution on the interference signal sequence to obtain the second time-frequency diagram; 3. Perform bilinear time-frequency transform based on Cui-Williams distribution on the interference signal sequence to obtain the third time-frequency diagram.

[0024] Step 32: Crop the first, second, and third time-frequency maps on the frequency axis respectively. By cropping, only the part of each time-frequency map within the communication bandwidth is retained, resulting in the cropped first, second, and third time-frequency maps.

[0025] Step 3: Perform grayscale processing on the cropped first time-frequency image, the cropped second time-frequency image, and the cropped third time-frequency image respectively to obtain the first grayscale time-frequency image, the second grayscale time-frequency image, and the third grayscale time-frequency image;

[0026] Steps 3 and 4: Normalize the first grayscale time-frequency image, the second grayscale time-frequency image, and the third grayscale time-frequency image respectively;

[0027] Step 35: Spatial dimension fusion is performed on the normalized first grayscale time-frequency image, the normalized second grayscale time-frequency image, and the normalized third grayscale time-frequency image to obtain the fused time-frequency image.

[0028] Furthermore, the specific process of steps three and four is as follows:

[0029] Map the pixels in the first grayscale time-frequency image to Within the range:

[0030]

[0031] In the formula, This represents the maximum grayscale value in the first grayscale time-frequency graph. This represents the minimum grayscale value in the first grayscale time-frequency plot. Represents the pixel points in the first grayscale time-frequency image. grayscale value, This represents the pixel in the first grayscale time-frequency image after normalization. The grayscale value.

[0032] Furthermore, in the fused time-frequency map, the R channel is the first gray-level time-frequency map after normalization, the G channel is the third gray-level time-frequency map after normalization, and the B channel is the second gray-level time-frequency map after normalization.

[0033] Furthermore, the teacher model inserts a fully connected layer before the classification layer of the ResNet-18 model.

[0034] Furthermore, the working process of the student model is as follows:

[0035] Within the student model, the input fused time-frequency map sequentially passes through the first CNN convolutional block, the second CNN convolutional block, the third CNN convolutional block, the global average pooling layer, the fully connected layer, and the SoftMax activation function layer. Each CNN convolutional block sequentially includes a depthwise convolutional layer, a pointwise convolutional layer, a BN layer, a ReLU activation function layer, and a max pooling layer.

[0036] Furthermore, the process of training the student model through knowledge distillation and a pre-trained teacher model specifically involves:

[0037] Step 1: Feed the fused time-frequency image into the student model for forward propagation, and then calculate the loss between the interference type identification result output by the student model and the actual interference type label. ;

[0038] Step 2: Feed the fused time-frequency image into the trained teacher model to obtain the feature vector corresponding to the fused time-frequency image;

[0039] Step 3: Calculate the feature vector output by the trained teacher model. The feature vector output by the student model Losses between :

[0040]

[0041] in, This represents the first feature vector in the output of the trained teacher model. One element, This represents the first feature vector in the output of the student model. One element, This indicates the calculation of the L2 norm. This represents the total number of elements in the eigenvector;

[0042] Step 4, according to and Calculate total loss :

[0043]

[0044] in, Indicates multiplication. and All are weighting coefficients; It is the input fused time-frequency diagram. This represents the parameters of the current student model. This indicates the label representing the actual interference type corresponding to the input fused time-frequency map. This represents the interference type identification result output by the student model;

[0045] Calculate total loss The gradients of all parameters of the student model are then used to update the student model parameters through backpropagation:

[0046]

[0047] in, Indicates total loss gradient, Indicates the learning rate. This represents the updated student model parameters;

[0048] Training continues until the termination condition is met, at which point a well-trained student model is obtained. The termination condition is that condition (1), condition (2), or condition (3) is met.

[0049] Condition (1) The number of training rounds reaches the preset maximum number of rounds;

[0050] Condition (2) The total loss function value on the validation set has not decreased for several consecutive rounds;

[0051] Condition (3): The learning rate drops to the preset minimum value.

[0052] The beneficial effects of this invention are:

[0053] This invention leverages the powerful automatic feature extraction capabilities of deep learning, employing time-frequency analysis to transform one-dimensional time-domain signals into two-dimensional time-frequency images. This method fully preserves the coupled features of the signal along both the time and frequency axes. By fusing multi-dimensional time-frequency information, it extracts robust representational information, providing richer raw input to deep convolutional neural networks and effectively improving the model's discrimination accuracy in low noise-to-interference ratio environments. Furthermore, this invention enhances the model's ability to perceive weak interference features by improving the generation method of time-frequency data. By introducing an improved knowledge distillation training framework, the teacher model guides the learning of a lightweight student model. In actual recognition, the complex mapping capabilities of deep networks can be transferred to a simplified network, significantly reducing the number of model parameters and computational overhead. Moreover, the extracted features effectively compensate for the performance loss caused by parameter reduction, ultimately achieving a balance between recognition accuracy and computational efficiency. Attached Figure Description

[0054] Figure 1 This is a flowchart of an interference identification method based on knowledge distillation and fusion of time-frequency maps according to the present invention;

[0055] Figure 2 This is a flowchart illustrating the generation of the interference signal dataset;

[0056] Figure 3 These are the recognition results of various commonly used classification models on the fused time-frequency map;

[0057] Figure 4 This is a flowchart for improving knowledge distillation training;

[0058] Figure 5 This is a comparison chart of the accuracy of knowledge distillation without knowledge distillation, traditional knowledge distillation based on soft tags, and the knowledge distillation method based on feature vectors of this invention. Detailed Implementation

[0059] Specific implementation method one: Combining Figure 1 and Figure 2 This embodiment describes an interference identification method based on knowledge distillation and time-frequency map fusion. The method specifically includes the following steps:

[0060] Step 1: Generate a dataset of interference signal sequences containing various interference types, including single-tone interference, multi-tone interference, linear frequency modulation interference, partial frequency band noise interference, noise frequency modulation interference, periodic rectangular pulse interference, single-tone and noise frequency modulation composite interference, multi-tone and linear frequency modulation composite interference, multi-tone and partial frequency band noise composite interference, and partial frequency band noise and noise frequency modulation composite interference, for a total of 10 interference types.

[0061] The communication system parameters were set as follows: communication bandwidth of 10MHz, sampling frequency of 30MHz, number of sampling points of 1024, and step size of 5dB. The parameters for various interference signals were set as follows:

[0062] The interference frequency of a single-tone interference follows Uniform distribution of MHz;

[0063] The number of interference frequencies in multi-tone interference follows The uniform distribution and the interference frequencies at each frequency point follow the rules of the system. Uniform distribution of MHz;

[0064] The sweep bandwidth of linear frequency modulation interference follows Uniform distribution of MHz, sweep period follows Uniform distribution, initial frequency follows Uniform distribution of MHz;

[0065] The noise bandwidth of partial frequency band noise interference follows Uniform distribution of MHz, noise power follows W is uniformly distributed, and its center frequency follows a constant. Uniform distribution of MHz;

[0066] The noise bandwidth of the frequency modulation interference is fixed at 1 MHz, and the noise power follows the law of... W is uniformly distributed, and its initial frequency follows a constant. Uniform distribution of MHz, frequency modulation index follows Uniform distribution of MHz / s;

[0067] The pulse period of periodic rectangular pulse interference follows Uniform distribution, duty cycle follows The uniform distribution.

[0068] Step 2: Add noise to the interference signal sequences in the dataset, and then normalize the power of each noise-added interference signal sequence to obtain the processed interference signal sequence dataset.

[0069] Specifically, for any one of the interference signal sequences in the dataset:

[0070] Step 21: Add Gaussian white noise to the interference signal sequence to obtain the noisy interference signal sequence;

[0071] Adding noise can alter the JNR (interference-to-noise ratio) of an interfering signal sequence. to Each interval We define one JNR level, with a total of 11 JNR levels. The sample size (conventionally referred to as the unit sample size of the training set) for each type of interference signal sequence under each JNR is 200. For 10 types of interference and 11 JNR levels, the total sample size in the training set is... The number of samples for each type of interference signal sequence in the validation set is 100 under each JNR, and the total number of samples is 11,000.

[0072] Step 22: Normalize the power of the noise-added interference signal sequence, specifically as follows:

[0073]

[0074] in, Indicates the first digit of the noise-added interference signal sequence. Power at each sampling point This indicates taking the absolute value. Indicates the total number of sampling points. This represents the average power across all sampling points;

[0075] Then the first noise-added interference signal sequence The normalized power of each sampling point is:

[0076]

[0077] in, Indicates the first digit of the noise-added interference signal sequence. Normalized power of each sampling point.

[0078] Step 3: Process the interference signal sequences in the processed interference signal sequence dataset to obtain the fused time-frequency diagram corresponding to each interference signal sequence. Use the fused time-frequency diagram corresponding to the interference signal sequence and the interference type to form a sample. Then divide all the obtained samples into two parts: training sample set and validation sample set.

[0079] For any given interference signal sequence, specifically:

[0080] Step 3: 1. Perform Synchronous Compressed Wavelet Transform (SWT) on the interference signal sequence to obtain the first time-frequency diagram; 2. Perform Bilinear Time-Frequency Transform based on Born-Jordan distribution (BJD) on the interference signal sequence to obtain the second time-frequency diagram; 3. Perform Bilinear Time-Frequency Transform based on Cui-Williams distribution (CWD) on the interference signal sequence to obtain the third time-frequency diagram.

[0081] The three time-frequency analysis methods correspond to different time-frequency analysis characteristics and can characterize the time-frequency features of a signal from multiple perspectives;

[0082] Step 3.2: Crop the first, second, and third time-frequency maps on the frequency axis respectively. By cropping, only the portion of each time-frequency map within the communication bandwidth is retained, resulting in the cropped first, second, and third time-frequency maps. On the time axis, the time-frequency maps are not cropped, and the values ​​of all sampling points are retained.

[0083] According to the Nyquist sampling theorem, the maximum value of each time-frequency graph on the frequency axis is obtained through time-frequency analysis. ,in, The sampling frequency is used. On the frequency axis, only the time-frequency plot is retained within the communication bandwidth. The inner part, the rest and Partial removal, of which, and These represent the lower and upper boundaries of the communication bandwidth, respectively.

[0084] Step 3: Perform grayscale processing on the cropped first time-frequency image, the cropped second time-frequency image, and the cropped third time-frequency image respectively to obtain the first grayscale time-frequency image, the second grayscale time-frequency image, and the third grayscale time-frequency image;

[0085] By converting the time-frequency graph to grayscale, it can be embedded into a single channel. However, different time-frequency analysis methods and different types of interference signals can cause significant differences in the grayscale values ​​of the grayscale time-frequency graph. To avoid feature imbalance caused by differences in grayscale values, it is necessary to perform a normalization operation on the grayscale time-frequency graph.

[0086] Steps 3 and 4: Normalize the first grayscale time-frequency image, the second grayscale time-frequency image, and the third grayscale time-frequency image respectively;

[0087] Taking the first grayscale time-frequency image as an example, specifically: the pixels in the first grayscale time-frequency image are mapped to... Within the range:

[0088]

[0089] In the formula, This represents the maximum grayscale value in the first grayscale time-frequency graph. This represents the minimum grayscale value in the first grayscale time-frequency plot. Represents the pixel points in the first grayscale time-frequency image. grayscale value, This represents the pixel in the first grayscale time-frequency image after normalization. grayscale value;

[0090] Step 35: Spatial dimension fusion is performed on the normalized first gray-level time-frequency image, the normalized second gray-level time-frequency image, and the normalized third gray-level time-frequency image to obtain a fused time-frequency image. In the fused time-frequency image, the R channel is the normalized first gray-level time-frequency image, the G channel is the normalized third gray-level time-frequency image, and the B channel is the normalized second gray-level time-frequency image.

[0091] In step three, a weighted fusion strategy with each channel having a weighting coefficient of 1 is used to stack the three normalized grayscale time-frequency images in the spatial dimension, forming a dimension of The fusion matrix. This fusion method does not lose any of the original features obtained by any transformation, and retains the original feature information of each transformation method. This allows the subsequent deep convolutional neural network to automatically adjust the attention of different channels through learning, thereby achieving deep fusion and utilization of heterogeneous time-frequency features.

[0092] The annotation information corresponding to each fused time-frequency plot is as follows: Where JNR is the JNR value corresponding to the fused time-frequency map, and index is the number of the fused time-frequency map under that JNR. For the training set file format: the first-level folder contains the interference signal type; the second-level folder contains the interference-to-noise ratio level; the third-level folder stores the fused time-frequency map generated after time-frequency transformation and multi-channel fusion processing, which facilitates data management and traceability during model training.

[0093] Step 4: Perform supervised end-to-end training on the teacher model using the training sample set. The training parameters are configured as follows: batch size is set to 16; stochastic gradient descent (SGD) is used as the optimizer; the maximum number of iterations is set to 50; and the initial learning rate is set to 0.005.

[0094] During the training of the teacher model, in each iteration, the model automatically extracts the time-frequency features of the signal through multi-layer convolutional operations and generates feature vectors. The cross-entropy loss function is used to calculate the deviation between the predicted results and the true labels, and the network weights are updated in reverse using stochastic gradient descent. Through multiple iterations, the model's feature extraction capability is gradually improved. Finally, after 50 rounds of training, the model weights with the best performance on the validation set are selected as the final teacher model weights. The feature representations learned by these weights will serve as the learning objective for the student model, used in the subsequent distillation process of the lightweight student model.

[0095] This invention employs a ResNet-18 residual network as the basic architecture of the teacher model. Its residual connection mechanism can effectively transfer gradient information, mitigating the gradient vanishing problem in deep networks while maintaining feature extraction depth. Furthermore, to achieve feature alignment between the teacher and student models during the distillation process, the original ResNet-18 residual network was structurally fine-tuned.

[0096] This invention employs a teacher model that inserts a fully connected layer before the classification layer of the ResNet-18 model. In other words, the ResNet-18 model with the inserted fully connected layer serves as the teacher model. The main function of this inserted fully connected layer is to uniformly map the deep abstract features extracted by the network into 256-dimensional feature vectors, achieving standardized feature representation. This dimensionality reduction and unification process ensures that the feature dimensions output by the teacher model remain consistent with those of the lightweight student model, thus laying the foundation for subsequent calculations of feature mapping loss and knowledge transfer.

[0097] Step 5: Build a student model. Train the student model using knowledge distillation and a trained teacher model to obtain a well-trained student model.

[0098] Specifically, the structure of the student model is shown in Table 1:

[0099] Table 1 Parameters of the Lightweight Student Network Model

[0100]

[0101] Table 1 provides a detailed description of the structural parameters of each layer in the student model. (224,224,3) represents a fused time-frequency map of size 224×224 with 3 channels as input; DepthwiseConv2D represents a depthwise convolution operation with a kernel size of 3×3 and a stride of 1, keeping the number of channels constant, used for spatial feature extraction within each channel; PointwiseConv2D represents a pointwise convolution operation with a kernel size of 1×1, expanding the number of channels to 32, used for feature fusion between channels; BN represents batch normalization, used to accelerate model convergence; ReLU represents the activation function, used to introduce nonlinearity; Maxpool2D represents a 2D max pooling operation with a kernel size of 2×2; GAP represents global average pooling, used to reduce the spatial dimension of the feature map; FC represents a fully connected layer, used to output a 256-dimensional feature vector; SoftMax represents the activation function of the output layer.

[0102] In the student model, the input fused time-frequency map passes through the first CNN convolutional block, the second CNN convolutional block, the third CNN convolutional block, the global average pooling layer, the fully connected layer, and the SoftMax activation function layer in sequence. Each CNN convolutional block includes a depthwise convolutional layer, a pointwise convolutional layer, a BN layer, a ReLU activation function layer, and a max pooling layer in sequence.

[0103] like Figure 4 As shown, the student model is trained using knowledge distillation and a pre-trained teacher model, specifically as follows:

[0104] Step 1: Feed the fused time-frequency image into the student model for forward propagation, and then calculate the loss between the interference type identification result output by the student model and the actual interference type label. The fused time-frequency map labels have been converted into one-hot encoded form during reading, so the actual labels are hard labels.

[0105] Step 2: Feed the fused time-frequency map into the trained teacher model to obtain the feature vector corresponding to the fused time-frequency map (i.e., the output of the added fully connected layer).

[0106] Step 3: Calculate the feature vector output by the trained teacher model. The feature vector output by the student model Losses between ,here:

[0107]

[0108] in, This represents the first feature vector in the output of the trained teacher model. One element, This represents the first feature vector in the output of the student model. One element, This indicates the calculation of the L2 norm. This represents the total number of elements in the feature vector. The feature vectors output by the trained teacher model and student model are both 256-dimensional.

[0109] Step 4, according to and Calculate total loss :

[0110]

[0111] in, Indicates multiplication. and All are weighting coefficients, and the range of values ​​for the weighting coefficients is [missing information]. ,and ; It is the input fused time-frequency diagram. This represents the parameters of the current student model. This indicates the label representing the actual interference type corresponding to the input fused time-frequency map. This represents the interference type identification result output by the student model. This represents the SoftMax activation function. This represents the regression result (logits) of the last layer before the SoftMax activation function in the student model.

[0112] Calculate total loss The gradients of all parameters of the student model are then used to update the student model parameters through backpropagation:

[0113]

[0114] in, Indicates total loss gradient, Indicates the learning rate. This represents the updated student model parameters;

[0115] Training continues until the termination condition is met, at which point a well-trained student model is obtained. The termination condition is that condition (1), condition (2), or condition (3) is met.

[0116] Condition (1) The number of training epochs reaches the preset maximum number of epochs.

[0117] Condition (2) The total loss function value on the validation set has not decreased for several consecutive rounds (early stopping strategy);

[0118] Condition (3): The learning rate drops to the preset minimum value.

[0119] After training the student model, the task of identifying the interference type of unknown signals can be completed using only the trained lightweight student model, without relying on the ResNet-18 teacher model. The specific implementation steps are as follows:

[0120] Step Six: Receive the interference type to be identified signal captured in the actual communication environment. Normalize the power of the received signal, and then process the normalized signal to obtain a fused time-frequency map of the interference type to be identified. Input the obtained fused time-frequency map into a trained student model. The trained student model outputs the interference type identification result. That is, after the convolutional layer of the student model, the feature data enters the global average pooling layer (GAP) and the fully connected layer (FC), outputting a 256-dimensional feature vector. Finally, this vector is transformed into a probability distribution vector through the SoftMax function. This vector represents the probability that the current signal belongs to one of 10 interference types. The category corresponding to the highest probability value is selected as the final interference type identification result for the unknown signal.

[0121] In step six, the method for processing the power-normalized signal is the same as in the training process:

[0122] (1) Perform synchronous compressed wavelet transform (SWT) on the power-normalized signal, perform bilinear time-frequency transform based on Born-Jordan distribution (BJD) on the power-normalized signal, and perform bilinear time-frequency transform based on Cui-Williams distribution (CWD) on the power-normalized signal to generate three original time-frequency diagrams.

[0123] (2) Crop the time-frequency graph on the frequency axis, keeping only the part within the communication bandwidth and removing the rest;

[0124] (3) The three cropped time-frequency images are converted to grayscale, and normalized according to their maximum and minimum grayscale values ​​to map the pixel values ​​to the range of [0,255].

[0125] Using a three-channel structure similar to RGB, the normalized SWT time-frequency map is embedded into the R channel, the CWD time-frequency map is embedded into the G channel, and the BJD time-frequency map is embedded into the B channel. A multi-channel fusion matrix with shape H×W×3 is constructed using an equal weight mapping strategy.

[0126] Simulation study:

[0127] The method of the present invention was simulated and verified using a fused time-frequency graph dataset. The simulation parameters are shown in Table 2.

[0128] Table 2 Simulation parameters of the distillation model

[0129]

[0130] Comparing the traditional knowledge distillation method with the method of this invention, the accuracy performance is as follows: Figure 5 As shown in the figure, the distilled model exhibits better performance than the undistilled model. The improved knowledge distillation method of this invention outperforms traditional knowledge distillation methods because traditional methods use soft labels, while the improved method uses feature vectors, which have higher dimensionality and more fine-grained knowledge. Furthermore, based on the obtained fused time-frequency graph, the recognition accuracy of the teacher model of this invention was compared with other commonly used classification models. The comparison results are as follows: Figure 3 As shown, therefore, this invention selects the ResNet-18 model as the teacher model.

[0131] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An interference identification method based on knowledge distillation and fusion of time-frequency maps, characterized in that, The method specifically includes the following steps: Step 1: Generate a dataset containing interference signal sequence sequences of various interference types; Step 2: Add noise to the interference signal sequences in the dataset, and then normalize the power of each noise-added interference signal sequence to obtain the processed interference signal sequence dataset. Step 3: Process the interference signal sequences in the processed interference signal sequence dataset to obtain the fused time-frequency diagram corresponding to each interference signal sequence. Use the fused time-frequency diagram corresponding to the interference signal sequence and the interference type to form a sample. Then divide all the obtained samples into two parts: training sample set and validation sample set. Step 4: Train the teacher model using the training sample set; Step 5: Build a student model. Train the student model using knowledge distillation and a trained teacher model to obtain a well-trained student model. Step 6: Receive the signal to be identified from the actual communication environment. Normalize the power of the received signal and then process the normalized signal to obtain the fused time-frequency map of the signal to be identified from the interference type. Then input the obtained fused time-frequency map into the trained student model and output the interference type identification result through the trained student model.

2. The interference identification method based on knowledge distillation and time-frequency graph fusion according to claim 1, characterized in that, The interference types include single-tone interference, multi-tone interference, linear frequency modulation interference, partial frequency band noise interference, noise frequency modulation interference, periodic rectangular pulse interference, single-tone and noise frequency modulation composite interference, multi-tone and linear frequency modulation composite interference, multi-tone and partial frequency band noise composite interference, and partial frequency band noise and noise frequency modulation composite interference.

3. The interference identification method based on knowledge distillation and time-frequency graph fusion according to claim 2, characterized in that, The specific process of step two is as follows: Step 21: Add Gaussian white noise to the interference signal sequence to obtain the noisy interference signal sequence; Step 22: Normalize the power of the noise-added interference signal sequence.

4. The interference identification method based on knowledge distillation and time-frequency graph fusion according to claim 3, characterized in that, The specific process of step two is as follows: in, Indicates the first digit of the noise-added interference signal sequence. Power at each sampling point This indicates taking the absolute value. Indicates the total number of sampling points. This represents the average power across all sampling points; Then the first noise-added interference signal sequence The normalized power of each sampling point is: in, Indicates the first digit of the noise-added interference signal sequence. Normalized power of each sampling point.

5. The interference identification method based on knowledge distillation and time-frequency graph fusion according to claim 4, characterized in that, The specific process of step three is as follows: Step 3:

1. Perform synchronous compressed wavelet transform on the interference signal sequence to obtain the first time-frequency diagram; 2. Perform bilinear time-frequency transform based on Born-Jordan distribution on the interference signal sequence to obtain the second time-frequency diagram; 3. Perform bilinear time-frequency transform based on Cui-Williams distribution on the interference signal sequence to obtain the third time-frequency diagram. Step 32: Crop the first, second, and third time-frequency maps on the frequency axis respectively. By cropping, only the part of each time-frequency map within the communication bandwidth is retained, resulting in the cropped first, second, and third time-frequency maps. Step 3: Perform grayscale processing on the cropped first time-frequency image, the cropped second time-frequency image, and the cropped third time-frequency image respectively to obtain the first grayscale time-frequency image, the second grayscale time-frequency image, and the third grayscale time-frequency image; Steps 3 and 4: Normalize the first grayscale time-frequency image, the second grayscale time-frequency image, and the third grayscale time-frequency image respectively; Step 35: Spatial dimension fusion is performed on the normalized first grayscale time-frequency image, the normalized second grayscale time-frequency image, and the normalized third grayscale time-frequency image to obtain the fused time-frequency image.

6. The interference identification method based on knowledge distillation and fusion of time-frequency maps according to claim 5, characterized in that, The specific process of steps three and four is as follows: Map the pixels in the first grayscale time-frequency image to Within the range: In the formula, This represents the maximum grayscale value in the first grayscale time-frequency graph. This represents the minimum grayscale value in the first grayscale time-frequency plot. Represents the pixel points in the first grayscale time-frequency image. grayscale value, This represents the pixel in the first grayscale time-frequency image after normalization. The grayscale value.

7. The interference identification method based on knowledge distillation and time-frequency graph fusion according to claim 6, characterized in that, In the fused time-frequency image, the R channel is the first gray-level time-frequency image after normalization, the G channel is the third gray-level time-frequency image after normalization, and the B channel is the second gray-level time-frequency image after normalization.

8. The interference identification method based on knowledge distillation and time-frequency graph fusion according to claim 7, characterized in that, The teacher model inserts a fully connected layer before the classification layer of the ResNet-18 model.

9. The interference identification method based on knowledge distillation and fusion of time-frequency maps according to claim 8, characterized in that, The working process of the student model is as follows: Within the student model, the input fused time-frequency map sequentially passes through the first CNN convolutional block, the second CNN convolutional block, the third CNN convolutional block, the global average pooling layer, the fully connected layer, and the SoftMax activation function layer. Each CNN convolutional block sequentially includes a depthwise convolutional layer, a pointwise convolutional layer, a BN layer, a ReLU activation function layer, and a max pooling layer.

10. The interference identification method based on knowledge distillation and fusion of time-frequency maps according to claim 9, characterized in that, The process of training the student model through knowledge distillation and a pre-trained teacher model is as follows: Step 1: Feed the fused time-frequency image into the student model for forward propagation, and then calculate the loss between the interference type identification result output by the student model and the actual interference type label. ; Step 2: Feed the fused time-frequency image into the trained teacher model to obtain the feature vector corresponding to the fused time-frequency image; Step 3: Calculate the feature vector output by the trained teacher model. The feature vector output by the student model Losses between : in, This represents the first feature vector in the output of the trained teacher model. One element, This represents the first feature vector in the output of the student model. One element, This indicates the calculation of the L2 norm. This represents the total number of elements in the eigenvector; Step 4, according to and Calculate total loss : in, Indicates multiplication. and All are weighting coefficients; It is the input fused time-frequency diagram. This represents the parameters of the current student model. This indicates the label representing the actual interference type corresponding to the input fused time-frequency map. This represents the interference type identification result output by the student model; Calculate total loss The gradients of all parameters of the student model are then used to update the student model parameters through backpropagation: in, Indicates total loss gradient, Indicates the learning rate. This represents the updated student model parameters; Training continues until the termination condition is met, at which point a well-trained student model is obtained. The termination condition is that condition (1), condition (2), or condition (3) is met. Condition (1) The number of training rounds reaches the preset maximum number of rounds; Condition (2) The total loss function value on the validation set has not decreased for several consecutive rounds; Condition (3): The learning rate drops to the preset minimum value.

Images