A Broadband Signal Detection Method Based on Unsupervised Deep Neural Networks

Abstract

This invention discloses a broadband signal detection method based on unsupervised deep neural networks. First, an autoencoder is trained on a power spectrum dataset of broadband signals as a feature extractor. Then, the feature extractor processes the broadband power spectrum dataset, and initial pseudo-labels are obtained through an unsupervised deep neural network and post-processing methods. Next, a supervised neural network is built on the feature extractor and trained using the initial pseudo-labels. Finally, the pseudo-labels are fine-tuned using the training output. After several iterations, the overall training is completed, achieving efficient and accurate broadband signal detection. This method employs unsupervised learning, effectively detecting broadband spectra without manual labeling, significantly reducing the degree of human intervention and complexity. It can also be used to improve the performance and robustness of other supervised learning algorithms, offering advantages such as high detection accuracy, good recognition effect, wide applicability, and no need for manual labeling.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a broadband signal detection method based on unsupervised deep neural networks. Background Technology

[0002] In broadband signal detection methods based on deep neural networks, obtaining accurate manual labels is crucial for the accuracy and robustness of the detection results. Existing supervised learning typically relies on large amounts of accurately labeled data to guide model parameter updates, where the accuracy of the labels directly impacts network performance. However, manually labeling accurate data is a massive undertaking and quite challenging. Furthermore, due to the lack of publicly available large datasets, the performance of many models on simulated datasets is not necessarily equivalent to their performance on real datasets. Even with a small amount of real data for fine-tuning, accurately labeling this data still requires significant time and effort.

[0003] Among existing methods, CN 112784690 A discloses a broadband signal parameter estimation method based on deep learning. This method generates a grayscale time-frequency map by segmenting and normalizing time-series data, and then uses a YOLOv4 network for signal detection. While effective, this method relies on manually labeled training data, which is not only time-consuming but may also limit the size and diversity of the dataset, thus affecting the overall performance and accuracy of the model. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a broadband signal detection method based on unsupervised deep neural networks, which solves the problem of requiring a large amount of accurately labeled data in broadband signal detection.

[0005] The technical solution adopted in this invention is: a broadband signal detection method based on unsupervised deep neural networks, the specific steps of which are as follows:

[0006] S1. Train an autoencoder using the power spectrum data of a broadband signal as a feature extractor;

[0007] S2. Use a feature extractor to obtain the features of the broadband power spectrum. Build an unsupervised deep neural network based on the feature extractor, train it using the features of the broadband power spectrum, obtain the optimized features, and then perform post-processing to generate initial pseudo-labels.

[0008] S3. Based on the initial pseudo-labels, train a supervised neural network to obtain preliminary broadband signal detection results;

[0009] S4. Use the detection results to iteratively fine-tune the pseudo-labels to obtain new pseudo-labels to continue training the supervised neural network. Iterate until the maximum number of optimizations is reached, and finally achieve efficient and accurate broadband signal detection.

[0010] Furthermore, step S1 is specifically as follows:

[0011] First, a broadband IQ signal containing multiple modulation methods is acquired and preprocessed to obtain broadband signal power spectrum data.

[0012] The modulation schemes include BPSK, QPSK, OQPSK, PI4QPSK, 8QAM, 16QAM, and GMSK; the preprocessing steps include resampling, amplitude adjustment, power spectrum calculation, and normalization; the Pwelch method is used when calculating the power spectrum.

[0013] Then, an autoencoder model is constructed, using the loss function L. MSE Optimize model parameters; the autoencoder model structure includes: an input layer, an encoding layer, and a decoding layer; specifically as follows:

[0014] Input layer: Designed to receive inputs of shape (1, L), where L represents the length of the input power spectrum;

[0015] Encoding layer: A residual network structure is adopted, with a total of 5 residual modules. The output length of each residual block is the input length / (2^number of layers).

[0016] Decoding layer: A residual network structure is adopted, which replaces the convolutional layer (Conv1d) in the encoding layer with the transposed convolutional layer (ConvTranspose1d), and performs upsampling at the corresponding downsampling positions to obtain an output with the same shape as the input data.

[0017] Finally, based on the preprocessed broadband signal power spectrum dataset, the autoencoder is trained until convergence and the error value is optimal. The encoding layer of the model at this point is then saved as a feature extractor.

[0018] Furthermore, step S2 is specifically as follows:

[0019] S21. Based on step S1, the obtained preprocessed broadband signal power spectrum dataset is processed by a feature extractor to obtain the corresponding feature vector.

[0020] S22. Based on the feature extractor, build an unsupervised deep neural network, optimize and adjust the feature vector, and obtain the output features;

[0021] The unsupervised deep neural network is specifically described below:

[0022] Input the preprocessed broadband signal power spectrum data, and based on the feature extractor obtained in step S1, obtain the outputs of its 3rd, 4th, and 5th layer residual modules.

[0023] Where i represents the output of the i-th residual module, n represents the number of channels of the output feature, and L i This represents the output length of the i-th residual module. Represents the real number field.

[0024] Then, use SEBlock on f i Perform feature optimization:

[0025] 1) In f i Global average pooling is performed on the last dimension to obtain the output.

[0026] Among them, f i ′ represents the output of global average pooling.

[0027] 2) f i Features are compressed and expanded through two fully connected layers;

[0028] 3) Pass the output of the fully connected layer through the sigmoid function to obtain an output located in the interval (0,1);

[0029] 4) Compare the output with f i Multiplying by the channel dimension yields the optimized features.

[0030] The output is then upsampled to the length of the input power spectrum to obtain the output. And calculate F i Rather than the mean in the length dimension difference

[0031] Recalculate The cosine similarity sim with the input broadband power spectrum is calculated.

[0032] Here, sgn(·) represents the sign function, which outputs 1 when its internal expression is greater than or equal to 0; otherwise, it outputs 0; and thus obtains the characteristic.

[0033]

[0034] Among them, F i n,l Representing feature F i The l-th element of all channels.

[0035] Finally, the features θ obtained from different residual modules are... i The features are concatenated along the channel dimension and passed through an SEBlock to obtain the output features of the unsupervised network. And adopts adaptive decision boundary loss L adb The network optimization expression is as follows:

[0036] D=|σ(θ)-σ(0)|

[0037]

[0038] Where σ represents the sigmoid function; N represents the batch size of the network input; ||·||1 and ||·||2 represent the L1 distance and L2 distance respectively, and α represents the hyperparameter, which represents the weight of the L1 distance.

[0039] S23. Based on the broadband signal power spectrum dataset preprocessed in step S1, freeze the autoencoder parameters, train the unsupervised neural network until the convergence and error values ​​are optimal, and save the model at this time as a pseudo-label generator.

[0040] S24. Post-process the output of the pseudo-tag generator obtained in step S23 to obtain the initial pseudo-tags;

[0041] The post-processing is specifically as follows:

[0042] The features θ obtained from the pseudo-label generator are input into a fully connected conditional random field (DenseCRF) to obtain the output.

[0043] Where X represents the output feature of a fully connected conditional random field.

[0044] Then, X is filtered using a median filter to obtain the initial pseudo-labels.

[0045] Furthermore, step S3 is specifically as follows:

[0046] A supervised deep neural network is built on the feature extractor. The input data is the broadband signal power spectrum dataset after preprocessing in step S1, and the label is Y obtained in step S2. The supervised neural network is trained using binary cross-entropy until the convergence and error values ​​are optimal to obtain the preliminary broadband signal detection result G.

[0047] The supervised deep neural network is specifically as follows:

[0048] Input the preprocessed broadband signal power spectrum data, and obtain the output of each residual module based on the feature extractor obtained in step S1.

[0049] Where i represents the output of the i-th residual module, n represents the number of channels of the output feature, and L i This represents the output length of the i-th residual module.

[0050] Then, f i Upsampled to the length of the input power spectrum, the feature is obtained.

[0051] Where, p i Indicates f i Features obtained after upsampling.

[0052] Then, subtract p5 from p1, p2, p3, and p4 respectively to obtain p′. i ; will p′ i Through a convolutional layer and a sigmoid function, and the corresponding p i Multiplying them together, we get P. i .

[0053] Where, p′ i Indicates the corresponding p i The difference between p5 and p5; P i p i With p i The product of ′.

[0054] Finally, P i By concatenating the data along the channel dimension and calculating the mean, we obtain...

[0055] Here, G represents the output of the supervised network, i.e., the preliminary broadband signal detection result.

[0056] Furthermore, step S4 is specifically as follows:

[0057] During the training of the supervised neural network based on step S3, starting from the third optimization of the network parameters, before each optimization, the previous pseudo-labels are fine-tuned using the new output, as shown in the following expression:

[0058] Y j+1 =λG j +(1-λ)Y j

[0059] Where j represents the j-th update of network parameters, Y j+1 G represents the pseudo-label used in the (j+1)th update of network parameters. j Y represents the output of the j-th network. j Y represents the pseudo-label used in the j-th update of network parameters. 1 =G, where λ represents the hyperparameter, i.e., the degree of each update.

[0060] Finally, based on the actual situation, set the maximum number of optimization attempts, optimize the network parameters until the specified number of optimization attempts are completed, and complete the broadband signal detection based on unsupervised deep neural networks.

[0061] The beneficial effects of this invention are as follows: The method of this invention first trains an autoencoder based on a power spectrum dataset of broadband signals, using it as a feature extractor. Then, the feature extractor is applied to process the broadband power spectrum dataset, and initial pseudo-labels are obtained through an unsupervised deep neural network and post-processing methods. Next, a supervised neural network is constructed based on the feature extractor, and trained using the initial pseudo-labels. Finally, the pseudo-labels are fine-tuned using the training output. After several iterations, the overall training is completed, achieving efficient and accurate broadband signal detection. This method employs unsupervised learning, effectively detecting broadband spectra without the need for manual labeling, significantly reducing the degree of human intervention and complexity. It can also be used to improve the performance and robustness of other supervised learning algorithms, offering advantages such as high detection accuracy, good recognition effect, wide applicability, and no need for manual labeling. Attached Figure Description

[0062] Figure 1 This is a flowchart of a broadband signal detection method based on an unsupervised deep neural network according to the present invention.

[0063] Figure 2 This is a diagram of the unsupervised deep neural network architecture described in an embodiment of the present invention.

[0064] Figure 3 This is a visualization simulation result of broadband signal power spectrum detection in an embodiment of the present invention. Detailed Implementation

[0065] The invention will now be further described with reference to the accompanying drawings and embodiments.

[0066] like Figure 1 The flowchart of a broadband signal detection method based on an unsupervised deep neural network according to the present invention is shown below. The specific steps are as follows:

[0067] S1. Train an autoencoder using the power spectrum data of a broadband signal as a feature extractor;

[0068] S2. Use a feature extractor to obtain the features of the broadband power spectrum. Build an unsupervised deep neural network based on the feature extractor, train it using the features of the broadband power spectrum, obtain the optimized features, and then perform post-processing to generate initial pseudo-labels.

[0069] S3. Based on the initial pseudo-labels, train a supervised neural network to obtain preliminary broadband signal detection results;

[0070] S4. Use the detection results to iteratively fine-tune the pseudo-labels to obtain new pseudo-labels to continue training the supervised neural network. Iterate until the maximum number of optimizations is reached, and finally achieve efficient and accurate broadband signal detection.

[0071] In this embodiment, step S1 is specifically as follows:

[0072] First, a broadband IQ signal containing multiple modulation methods is acquired and preprocessed to obtain broadband signal power spectrum data.

[0073] The modulation schemes include BPSK, QPSK, OQPSK, PI4QPSK, 8QAM, 16QAM, and GMSK, aiming to enhance the versatility of the recognition algorithm. The preprocessing steps include resampling, amplitude adjustment, power spectrum calculation, and normalization, designed to optimize signal data for subsequent feature extraction and analysis, thereby improving the performance of the recognition algorithm. When calculating the power spectrum, the Pwelch method is used; in this embodiment, the Fourier transform length is set to 16384, a Hamming window function is employed, and the overlap length is set to 8192.

[0074] Then, an autoencoder model is constructed, using the loss function L. MsE Used to measure the difference between model predictions and the original input, in order to optimize model parameters; the autoencoder model structure includes: an input layer, an encoding layer, and a decoding layer; specifically as follows:

[0075] Input layer: Designed to receive inputs of shape (1, L), where L represents the length of the input power spectrum;

[0076] Encoding layer: A residual network structure is adopted, with a total of 5 residual modules. The output length of each residual block is the input length / (2^layer number), thereby gradually reducing the feature dimension and deepening feature extraction.

[0077] Decoding layer: A residual network structure is adopted, which replaces the convolutional layer (Conv1d) in the encoding layer with the transposed convolutional layer (ConvTranspose1d), and performs upsampling at the corresponding downsampling positions to obtain an output with the same shape as the input data.

[0078] Finally, based on the preprocessed broadband signal power spectrum dataset, the autoencoder is trained until convergence and the error value is optimal. The encoding layer of the model at this point is then saved as a feature extractor.

[0079] In this embodiment, step S2 is specifically as follows:

[0080] S21. Based on step S1, the obtained preprocessed broadband signal power spectrum dataset is processed by a feature extractor to obtain the corresponding feature vector.

[0081] S22. Based on the feature extractor, build an unsupervised deep neural network, optimize and adjust the feature vector, and obtain the output features;

[0082] The unsupervised deep neural network architecture diagram is as follows: Figure 2 As shown, the unsupervised deep neural network is specifically as follows:

[0083] Input the preprocessed broadband signal power spectrum data, and based on the feature extractor obtained in step S1, obtain the outputs of its 3rd, 4th, and 5th layer residual modules.

[0084] Where i represents the output of the i-th residual module, n represents the number of channels of the output feature, and L i This represents the output length of the i-th residual module. Represents the real number field.

[0085] Then, use SEBlock on f i Perform feature optimization:

[0086] 1) In f i Global average pooling is performed on the last dimension to obtain the output.

[0087] Among them, f i ′ represents the output of global average pooling.

[0088] 2) f i Features are compressed and expanded through two fully connected layers;

[0089] 3) Pass the output of the fully connected layer through the sigmoid function to obtain an output located in the interval (0,1);

[0090] 4) Compare the output with f i Multiplying by the channel dimension yields the optimized features.

[0091] The output is then upsampled to the length of the input power spectrum to obtain the output. And calculate F i Rather than the mean in the length dimension difference

[0092] To prevent the network from identifying noise as a salient feature, further calculations are needed. The cosine similarity sim with the input broadband power spectrum is calculated.

[0093] Here, sgn(·) represents the sign function, which outputs 1 when its internal expression is greater than or equal to 0; otherwise, it outputs 0; and thus obtains the characteristic.

[0094]

[0095] Among them, F i n,l Representing feature F i The l-th element of all channels.

[0096] Finally, the features θ obtained from different residual modules are... i The features are concatenated along the channel dimension and passed through an SEBlock to obtain the output features of the unsupervised network. And adopts adaptive decision boundary loss L adb The network optimization expression is as follows:

[0097] D=|σ(θ)-σ(0)|

[0098]

[0099] Where σ represents the sigmoid function; N represents the batch size of the network input; ||·||1 and ||·||2 represent the L1 distance and L2 distance respectively, and α represents the hyperparameter, which represents the weight of the L1 distance.

[0100] S23. Based on the broadband signal power spectrum dataset preprocessed in step S1, freeze the autoencoder parameters, train the unsupervised neural network until the convergence and error values ​​are optimal, and save the model at this time as a pseudo-label generator.

[0101] S24. Post-process the output of the pseudo-tag generator obtained in step S23 to obtain the initial pseudo-tags;

[0102] The post-processing is specifically as follows:

[0103] The features θ obtained from the pseudo-label generator are input into a fully connected conditional random field (DenseCRF) to obtain the output.

[0104] Where X represents the output feature of a fully connected conditional random field.

[0105] To prevent interference spikes, a median filter is used to filter X to obtain the initial pseudo-labels. The median filter used in this embodiment has a window size of 101.

[0106] In this embodiment, step S3 is specifically as follows:

[0107] A supervised deep neural network is built on the feature extractor. The input data is the broadband signal power spectrum dataset after preprocessing in step S1, and the label is Y obtained in step S2. The supervised neural network is trained using binary cross-entropy until the convergence and error values ​​are optimal to obtain the preliminary broadband signal detection result G.

[0108] The supervised deep neural network is specifically as follows:

[0109] Input the preprocessed broadband signal power spectrum data, and obtain the output of each residual module based on the feature extractor obtained in step S1.

[0110] Where i represents the output of the i-th residual module, n represents the number of channels of the output feature, and L i This represents the output length of the i-th residual module.

[0111] Then, f i Upsampled to the length of the input power spectrum, the feature is obtained.

[0112] Where, p i Indicates f i Features obtained after upsampling.

[0113] Then, subtract p5 from p1, p2, p3, and p4 respectively to obtain p′. i ; will p′ i Through a convolutional layer and a sigmoid function, and the corresponding p i Multiplying them together, we get P. i .

[0114] Where, p′ i Indicates the corresponding p i The difference between p5 and p5; P i p i With p i The product of ′.

[0115] Finally, P i By concatenating the data along the channel dimension and calculating the mean, we obtain...

[0116] Here, G represents the output of the supervised network, i.e., the preliminary broadband signal detection result.

[0117] In this embodiment, step S4 is specifically as follows:

[0118] During the training of the supervised neural network based on step S3, starting from the third optimization of the network parameters, before each optimization, the previous pseudo-labels are fine-tuned using the new output, as shown in the following expression:

[0119] Y j+1 =λG j +(1-λ)Y j

[0120] Where j represents the j-th update of network parameters, Y j+1 G represents the pseudo-label used in the (j+1)th update of network parameters. j Y represents the output of the j-th network. j Y represents the pseudo-label used in the j-th update of network parameters. 1 =G, where λ represents the hyperparameter, i.e., the degree of each update.

[0121] Finally, based on the actual situation, set the maximum number of optimization attempts, optimize the network parameters until the specified number of optimization attempts are completed, and complete the broadband signal detection based on unsupervised deep neural networks.

[0122] Based on the above, this embodiment also includes simulation experiments to further illustrate the method of the present invention.

[0123] The hardware environment for the simulation experiment in this embodiment is: GeForce RTX 3080Ti, Intel(R)Xeon(R)Bronze3204CPU@1.90GHz*12.

[0124] The software environment for the simulation experiment in this embodiment is: Ubuntu 20.04.2LTS, Visual Studio Code, Python 3.8, and cuda 12.0.

[0125] The input data used in the simulation experiment of this embodiment is as follows: 10,000 power spectra of broadband signals containing the above modulation types are obtained by simulating and preprocessing with broadband signal simulation code, and these are used as the training set; at the same time, 100 power spectra are generated and used as the test set. The length of each broadband power spectrum is 16384, and the signal-to-noise ratio ranges from 6dB to 15dB and is randomly distributed.

[0126] The parameters for training the simulation experiment network in this embodiment are as follows: the batch size of all networks is uniformly 64, the network input shape is 1*16384, the number of iterations is uniformly 20, Adam is used as the optimizer for all networks, and the learning rate is 0.001.

[0127] Figure 3 This is a visualization of the simulation results for broadband signal power spectrum detection. (From...) Figure 3 It can be seen that all subcarriers appearing in the broadband power spectrum were detected and enclosed in the detection frame.

[0128] Table 1 is a quantitative analysis table of the broadband signal detection results in this embodiment. The updated broadband specific signal estimation parameters are evaluated using the following evaluation indicators: true positives (TP), false positives (FP), true negatives (TN), average precision (AP), and average recall (AR).

[0129] Table 1

[0130]

[0131] As can be seen from Table 1, the AP and AR values ​​of the detection results on the test set in this embodiment are both greater than 0.85, proving that the method of the present invention can obtain better broadband signal detection results.

[0132] In summary, the method of this invention utilizes unsupervised deep neural networks for broadband signal detection, which has the advantages of high detection accuracy, good recognition effect, wide applicability, and no need for manual annotation. It solves the problem of existing broadband signal processing methods requiring the annotation of a large amount of accurate data, and is a simple and practical broadband signal detection method.

[0133] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A broadband signal detection method based on unsupervised deep neural networks, the specific steps of which are as follows: S1. Train an autoencoder using the power spectrum data of a broadband signal as a feature extractor; S2. Use a feature extractor to obtain the features of the broadband power spectrum. Build an unsupervised deep neural network based on the feature extractor, train it using the features of the broadband power spectrum, obtain the optimized features, and then perform post-processing to generate initial pseudo-labels. S21. Based on step S1, the obtained preprocessed broadband signal power spectrum dataset is processed by a feature extractor to obtain the corresponding feature vector. S22. Based on the feature extractor, build an unsupervised deep neural network, optimize and adjust the feature vector, and obtain the output features; wherein The unsupervised deep neural network is as follows: Input the pre-processed wideband signal power spectrum data, based on the feature extractor obtained in step S1, obtain the output of the 3rd, 4th and 5th layer residual modules ; Where i represents the output of the i-th residual module. This represents the number of channels in the output feature. This represents the output length of the i-th residual module. Represents the real number field; Then, use SEBlock on Perform feature optimization: 1) In Global average pooling is performed on the last dimension to obtain the output. ; in, This represents the output of global average pooling; 2) Feature compression and expansion are achieved through two fully connected layers; 3) Pass the output of the fully connected layer through the sigmoid function to obtain an output located in the interval (0,1); 4) Connect the output with Multiplying by the channel dimension yields the optimized features. ; The output is then upsampled to the length of the input power spectrum to obtain the output. ; and calculate Rather than the mean in the length dimension difference ; Recalculate The cosine similarity sim with the input broadband power spectrum is calculated. ; in, This represents a symbolic function that outputs 1 when its internal expression is greater than or equal to 0, and 0 otherwise; thus, it obtains the characteristic... : ； in, Representation of features The l-th element of all channels; Finally, the features obtained from different residual modules are... The output features of the unsupervised network are obtained by concatenating the data along the channel dimension and passing it through an SEBlock. And it adopts adaptive decision boundary loss. The network optimization expression is as follows: ； ； in, This represents the sigmoid function; N represents the batch size of the network input. and These represent the L1 distance and L2 distance, respectively. This represents the hyperparameters, indicating the weights of the L1 distance; S23. Based on the broadband signal power spectrum dataset preprocessed in step S1, freeze the autoencoder parameters, train the unsupervised neural network until the convergence and error values ​​are optimal, and save the model at this time as a pseudo-label generator. S24. Post-process the output of the pseudo-tag generator obtained in step S23 to obtain the initial pseudo-tags; The post-processing is specifically as follows: Features obtained based on pseudo-label generator The input is fed into a fully connected conditional random field, and the output is obtained. ; in, This represents the output characteristics of a fully connected conditional random field; Then filter using a median filter. , obtain the initial pseudo-label ; S3. Based on the initial pseudo-labels, train a supervised neural network to obtain preliminary broadband signal detection results; S4. Use the detection results to iteratively fine-tune the pseudo-labels to obtain new pseudo-labels to continue training the supervised neural network. Iterate until the maximum number of optimizations is reached, and finally achieve efficient and accurate broadband signal detection.

2. The broadband signal detection method based on unsupervised deep neural networks according to claim 1, characterized in that, The specific steps of S1 are as follows: First, a broadband IQ signal containing multiple modulation methods is acquired and preprocessed to obtain broadband signal power spectrum data. The modulation schemes include BPSK, QPSK, OQPSK, PI4QPSK, 8QAM, 16QAM, and GMSK; the preprocessing steps include resampling, amplitude adjustment, power spectrum calculation, and normalization; the Pwelch method is used when calculating the power spectrum. Then, an autoencoder model is constructed, using a loss function. Optimize model parameters; the autoencoder model structure includes: an input layer, an encoding layer, and a decoding layer; specifically as follows: Input layer: Designed to receive inputs of shape (1, L), where L represents the length of the input power spectrum; Encoding layer: A residual network structure is adopted, with a total of 5 residual modules. The output length of each residual block is the input length / (2^number of layers). Decoding layer: A residual network structure is adopted, which replaces the Conv1d convolutional layer in the encoding layer with the transposed convolutional layer ConvTranspose1d, and performs upsampling at the corresponding downsampling positions to obtain an output with the same shape as the input data; Finally, based on the preprocessed broadband signal power spectrum dataset, the autoencoder is trained until convergence and the error value is optimal. The encoding layer of the model at this point is then saved as a feature extractor.

3. The broadband signal detection method based on unsupervised deep neural networks according to claim 1, characterized in that, Step S3 is as follows: A supervised deep neural network is built on top of the feature extractor. The input data is the broadband signal power spectrum dataset preprocessed in step S1, and the labels are those obtained in step S2. The supervised neural network was trained using binary cross-entropy until convergence and the error value was optimal, thus obtaining the preliminary broadband signal detection result G. The supervised deep neural network is specifically as follows: Input the preprocessed broadband signal power spectrum data, and obtain the output of each residual module based on the feature extractor obtained in step S1. ; Where i represents the output of the i-th residual module. This represents the number of channels in the output feature. This represents the output length of the i-th residual module; Then, Upsampled to the length of the input power spectrum, the feature is obtained. ; in, Indicates will Features obtained after upsampling; Then the first , , , respectively with To do the work, one gets ;Will Through a convolutional layer and a sigmoid function, and the corresponding Multiply, we get ; in, Indicates the corresponding and The difference; express and The product; Finally, By concatenating the data along the channel dimension and calculating the mean, we obtain... ; in, This represents the output of the supervised network, i.e., the preliminary broadband signal detection results.

4. The broadband signal detection method based on unsupervised deep neural networks according to claim 1, characterized in that, Step S4 is as follows: During the training of the supervised neural network based on step S3, starting from the third optimization of the network parameters, before each optimization, the previous pseudo-labels are fine-tuned using the new output, as shown in the following expression: ； Where j represents the j-th update of network parameters, This represents the pseudo-label used in the (j+1)th update of network parameters. This represents the output of the j-th network iteration. This represents the pseudo-label used in the j-th update of network parameters. , This represents the hyperparameter, i.e., the degree of update in each iteration; Finally, based on the actual situation, set the maximum number of optimization attempts, optimize the network parameters until the specified number of optimization attempts are completed, and complete the broadband signal detection based on unsupervised deep neural networks.

Images