An interference pattern recognition method based on threshold denoising

Abstract

The application discloses a kind of interference mode identification methods based on threshold denoising, comprising: based on the baseband signal of interference signal, obtain preset format interference mode data;Using the interference mode identification model trained, the convolution processing of preset format interference mode data is carried out, and convolution processing data is obtained;Threshold extraction is carried out to convolution processing data, and based on the threshold after extraction, soft threshold processing is carried out, and the data after denoising is obtained;The classification processing of the data after denoising is carried out, and the classification result of preset format interference mode data is obtained;The interference mode identification model with threshold denoising module is used in the application, and the classification identification of preset format interference mode data set is carried out, to improve the interference mode identification accuracy under noise condition.

Description

Technical Field

[0001] This invention relates to an interference pattern recognition method based on threshold denoising, belonging to the field of wireless communication anti-interference technology. Background Technology

[0002] The open and shared nature of wireless communication makes it vulnerable to malicious interference attacks, severely impacting its security and reliability. Effective interference identification is crucial for implementing appropriate anti-interference measures. These measures aim to minimize or suppress interference, thereby reducing its adverse effects on communication quality, improving the anti-interference performance of communication systems, ensuring the stability, security, and reliability of wireless communication, and meeting ever-increasing communication demands.

[0003] Existing research largely focuses on identifying interference signals, aiming to mitigate their impact. While these methods perform well in certain situations, they cannot directly identify and classify different interference patterns. For technologies like intelligent anti-interference spectrum decision-making based on machine learning, interference pattern identification is more practically significant, as distinguishing different interference patterns is crucial for effectively managing and maintaining communication quality. However, despite the increasing demand in practical applications, in-depth research on interference pattern identification remains relatively limited. In recent years, with the development of machine learning and deep learning, interference pattern identification has achieved new breakthroughs, bringing more new possibilities to the development of intelligent anti-interference technologies.

[0004] Existing research has yielded significant results in the field of interference pattern recognition, but several issues remain to be addressed. First, most current interference pattern recognition methods are based on idealized high signal-to-noise ratio (SNR) conditions, neglecting the diversity and complexity of noise in real-world environments. Real-world environments are typically highly complex, containing various noise sources. These sources can introduce different frequencies and types of interference into signals. Therefore, considering noise in interference pattern recognition is essential for reliable operation in complex environments. A robust interference pattern recognition model should be able to handle diverse noise types, not just high SNR conditions. By considering noise, the model can better adapt to different environments, maintain performance in the presence of noise, and improve its generalization ability. Therefore, considering noise in the design and implementation of interference pattern recognition systems is crucial for better approximation of real-world application scenarios and for enhancing the model's robustness and generalization capabilities.

[0005] Secondly, current research focuses more on identifying interference signals than on identifying interference patterns, and in-depth research on interference pattern identification remains relatively limited. In real-world scenarios, interference signals may simultaneously originate from multiple sources, each with different characteristics and frequencies. Identifying interference patterns helps handle multi-source interference, improving system robustness. Furthermore, interference pattern identification goes beyond just frequency analysis; it considers multiple aspects, including the time and frequency domains, making the identification more comprehensive. Therefore, interference pattern identification is more adaptable, comprehensive, and robust, better suited to the diverse and complex interference conditions encountered in practical applications. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide an interference pattern recognition method based on threshold denoising. The method utilizes an interference pattern recognition model with a threshold denoising module to classify and recognize interference pattern datasets in a preset format. It reduces the impact of noise features on recognition in the feature space and proposes to perform threshold denoising in the shallow layer of the interference pattern recognition model to overcome the influence of noise on shallow features and the further influence of nonlinear propagation on deep features, thereby improving the accuracy of interference pattern recognition under noisy conditions.

[0007] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0008] This invention provides a threshold-based interference pattern recognition method, comprising:

[0009] Based on the baseband signal of the interference signal, obtain interference mode data in a preset format;

[0010] Using a trained interference pattern recognition model, convolution processing is performed on interference pattern data in a preset format to obtain convolutional processed data.

[0011] Threshold extraction is performed on the convolutional data, and soft thresholding is then applied based on the extracted thresholds to obtain the denoised data.

[0012] The denoised data is then classified to obtain the classification results of the interference pattern data in the preset format.

[0013] Optionally, the training method for the interference pattern recognition model includes:

[0014] Obtain a dataset of interference patterns in a preset format;

[0015] Construct an initial interference pattern recognition model;

[0016] Based on the preset format interference pattern dataset, the initial interference pattern recognition model is trained to obtain a trained interference pattern recognition model.

[0017] Optionally, the method for obtaining the preset format interference pattern dataset includes:

[0018] Obtain the baseband signal set of the interference signal;

[0019] By using a method to acquire interference pattern data in a preset format, the baseband signal set of the interference signal is transformed into an interference pattern dataset in a preset format.

[0020] Optionally, the method for obtaining interference pattern data in a preset format includes:

[0021] The baseband signal of the interference signal is modulated to obtain a modulated interference signal;

[0022] The modulated interference signal is filtered by a square root raised cosine filter and upsampled to obtain a sampled signal;

[0023] Based on the sampling signal, the gain between the interference transmitting device and the receiving device, and the noise at the receiving point, the signal received at the receiving point is obtained;

[0024] Perform a short-time Fourier transform on the signal received at the receiving point to obtain interference pattern data in a preset format.

[0025] The method of generating interference pattern data transforms the baseband signal set of interference signals into a interference pattern dataset in a preset format.

[0026] Optionally, the convolution processing of the interference pattern data in the preset format to obtain convolutional processed data includes:

[0027] The preset format interference pattern data is processed through the first convolutional layer to extract features, resulting in the first extracted data.

[0028] The first extracted data is normalized by passing it through a BN layer to obtain normalized data;

[0029] The normalized data is processed through a second convolutional layer to extract features, resulting in the second extracted data.

[0030] The second extracted data is pooled through a pooling layer to reduce the dimension of the feature map, resulting in convolutional processed data.

[0031] Optionally, the step of extracting a threshold from the convolutional data and performing soft thresholding based on the extracted threshold to obtain denoised data includes:

[0032] The first part of the result is obtained by taking the convolutional data as input and averaging it along different feature channels at a certain position in the feature map.

[0033] After performing a one-dimensional convolution on the convolutional data, the result is multiplied by the Softmax function, then another one-dimensional convolution is performed, and finally the result is passed through the Sigmoid function to obtain the second part of the result.

[0034] The average of each feature map in the convolutional data is then performed, followed by a one-dimensional convolution. Finally, the average of each feature map is multiplied by the Sigmoid function to obtain the third part of the result.

[0035] Multiply the first part of the result, the second part of the result, and the third part of the result together, and use this to extract the threshold;

[0036] The threshold and the convolutional data are used to perform a soft threshold calculation, which involves setting the data in the convolutional data whose absolute value is less than the threshold to zero, thus obtaining the denoised data.

[0037] Optionally, the method for extracting the threshold includes:

[0038] Based on the convolution processing of data, the mean C along the feature channel at location a in the feature map is obtained. a The formula is:

[0039]

[0040] Where: a is the feature map location, n is the number of feature channels, and C a x represents the mean value along the feature channel at location a in the feature map, and also represents the first part of the result. ja Let a be the feature value at position a of the j-th feature map;

[0041] Based on the convolutional processing data, a weighted average is calculated along different feature channels at feature map location a, and then compared with the aforementioned C. a Multiply to obtain the partial threshold Z at location a in the feature map. a The formula is:

[0042]

[0043] Where: Z a The threshold value at location a along the feature channel in the feature map is σ(·), which is a normalization function using the sigmoid function. The purpose of this function is to ensure that the obtained threshold value is between 0 and 1. v x j W represents the weighted average of the feature map at location a along different feature channels, and also represents the result of the second part. v W represents the relationship between different locations in the feature map. k For the weight matrix to be learned, x j Let λ be the eigenvalue of the j-th feature channel, and λ be the weighting coefficient.

[0044] The cross-channel relationship matrix W l The mean m of the j-th feature map at position a ja After normalizing the product, it is compared with the mean m at position a of the feature map. ja Multiplying them yields the third part of the result, which is then compared with the aforementioned partial threshold Z. a Multiply to obtain the threshold Z at location a in the feature map. a The formula is:

[0045] Z” a =Z a ·m ja ·σ(W l ·m ja )

[0046] Where: Z” a W is the threshold at location a in the feature map. l For the cross-channel relationship matrix, m ja Let m be the mean of the j-th feature map at position a on the feature map. ja ·σ(W l ·m ja () represents the third part of the result.

[0047] Optionally, the method for obtaining the weighting coefficient λ includes:

[0048] By embedding a Gaussian function to capture the relationship between the current channel and other channels, the weight coefficient λ of that channel is determined. i The formula is:

[0049]

[0050] Among them: W q Let λ be the weight matrix to be learned. i Let x be the i-th weighting coefficient, where i is the i-th feature channel out of n feature channels. i Let x be the eigenvalue of the i-th feature channel. m The eigenvalue of the m-th feature channel;

[0051] Based on the iterative training process and the continuous updating of the weight matrix, the global information obtained after training is ultimately independent of the query position, and the weighting coefficient λ is expressed as:

[0052]

[0053] Where: ρ(·) is the softmax function.

[0054] Optionally, the denoised data can be classified to obtain classification results for the interference pattern data in a preset format, including:

[0055] The denoised data is subjected to two convolution processes to obtain the pooled retained data;

[0056] The pooled data is then processed by a Dropout layer to suppress overfitting.

[0057] The data processed by the Dropout layer is made one-dimensional by the Flatten layer, and the one-dimensional data is then classified by the fully connected layer to obtain the classified output data.

[0058] The classification output data is converted into predicted probabilities using the Softmax activation function, and the classification results of the interference pattern data in a preset format are output based on the predicted probabilities.

[0059] Optionally, the interference pattern recognition model includes a first sub-module, a threshold denoising module, and a second sub-module;

[0060] The first submodule includes a first convolutional block, used to perform convolution processing on interference pattern data in a preset format;

[0061] The threshold denoising module is used to extract thresholds from the convolutional data and perform soft thresholding based on the extracted thresholds.

[0062] The second submodule includes a second convolutional block, a third convolutional block, a Dropout layer, a Flatten layer, and a fully connected layer, which are used to classify the denoised data to obtain the classification results of the interference pattern data in the preset format.

[0063] The first convolutional block, the second convolutional block, and the third convolutional block each include a first convolutional layer, a BN layer, a second convolutional layer, and a pooling layer;

[0064] The second and third convolutional blocks are used to perform two convolutional processes on the denoised data;

[0065] The Dropout layer is used to suppress overfitting;

[0066] The Flatten layer is used to make the data processed by the Dropout layer one-dimensional;

[0067] The fully connected layer is used to classify the one-dimensional data.

[0068] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0069] This invention addresses the problem of interference pattern recognition under noisy conditions by proposing a threshold-based interference pattern recognition method. The method utilizes an interference pattern recognition model with a threshold denoising module to classify and recognize interference pattern datasets in a preset format, effectively reducing the impact of noise on interference model recognition and demonstrating good performance in handling interference pattern recognition tasks.

[0070] This invention performs threshold denoising on the shallow layer of the interference pattern recognition model, overcoming the influence of noise on shallow features and the further influence of nonlinear propagation on deep features, thereby improving the accuracy of interference pattern recognition under noisy conditions. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0072] Figure 1 The flowchart shown is a process for the interference pattern recognition method based on threshold denoising according to the present invention.

[0073] Figure 2 The diagram shown is a structural schematic of the interference pattern recognition model of the present invention;

[0074] Figure 3 The diagram shown is a schematic of the threshold denoising module processing procedure of the present invention.

[0075] Figure 4 The diagram shown is a schematic diagram of the confusion matrix when the interference-to-noise ratio is 8dB in one embodiment of the present invention;

[0076] Figure 5 The figure shown is a graph illustrating the recognition accuracy at different interference-to-noise ratios after adding a threshold denoising module to the shallow layer of the model in one embodiment of the present invention. Detailed Implementation

[0077] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0078] The application principle of the present invention will be described in detail below with reference to the accompanying drawings.

[0079] like Figure 1 As shown, this embodiment of the invention provides an interference pattern recognition method based on threshold denoising, comprising:

[0080] S110: Based on the baseband signal of the interference signal, acquire interference mode data in a preset format;

[0081] S120: Using the trained interference pattern recognition model, perform convolution processing on the interference pattern data in the preset format to obtain convolutional processed data.

[0082] S130: Extract thresholds from the convolutional data and perform soft thresholding based on the extracted thresholds to obtain denoised data;

[0083] S140: Classify the denoised data to obtain the classification results of the interference pattern data in the preset format.

[0084] The threshold-based noise reduction interference pattern recognition method in this invention effectively reduces the impact of noise on interference pattern recognition and demonstrates good performance in handling interference pattern recognition tasks.

[0085] In one specific embodiment of the present invention, an interference pattern recognition model needs to be constructed before interference pattern recognition. The steps are as follows:

[0086] Obtain a dataset of interference patterns in a preset format;

[0087] Construct an initial interference pattern recognition model;

[0088] Based on the preset format interference pattern dataset, the initial interference pattern recognition model is trained to obtain the interference pattern recognition model.

[0089] Specifically, the step of obtaining the preset format interference pattern dataset includes:

[0090] Obtain the baseband signal x of the nth interference signal at time t. n (t); n represents the type of interference signal;

[0091] The baseband signal x of the nth interference signal n (t) After modulation by f(·), the modulation pattern s is obtained. n (t), i.e., s n (t)=f(x n (t))

[0092] Modulation mode s n (t) The sampled signal a is obtained after filtering by the square root raised cosine filter and upsampling g(·). n (t), i.e., a n (t)=g(f(x n (t)));

[0093] Obtain the gain d between the jamming transmitter and receiver n and the noise at the receiving point w n (t);

[0094] Based on gain d n and the noise at the receiving point w n (t) and the sampled signal a n (t) The received signal is represented as J n (t)=d n ·a n (t)+w n (t);

[0095] Perform a short-time Fourier transform on the signal at the receiving point to obtain interference mode data (IQ data) in a preset format, i.e., IQ = STFT(J n (t));

[0096] Based on the obtained interference pattern data in a preset format, the interference pattern dataset in a preset format, namely the IQ dataset, is obtained.

[0097] In this embodiment, the preset format interference pattern dataset includes 12 interference patterns, detailed as follows:

[0098] (1) Single sweep frequency I (SJ_I): Single tone interference with a sweep frequency speed of 0.2 GHz / us and a bandwidth of 10 MHz;

[0099] (2) Single sweep frequency II (SJ_II): Single tone interference with a sweep frequency speed of 0.1 GHz / us and a bandwidth of 10 MHz;

[0100] (3) Single sweep frequency Ⅲ (SJ_Ⅲ): Single tone interference with a sweep frequency speed of 0.2GHZ / us and a bandwidth of 10MHZ;

[0101] (4) Single sweep frequency IV (SJ_Ⅳ): Single tone interference with a sweep frequency speed of 0.1 GHz / us and a bandwidth of 10 MHz;

[0102] (5) Comb Interference I (CJ_I): Center frequencies are 640MHz and 690MHz, and bandwidth is 40MHz;

[0103] (6) Comb Interference II (CJ_II): The center frequencies are 890MHz and 835MHz, and the bandwidth is 25MHz;

[0104] (7) Comb Interference III (CJ_III): The center frequencies are 735MHz and 790MHz, and the bandwidth is 16MHz;

[0105] (8) Comb Interference IV (CJ_Ⅳ): Center frequencies are 650MHz and 720MHz, and bandwidth is 10MHz;

[0106] (9) Dual sweep frequency I (DSJ_I): Dual-tone interference with a sweep frequency rate of 0.1 GHz / µs and a bandwidth of 10 MHz;

[0107] (10) Dual sweep frequency II (DSJ_II): Dual-tone interference with a sweep frequency rate of 0.15 GHz / us and a bandwidth of 10 MHz;

[0108] (11) Dual sweep frequency III (DSJ_III): Dual-tone interference with a sweep frequency rate of 0.15 GHz / us and a bandwidth of 10 MHz;

[0109] (12) Dual sweep frequency IV (DSJ_Ⅳ): Dual-tone interference with a sweep frequency speed of 0.1 GHz / µs and a bandwidth of 10 MHz.

[0110] Specifically, based on the preset format interference pattern dataset, the initial interference pattern recognition model is trained to obtain the interference pattern recognition model. During the model training process, the loss function is used as the objective function to measure the difference between the model output and the actual label. By minimizing the loss function, the model parameters will be adjusted to improve the prediction accuracy of the input data.

[0111] As shown in the following equation, the optimization objective of this paper is to minimize the loss function.

[0112]

[0113] in Let y represent the predicted value, y represent the actual value, and k represent the number of categories. Represents the loss function. This indicates minimizing the loss function, where min means minimizing. The smaller the value, the smaller the difference between the model's predicted value and the true value, indicating that the model has a higher recognition ability.

[0114] In one specific embodiment of the present invention, the method for obtaining interference pattern data in a preset format includes:

[0115] The baseband signal of the interference signal is modulated to obtain a modulated interference signal;

[0116] The modulated interference signal is filtered by a square root raised cosine filter and upsampled to obtain a sampled signal;

[0117] Obtain the gain between the jamming transmitter and receiver, as well as the noise at the receiver point;

[0118] Based on the sampled signal, gain, and noise at the receiving point, the signal received by the receiving point is obtained;

[0119] Perform a short-time Fourier transform on the signal received at the receiving point to obtain interference pattern data (IQ data) in a preset format;

[0120] For specific steps, please refer to the above description of the steps to obtain the preset format interference pattern dataset.

[0121] In one specific embodiment of the present invention, the step of performing convolution processing on the interference pattern data in a preset format to obtain convolutional processed data includes:

[0122] The preset format interference pattern data is processed through the first convolutional layer to extract features, resulting in the first extracted data.

[0123] The first extracted data is normalized by passing it through a BN layer to obtain normalized data;

[0124] The normalized data is processed through a second convolutional layer to extract features, resulting in the second extracted data.

[0125] The second extracted data is pooled through a pooling layer to reduce the dimension of the feature map, resulting in convolutional processed data.

[0126] Specifically, the first and second convolutional layers extract features from the input data through convolution operations, which helps the model identify important information in the input data and achieve more accurate predictions and decisions. The Batch Normalization (BN) layer normalizes the input by calculating the mean and variance of each layer during training, and then performs translation and scaling operations to preserve the model's representational power, which helps accelerate the training process of deep neural networks and improve the model's generalization ability. The pooling layer reduces the dimensionality of the feature map by retaining the maximum value of the data points in each pooling window and discarding other values, thus preserving the most important features. This helps reduce the number of model parameters and computational overhead, simplifying the complexity of the network.

[0127] In one specific embodiment of the present invention, such as Figure 3 As shown, convolution processing is performed on the interference pattern data in a preset format to obtain convolutional processed data, including:

[0128] The first part of the result is obtained by taking the convolutional data as input and averaging it along different feature channels at a certain position in the feature map.

[0129] After performing a one-dimensional convolution on the convolutional data, the result is multiplied by the Softmax function, then another one-dimensional convolution is performed, and finally the result is passed through the Sigmoid function to obtain the second part of the result.

[0130] The average of each feature map in the convolutional data is then performed, followed by a one-dimensional convolution. Finally, the average of each feature map is multiplied by the Sigmoid function to obtain the third part of the result.

[0131] Multiply the first part of the result, the second part of the result, and the third part of the result together, and use this to extract the threshold;

[0132] The threshold and the convolutional data are used to perform a soft threshold calculation, which involves setting the data in the convolutional data whose absolute value is less than the threshold to zero, thereby obtaining the denoised data. Figure 2 middle, This indicates element-wise multiplication. Indicates soft threshold calculation;

[0133] In one specific embodiment of this invention, to achieve more accurate identification of interference patterns under noisy conditions, it is necessary to consider how to extract a suitable threshold to maintain the integrity of information while reducing noise interference. With the development of deep learning, denoising models are increasingly widely used. Currently, most deep learning-based denoising models use global average pooling to calculate a general threshold, and then apply this threshold to each location in the feature map. While this method is effective in some cases, it still has limitations. Because each location in the feature map contains different information, using the same threshold for soft thresholding at different locations may filter out features that are crucial for identification, leading to information loss and affecting the identification process. Therefore, this invention proposes a new threshold denoising method to better handle noise in the data. This method differs from traditional methods in that it considers the different information contained in different feature channels and different locations in the feature map, setting different thresholds for different locations in the feature map to reduce the loss of effective information, improve the model's noise resistance, and bring new possibilities for interference pattern recognition under noisy conditions.

[0134] Methods for extracting thresholds include:

[0135] Based on the convolution processing of data, the mean C along the feature channel at location a in the feature map is obtained. a The formula is:

[0136]

[0137] Where: a is the feature map location, n is the number of feature channels, and C a x represents the mean value along the feature channel at location a in the feature map, and also represents the first part of the result. ja Let a be the feature value at position a of the j-th feature map;

[0138] Based on the convolutional processing data, a weighted average is calculated along different feature channels at feature map location a, and then compared with the aforementioned C. a Multiply to obtain the partial threshold Z at location a in the feature map. a The formula is:

[0139]

[0140] Where: Z a The threshold value at location a along the feature channel in the feature map is σ(·), which is a normalization function using the sigmoid function. The purpose of this function is to ensure that the obtained threshold value is between 0 and 1. v x j W represents the weighted average of the feature map at location a along different feature channels, and also represents the result of the second part. v W represents the relationship between different locations in the feature map. k For the weight matrix to be learned, x j Let λ be the eigenvalue of the j-th feature channel, and λ be the weighting coefficient.

[0141] The method for obtaining λ is as follows:

[0142] By embedding a Gaussian function to capture the relationship between the current channel and other channels, the weight coefficient λ of that channel is determined. i The formula is:

[0143]

[0144] Among them: W q Let λ be the weight matrix to be learned. i Let x be the i-th weighting coefficient, where i is the i-th feature channel out of n feature channels. i Let x be the eigenvalue of the i-th feature channel. m The eigenvalue of the m-th feature channel;

[0145] Based on the iterative training process and the continuous updating of the weight matrix, the global information obtained after training is ultimately independent of the query position, and the weighting coefficient λ is expressed as:

[0146]

[0147] Where: ρ(·) is the softmax function.

[0148] The above implementation calculates thresholds at different locations in the feature map, resulting in a threshold map with the same dimensions as the input feature map. However, using the same threshold map for different feature channels cannot meet complex practical needs because different feature channels have different importance for classification. Therefore, to calculate thresholds more accurately, this patent assigns different weights to each channel based on their different importance. To achieve this, a one-dimensional convolution is introduced to capture the correlation between different channels. Through weight allocation, channels that contribute more to the classification task have higher weights.

[0149] The cross-channel relationship matrix W l The mean m of the j-th feature map at position a ja After normalizing the product, it is compared with the mean m at position a of the feature map. ja Multiplying them yields the third part of the result, which is then compared with the aforementioned partial threshold Z. a Multiply to obtain the threshold Z at location a in the feature map. a The formula is:

[0150] Z” a =Z a ·m ja ·σ(W l ·m ja )

[0151] Where: Z” a W is the threshold at location a in the feature map. l For the cross-channel relationship matrix, m ja Let m be the mean of the j-th feature map at position a on the feature map. ja ·σ(W l ·m ja () represents the third part of the result.

[0152] In one specific embodiment of this invention, as the network depth increases, noisy features in the data are still activated when entering deeper layers, which can adversely affect the recognition task. To address this issue, this invention introduces a threshold denoising module after the first convolutional block of the interference pattern recognition model. The purpose is to reduce the activation of noisy features in the feature space, thereby mitigating their impact on the classification task. By performing denoising at shallow layers of the network, the network can better learn useful features, improving the model's recognition performance.

[0153] In one specific embodiment of the present invention, the threshold and the convolutional data are subjected to soft thresholding calculation. This involves setting the data in the convolutional data whose absolute value is less than the threshold to zero, thus obtaining the denoised data. Soft thresholding is a crucial step in many signal denoising methods. Its purpose is to set values ​​in the original signal whose absolute value is less than a set threshold to zero, while keeping other values ​​unchanged. Soft thresholding is commonly used to reduce noise components in a signal. This is achieved by turning near-zero values ​​to zero. Simultaneously, this process causes those near-zero values, i.e., relatively small amplitude fluctuations, to shrink towards zero, effectively improving signal quality, reducing the impact of noise, and preserving important information in the signal. The soft thresholding function expression is shown below.

[0154]

[0155] Where x represents the input feature (data processed by convolution), y represents the output feature, and τ represents the threshold (the threshold Z extracted above). a ).

[0156] In one specific embodiment of the present invention, the denoised data is classified to obtain a classification result of the interference pattern data in a preset format, including:

[0157] The denoised data is subjected to two convolution processes to obtain the pooled retained data;

[0158] The pooled data is then processed by a Dropout layer to suppress overfitting.

[0159] The data processed by the Dropout layer is made one-dimensional by the Flatten layer, and the one-dimensional data is then classified by the fully connected layer to obtain the classified output data.

[0160] The classification output data is converted into predicted probabilities using the Softmax activation function, and the classification results of the interference pattern data in a preset format are output based on the predicted probabilities.

[0161] Specifically, the Dropout layer randomly sets the output of some neurons to zero during the training of the neural network. This helps prevent neurons from becoming overly reliant on specific information in the network, thereby reducing the risk of overfitting. The Flatten layer transforms the input multidimensional feature map into a one-dimensional vector, which is then fed into the fully connected layer for further processing. The fully connected layer usually appears in the last layer of the network, connecting all neurons in the previous layer with all neurons in the current layer, integrating local features into global features, and finally sending the calculated response value to the output.

[0162] In one specific embodiment of the present invention, such as Figure 2 As shown, the interference pattern recognition model includes a first submodule, a threshold denoising module, and a second submodule;

[0163] The first submodule includes a first convolutional block, which is used to perform convolution processing on interference pattern data in a preset format;

[0164] The threshold denoising module is used to extract thresholds from the convolutional data and perform soft thresholding based on the extracted thresholds.

[0165] The second submodule includes a second convolutional block, a third convolutional block, a Dropout layer, a Flatten layer, and a fully connected layer, which are used to classify the denoised data to obtain the classification results of the interference pattern data in the preset format.

[0166] The first convolutional block, the second convolutional block, and the third convolutional block all include a first convolutional layer, a BN layer, a second convolutional layer, and a pooling layer;

[0167] The second and third convolutional blocks are used to perform two convolutional processes on the denoised data;

[0168] Dropout layers are used to suppress overfitting;

[0169] The Flatten layer is used to flatten the data processed by the Dropout layer into one dimension.

[0170] A fully connected layer is used to classify one-dimensional data.

[0171] In one specific embodiment of the present invention, such as Figure 4 The confusion matrix shown is for an interference-to-noise ratio (INR) of 8dB. Rows represent true values, columns represent predicted values, and subscripts indicate the categories of the 12 interference patterns (from the aforementioned IQ dataset). At an INR of 8dB, the algorithm proposed in this invention can completely identify comb interference I, achieve a 99% recognition rate for comb interference III, a 97% recognition rate for dual-scan frequency II and III, a 95% recognition rate for single-scan frequency I, dual-scan frequency I, and dual-scan frequency IV, an 89% recognition rate for single-scan frequency III, an 86% recognition rate for comb interference IV, and a 75% recognition rate for single-scan frequency IV. The confusion matrix results demonstrate the effectiveness of this invention in interference pattern recognition.

[0172] In one specific embodiment of the present invention, such as Figure 5 The accuracy data curves shown are for different interference-to-noise ratios (SNR). The horizontal axis SNR represents the SNR, and the vertical axis represents the accuracy. When SNR = 6dB to SNR = 10dB, the shallow denoising proposed in this invention improves the classification accuracy by an average of 10% and 7% compared to no denoising and deep denoising, respectively. This invention demonstrates good performance in handling interference pattern recognition tasks.

[0173] In one specific embodiment of the present invention, to further evaluate the proposed method, the present invention introduces an F1-score to comprehensively understand the performance of the model in the recognition task. For the i-th category, the F1-score is:

[0174]

[0175] P i R represents accuracy, which indicates the proportion of samples correctly predicted as positive out of all samples predicted as positive. i Recall represents the proportion of samples correctly predicted as positive by the model out of all actual positive samples. TPi FP represents the number of correct positive examples in the model. i This indicates the number of instances where the model incorrectly predicted as positive. FN i This indicates the number of instances where the model incorrectly predicted as negative.

[0176]

[0177] Table 1 compares the F1-score (%) of different models when the noise-to-interference ratio is 8dB.

[0178] As shown in Table 1, at an interference-to-noise ratio of 8dB, the F1-score of the proposed shallow denoising method was compared with that of no denoising and deep denoising. For comb interference II, the F1-score of this invention is lower than the other two. However, for single-sweep I, single-sweep II, single-sweep III, single-sweep IV, comb interference I, comb interference III, comb interference IV, dual-sweep I, dual-sweep II, dual-sweep III, and dual-sweep IV, the F1-score of the proposed method is higher than the other two algorithms.

[0179] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for interference pattern recognition based on threshold denoising, characterized in that, include: Based on the baseband signal of the interference signal, obtain interference mode data in a preset format; Using a trained interference pattern recognition model, convolution processing is performed on interference pattern data in a preset format to obtain convolutional processed data. Threshold extraction is performed on the convolutional data, and soft thresholding is then applied based on the extracted thresholds to obtain the denoised data. The denoised data is then classified to obtain the classification results of the interference pattern data in the preset format. The convolution processing of the interference pattern data in the preset format to obtain convolutional processed data includes: The preset format interference pattern data is processed through the first convolutional layer to extract features, resulting in the first extracted data. The first extracted data is normalized by passing it through a BN layer to obtain normalized data; The normalized data is processed through a second convolutional layer to extract features, resulting in the second extracted data. The second extracted data is pooled through a pooling layer to reduce the dimension of the feature map, resulting in convolutional processed data. The process of extracting a threshold from the convolutional data and then performing soft thresholding based on the extracted threshold to obtain denoised data includes: The first part of the result is obtained by taking the convolutional data as input and averaging it along different feature channels at a certain position in the feature map. After performing a one-dimensional convolution on the convolutional data, the result is multiplied by the Softmax function, then another one-dimensional convolution is performed, and finally the result is passed through the Sigmoid function to obtain the second part of the result. The average of each feature map in the convolutional data is then performed, followed by a one-dimensional convolution. Finally, the average of each feature map is multiplied by the Sigmoid function to obtain the third part of the result. Multiply the first part of the result, the second part of the result, and the third part of the result together, and use this to extract the threshold; The threshold and the convolutional data are used to perform a soft threshold calculation, which involves setting the data in the convolutional data whose absolute value is less than the threshold to zero, thus obtaining the denoised data.

2. The interference pattern recognition method based on threshold denoising according to claim 1, characterized in that, The training method for the interference pattern recognition model includes: Obtain a dataset of interference patterns in a preset format; Construct an initial interference pattern recognition model; Based on the preset format interference pattern dataset, the initial interference pattern recognition model is trained to obtain a trained interference pattern recognition model.

3. The interference pattern recognition method based on threshold denoising according to claim 2, characterized in that, The method for obtaining the preset format interference pattern dataset includes: Obtain the baseband signal set of the interference signal; By using a method to acquire interference pattern data in a preset format, the baseband signal set of the interference signal is transformed into an interference pattern dataset in a preset format.

4. The interference pattern recognition method based on threshold denoising according to any one of claims 1 to 3, characterized in that, The method for obtaining interference pattern data in a preset format includes: The baseband signal of the interference signal is modulated to obtain a modulated interference signal; The modulated interference signal is filtered by a square root raised cosine filter and upsampled to obtain a sampled signal; Based on the sampling signal, the gain between the interference transmitting device and the receiving device, and the noise at the receiving point, the signal received at the receiving point is obtained; Perform a short-time Fourier transform on the signal received at the receiving point to obtain interference pattern data in a preset format; The method of generating interference pattern data transforms the baseband signal set of interference signals into a interference pattern dataset in a preset format.

5. The interference pattern recognition method based on threshold denoising according to claim 1, characterized in that, The method for extracting the threshold includes: The mean value along the feature channel at location a in the feature map is obtained by processing the data through convolution. The formula is: ； Where: 'a' represents the feature map position. The number of feature channels, This represents the mean value along the feature channel at location a in the feature map, and also represents the first part of the result. For the first The feature value at position a in the feature map; Based on the convolutional processing data, a weighted average is calculated along different feature channels at feature map location a, and then compared with the aforementioned... Multiply to obtain the feature map location partial threshold at the location The formula is: ； in: For feature map location At a certain threshold along the feature channel, For normalization, adopt The function aims to ensure that the obtained threshold is between 0 and 1. This represents the weighted average value along different feature channels at location a in the feature map, and also represents the result of the second part. This describes the relationship between different locations in the feature map. The weight matrix to be learned. For the first The feature values ​​of each feature channel These are weighting coefficients. for function; Cross-channel relationship matrix With the Each feature map at location mean at The product of these two values, after normalization, is compared with the mean value at position a in the feature map. Multiplying them yields the third part of the result, which is then combined with the aforementioned partial threshold. Multiply to obtain the feature map location Threshold at The formula is: ； in: The threshold at location a in the feature map. This is a cross-channel relationship matrix. For the first Each feature map at the feature map location The mean at that point, This refers to the third part of the result.

6. The interference pattern recognition method based on threshold denoising according to claim 5, characterized in that, The weighting coefficients The methods for obtaining it include: By embedding a Gaussian function to capture the relationship between the current channel and other channels, the weighting coefficient of that channel can be determined. The formula is: ； in: The weight matrix to be learned. For the first Each weighting coefficient for The first feature channel Feature channels, For the first The feature values ​​of each feature channel For the first The eigenvalues ​​of each feature channel; Based on the iterative training process and the continuous updating of the weight matrix, the final global information obtained through training is independent of the query location, and the weighting coefficients... Represented as: 。 7. The interference pattern recognition method based on threshold denoising according to claim 1, characterized in that, The denoised data is then classified to obtain classification results for the interference pattern data in a preset format, including: The denoised data is subjected to two convolution processes to obtain the pooled retained data; The pooled data is then processed by a Dropout layer to suppress overfitting. The data processed by the Dropout layer is made one-dimensional by the Flatten layer, and the one-dimensional data is then classified by the fully connected layer to obtain the classified output data. The classification output data is converted into predicted probabilities using the Softmax activation function, and the classification results of the interference pattern data in a preset format are output based on the predicted probabilities.

8. The interference pattern recognition method based on threshold denoising according to claim 1, characterized in that, The interference pattern recognition model includes a first sub-module, a threshold denoising module, and a second sub-module; The first submodule includes a first convolutional block, used to perform convolution processing on interference pattern data in a preset format; The threshold denoising module is used to extract thresholds from the convolutional data and perform soft thresholding based on the extracted thresholds. The second submodule includes a second convolutional block, a third convolutional block, a Dropout layer, a Flatten layer, and a fully connected layer, which are used to classify the denoised data to obtain the classification results of the interference pattern data in the preset format. The first convolutional block, the second convolutional block, and the third convolutional block each include a first convolutional layer, a BN layer, a second convolutional layer, and a pooling layer; The second and third convolutional blocks are used to perform two convolutional processes on the denoised data; The Dropout layer is used to suppress overfitting; The Flatten layer is used to make the data processed by the Dropout layer one-dimensional; The fully connected layer is used to classify the one-dimensional data.

Images