Improved snr grading spectrum sensing method based on improved densenet model

Abstract

The present application relates to the technical field of spectrum sensing, and particularly relates to a signal-to-noise ratio grading spectrum sensing method based on an improved DenseNet model, which solves the technical defects of large errors of existing traditional signal-to-noise ratio grading due to the factors of being difficult to grasp the distance and the dynamic movement of secondary users in the mobile communication process, specifically, when the secondary user performs local sensing, the sensing signal is first graded into high and low signal-to-noise ratios, the high signal-to-noise ratio signal can directly obtain a detection result from a CNN signal-to-noise ratio grading model, and the low signal-to-noise ratio signal is input into an improved DenseNet low signal-to-noise ratio model to perform deep feature mining to obtain a detection result. The CNN signal-to-noise ratio grading model can perform high-precision and low-latency signal-to-noise ratio grading on the sensing signal, the model structure is simple, the sensing speed is fast, the sensing latency of the high signal-to-noise ratio signal is greatly reduced, and the improved DenseNet low signal-to-noise ratio model still retains the advantage of strengthening feature transmission while reducing parameters.

Description

Technical Field

[0001] This invention relates to the field of spectrum sensing technology, and in particular to a signal-to-noise ratio hierarchical spectrum sensing method based on an improved DenseNet model. Background Technology

[0002] With the rapid development of communication technology, research on sixth-generation mobile communication technology (6G) has gradually attracted widespread attention in the industry. Research indicates that 6G will achieve leapfrog communication in terms of communication speed, user capacity, and propagation latency, building upon 5G. These improvements are inseparable from the huge demand for spectrum resources.

[0003] Spectrum resources are still allocated using the traditional "fixed allocation" method. This means that a segment of spectrum is allocated to a specific licensed user (primary user) based on demand. This spectrum segment is only permitted for data communication transmission by the licensed user; unlicensed users (secondary users) cannot access this spectrum for data communication. However, in real-world scenarios, it often happens that only a very small portion of the allocated spectrum is constantly used for information transmission, leaving other communication devices with no spectrum available, resulting in wasted spectrum resources.

[0004] Spectrum sensing can detect whether licensed frequency bands are idle. If they are idle, transmission can be carried out on that channel using spectrum access technology. Therefore, accurate detection of licensed frequency bands is crucial.

[0005] As deep learning becomes increasingly mature, and with the development of various network models, the detection accuracy of spectrum sensing technology is becoming higher and higher.

[0006] Traditional signal-to-noise ratio (SNR) classification is based on calculations of the SNR of the sensed signal according to factors such as distance, primary user transmit power, and path loss. However, the actual environment is complex and variable, and secondary users move dynamically during mobile communication, making it difficult to accurately grasp factors such as distance, resulting in significant errors. Summary of the Invention

[0007] To overcome the shortcomings of traditional signal-to-noise ratio (SNR) classification methods, which suffer from significant errors due to factors such as the dynamic movement of secondary users and the difficulty in accurately determining distances during mobile communication, this invention provides an SNR classification spectrum sensing method based on an improved DenseNet model. Specifically, when a secondary user performs local sensing, the sensing signal is first classified into high and low SNR categories. High SNR signals can be directly detected by a CNN SNR classification model, while low SNR signals are input into an improved DenseNet low SNR model for deep feature mining to obtain the detection result.

[0008] This invention provides a signal-to-noise ratio hierarchical spectrum sensing method based on an improved DenseNet model, the steps of which are as follows:

[0009] Step 1: Collect modulation signal datasets with different modulation methods and signal-to-noise ratios from GNU Radio for experiments. The type of modulation signal dataset is binary IQ signal.

[0010] Step 2: For each modulated signal dataset, preprocess it using the min-max normalization method to map the binary IQ signal to the [0,1] interval for feature extraction.

[0011] Step 3: Divide the preprocessed IQ signals into training set, validation set and test set according to a certain ratio;

[0012] Step 4: Construct the DenseNet multi-scale signal-to-noise ratio (SNR) classification model, which consists of a CNN SNR classification model and an improved DenseNet low SNR model. The CNN SNR classification model includes two sequentially connected one-dimensional convolutional neural network layers, two fully connected (fc) layers, and a softmax classifier. The preprocessed IQ signal is input into the CNN SNR classification model, and the classification reference points are obtained. If it is a high SNR signal, a high SNR signal detection result is output; if it is a low SNR signal, a low SNR index is output. The improved DenseNet model adds three final output modules to the original DenseNet model. These three final output modules are connected to the three output modules of the original DenseNet model. The output of one output module interacts with the outputs of the other two output modules within the corresponding final output module. The outputs of the three final output modules are processed within the original DenseNet model's Pooling and Normalizing and... The Concat module fuses the outputs; the low signal-to-noise ratio index obtained from the CNN signal-to-noise ratio classification model is multiplied by the preprocessed IQ signal, and then input into the improved DenseNet low signal-to-noise ratio model for detection, finally outputting the low signal-to-noise ratio signal detection result;

[0013] Step 5: First, input the training set into the CNN signal-to-noise ratio (SNR) classification model. By comparing the classification accuracy of different modulation methods under the same SNR, select the most suitable threshold as the SNR classification point. Then, input the validation set into the CNN SNR classification model for validation. Adjust the hyperparameters to obtain a trained and validated CNN SNR classification model.

[0014] Step 6: Then input the training set into the improved DenseNet low signal-to-noise ratio model to perform low signal-to-noise ratio awareness; then input the validation set into the improved DenseNet low signal-to-noise ratio model for validation, adjust the hyperparameters, and obtain the trained and validated improved DenseNet low signal-to-noise ratio model.

[0015] Step 7: Combine the final CNN signal-to-noise ratio (SNR) classification model from Steps 5 and 6 with the final improved DenseNet low SNR model to obtain a trained and validated DenseNet multi-scale SNR classification model. Input the test set from Step 3 into the obtained DenseNet multi-scale SNR classification model to test the model's accuracy on the test set. When the accuracy meets the requirements, the optimal DenseNet multi-scale SNR classification model is obtained.

[0016] Step 8: Process the licensed frequency band data to be sensed using the optimal DenseNet multi-scale signal-to-noise ratio (SNR) classification model obtained in Step 7. After data preprocessing, the licensed frequency band data to be sensed is first input into the CNN SNR classification model to analyze and obtain classification reference points. If it is a high SNR signal, output whether the licensed frequency band is in an idle state; if it is a low SNR signal, output a low SNR index. The low SNR index obtained by the CNN SNR classification model is multiplied by the preprocessed licensed frequency band data to be sensed, and then input into the improved DenseNet low SNR model for detection. Finally, output whether the licensed frequency band is in an idle state.

[0017] The CNN signal-to-noise ratio (SNR) ranking model works by using a simple one-dimensional convolutional neural network to extract features from the signal, then weighting the features from the previous layers through a fully connected (FC) layer, and finally converting the features into detection results through a softmax classifier. This model has a simple structure, fast perception speed, and significantly reduces the perception latency for high SNR signals. The output of one output module interacts with the outputs of the other two output modules in their corresponding final output modules. Outputs from output modules that do not interact directly pass through their respective final output modules. Finally, the three final output modules fuse the output. This approach reduces the number of parameters while still retaining the advantage of enhanced feature transfer.

[0018] Preferably, in step three, the preprocessed IQ signals are divided into training set, validation set and test set in a ratio of 8:1:1.

[0019] Excellently, in step six, when performing low signal-to-noise ratio sensing, the Adam optimizer is used with an initial learning rate set to 0.001.

[0020] The technical solution provided by this invention has the following technical effects compared with the prior art:

[0021] This invention designs a lightweight CNN signal-to-noise ratio (SNR) classification model, capable of performing high-precision, low-latency SNR classification of perceived signals. The model has a simple structure and fast sensing speed, significantly reducing the sensing latency of high SNR signals. Furthermore, this invention improves upon the existing DenseNet model by adding three final output modules. The output of one of these modules interacts with the outputs of the other two, allowing for the extraction of more features. Simultaneously, it introduces noise into the outputs of modules without interaction, making the results more robust. The improved DenseNet low SNR model abandons the original pairwise interaction of the three outputs, performing only two interactions instead, reducing parameters while retaining the advantage of enhanced feature transfer. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the structure of the DensetNet multi-scale signal-to-noise ratio hierarchical model according to a certain embodiment of the present invention;

[0025] Figure 2 This is a flowchart of a signal-to-noise ratio hierarchical spectrum sensing method based on an improved DenseNet model in a certain embodiment of the present invention. Detailed Implementation

[0026] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0028] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] In one embodiment, such as Figure 1As shown, the signal-to-noise ratio hierarchical spectrum sensing method based on the improved DenseNet model comprises the following steps:

[0030] Step 1: Collect modulation signal datasets with different modulation methods and signal-to-noise ratios from GNU Radio for experiments. The type of modulation signal dataset is binary IQ signal.

[0031] Step 2: For each modulated signal dataset, preprocess it using the min-max normalization method to map the binary IQ signal to the [0,1] interval for feature extraction.

[0032] Step 3: Divide the preprocessed IQ signals into training set, validation set and test set according to a certain ratio; in a specific embodiment, the ratio is 8:1:1.

[0033] Step 4: Construct the DenseNet multi-scale signal-to-noise ratio (SNR) classification model, which consists of a CNN SNR classification model and an improved DenseNet low SNR model. The CNN SNR classification model includes two sequentially connected one-dimensional convolutional neural network layers, two fully connected (fc) layers, and a softmax classifier. The preprocessed IQ signal is input into the CNN SNR classification model, and the classification reference points are obtained. If it is a high SNR signal, a high SNR signal detection result is output; if it is a low SNR signal, a low SNR index is output. The improved DenseNet model adds three final output modules to the original DenseNet model. These three final output modules are connected to the three output modules of the original DenseNet model. The output of one output module interacts with the outputs of the other two output modules within the corresponding final output module. The outputs of the three final output modules are processed within the original DenseNet model's Pooling and Normalizing and... The Concat module fuses the outputs; the low signal-to-noise ratio index obtained from the CNN signal-to-noise ratio classification model is multiplied by the preprocessed IQ signal, and then input into the improved DenseNet low signal-to-noise ratio model for detection, finally outputting the low signal-to-noise ratio signal detection result;

[0034] Step 5: First, input the training set into the CNN signal-to-noise ratio (SNR) classification model. By comparing the classification accuracy of different modulation methods under the same SNR, select the most suitable threshold as the SNR classification point. Then, input the validation set into the CNN SNR classification model for validation. Adjust the hyperparameters to obtain a trained and validated CNN SNR classification model.

[0035] Step Six: Then, input the training set into the improved DenseNet low signal-to-noise ratio model, specifically using the Adam optimizer with an initial learning rate of 0.001 for low signal-to-noise ratio awareness; then input the validation set into the improved DenseNet low signal-to-noise ratio model for validation, adjusting the hyperparameters to obtain a trained and validated improved DenseNet low signal-to-noise ratio model; Step Seven: Combine the CNN signal-to-noise ratio classification model from Steps Five and Six with the final improved DenseNet low signal-to-noise ratio model to obtain a trained and validated DenseNet multi-scale signal-to-noise ratio classification model. Input the test set from Step Three into the obtained DenseNet multi-scale signal-to-noise ratio classification model to test the model's accuracy on the test set. When the accuracy meets the requirements, the optimal DenseNet multi-scale signal-to-noise ratio classification model is obtained.

[0036] Step 8: Process the licensed frequency band data to be sensed using the optimal DenseNet multi-scale signal-to-noise ratio (SNR) classification model obtained in Step 7. After data preprocessing, the licensed frequency band data to be sensed is first input into the CNN SNR classification model to analyze and obtain classification reference points. If it is a high SNR signal, output whether the licensed frequency band is in an idle state; if it is a low SNR signal, output a low SNR index. The low SNR index obtained by the CNN SNR classification model is multiplied by the preprocessed licensed frequency band data to be sensed, and then input into the improved DenseNet low SNR model for detection. Finally, output whether the licensed frequency band is in an idle state.

[0037] The CNN signal-to-noise ratio (SNR) ranking model works by using a simple one-dimensional convolutional neural network to extract features from the signal, then weighting the features from the previous layers through a fully connected (FC) layer, and finally converting the features into detection results through a softmax classifier. This model has a simple structure, fast perception speed, and significantly reduces the perception latency for high SNR signals. The output of one output module interacts with the outputs of the other two output modules in their corresponding final output modules. Outputs from output modules that do not interact directly pass through their respective final output modules. Finally, the three final output modules fuse the output. This approach reduces the number of parameters while still retaining the advantage of enhanced feature transfer.

[0038] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.

Claims

1. A signal-to-noise ratio hierarchical spectrum sensing method based on an improved DenseNet model, characterized in that, The steps are as follows: Step 1: Collect modulation signal datasets with different modulation methods and signal-to-noise ratios from GNU Radio for experiments. The type of modulation signal dataset is binary IQ signal. Step 2: For each modulated signal dataset, preprocess it using the min-max normalization method to map the binary IQ signal to the [0,1] interval for feature extraction. Step 3: Divide the preprocessed IQ signals into training set, validation set and test set according to a certain ratio; Step 4: Construct the DenseNet multi-scale signal-to-noise ratio (SNR) classification model, which consists of a CNN SNR classification model and an improved DenseNet model. The CNN SNR classification model includes two sequentially connected one-dimensional convolutional neural network layers, two fully connected (fc) layers, and a Softmax classifier. The preprocessed IQ signal is input into the CNN SNR classification model, which analyzes and obtains classification reference points. If it is a high SNR signal, a high SNR signal detection result is output; if it is a low SNR signal, a low SNR index is output. The improved DenseNet model adds three final output modules to the original DenseNet model. These three final output modules are connected to the three output modules of the original DenseNet model. The output of one output module interacts with the outputs of the other two output modules in the corresponding final output module. The outputs of the three final output modules are fused and output in the Pooling, Normalize, and Concat modules of the original DenseNet model. The low SNR index obtained from the CNN SNR classification model is multiplied by the preprocessed IQ signal and then input into the improved DenseNet model for detection, finally outputting the low SNR signal detection result. Step 5: First, input the training set into the CNN signal-to-noise ratio (SNR) classification model. By comparing the classification accuracy of different modulation methods under the same SNR, select the most suitable threshold as the SNR classification point. Then, input the validation set into the CNN SNR classification model for validation. Adjust the hyperparameters to obtain a trained and validated CNN SNR classification model. Step 6: Then input the training set into the improved DenseNet model for low signal-to-noise ratio sensing; The validation set is then input into the improved DenseNet model for validation, and the hyperparameters are adjusted to obtain a trained and validated improved DenseNet model. Step 7: Combine the final CNN signal-to-noise ratio (SNR) classification model from Steps 5 and 6 with the final improved DenseNet model to obtain a trained and validated DenseNet multi-scale SNR classification model. Input the test set from Step 3 into the obtained DenseNet multi-scale SNR classification model to test the model's accuracy in judging the test set. When the accuracy meets the requirements, the optimal DenseNet multi-scale SNR classification model is obtained. Step 8: Process the licensed frequency band data to be sensed using the optimal DenseNet multi-scale signal-to-noise ratio (SNR) classification model obtained in Step 7. After data preprocessing, the licensed frequency band data to be sensed is first input into the CNN SNR classification model to analyze and obtain classification reference points. If it is a high SNR signal, the model outputs whether the licensed frequency band is in an idle state; if it is a low SNR signal, the model outputs a low SNR index. The low SNR index obtained from the CNN SNR classification model is multiplied by the preprocessed licensed frequency band data to be sensed, and then input into the improved DenseNet model for detection. Finally, the model outputs whether the licensed frequency band is in an idle state.

2. The signal-to-noise ratio hierarchical spectrum sensing method based on the improved DenseNet model according to claim 1, characterized in that, In step three, the preprocessed IQ signals are divided into training set, validation set and test set in a ratio of 8:1:

1.

3. The signal-to-noise ratio hierarchical spectrum sensing method based on the improved DenseNet model according to claim 2, characterized in that, In step six, when performing low signal-to-noise ratio sensing, the Adam optimizer is used with an initial learning rate of 0.001.

Images