An interference signal identification method based on two-stage fine tuning

Abstract

The application discloses a two-stage fine-tuning-based interference signal identification method and belongs to the technical field of communication.The TSFTJR-Net interference signal identification network model is designed and built, and the detection accuracy of interference types and interference frequency ranges is significantly improved; a target detection algorithm is introduced, the position of the interference signal in the interference time-frequency diagram is predicted, and the frequency range of the interference signal is obtained; two-stage training is used in model training, including basic model training and fine-tuning model training, so that the model can still maintain good generalization performance under the condition of a small amount of new class data.The algorithm has obvious advantages compared with existing algorithms, and can solve the problems of insufficient feature extraction of traditional interference identification methods and the need for a large amount of training data of existing deep learning-based interference identification methods.

Description

Technical Field

[0001] This invention relates to a method for identifying interference signals in a wireless communication system, specifically a two-stage fine-tuning jam recognition (TSFTJR) method, belonging to the field of communication technology. Background Technology

[0002] With the continuous development of communication technology, from the initial wired telephone to the emergence of the Internet, and now to the widespread use of mobile wireless devices, the application of communication technology has brought tremendous changes to people's lives, propelling society into the information age. Wireless communication is an important research direction in the current field of communication. One serious problem affecting communication security and quality in wireless communication is interference. Malicious parties send interference signals to disrupt one's own wireless communication link, making it impossible for one's own party to receive accurate information. Therefore, it is necessary to analyze and study interference in communication to ensure communication security. Interference identification is the prerequisite and foundation for achieving intelligent anti-interference. In communication, if the type of interference can be identified in a timely and accurate manner, the receiving end can take appropriate anti-interference measures to suppress or avoid interference as much as possible, thereby reducing the impact of interference on communication quality. However, interference in wireless communication is generally not of one type; multiple types of interference often affect normal communication. On the other hand, it is difficult to collect enough malicious interference signal samples in real-world scenarios, which significantly affects the performance of interference signal identification algorithms. Therefore, research on interference identification technology is extremely important.

[0003] Traditional interference identification techniques can be broadly categorized into two types, both requiring data preprocessing and manual extraction of interference feature parameters before identification. The first type uses threshold comparison to classify interference signals; the second uses machine learning algorithms. Methods involving manual extraction of interference feature parameters typically employ decision trees, support vector machines, or backpropagation (BP) neural networks. However, these algorithms often suffer from loss of useful information due to improper extraction methods, ultimately affecting the accuracy of interference classification. Clearly, to meet the communication requirements for accurate interference signal identification, interference identification technology needs further improvement and optimization.

[0004] With the rapid development of artificial intelligence technology, researchers have introduced deep learning into interference identification technology. Compared with traditional interference identification algorithms, deep learning-based interference signal identification can better extract the features of interference data, solving the problem of insufficient data extraction in traditional interference identification algorithms. Some scholars have combined two different methods, LSTM (Long-Short-Term Memory) and CNN (Convolutional Neural Network), which outperforms individual networks in interference signal identification. Others have used time-frequency analysis with a smoothed pseudo-Wigner-Ville distribution to transform the time-frequency feature domain, generating time-frequency images, and then using lightweight convolutional neural networks to extract interference signal features from these images, thus achieving automatic, accurate, and rapid identification of common interference signals. However, most of the aforementioned deep learning-based interference identification algorithms require a large amount of data for model training; without sufficient data, the model performance is unsatisfactory. Furthermore, these algorithms lack generalization ability for unknown interference, and their generalization capabilities need to be improved.

[0005] This invention relates to the concepts of datasets and models, which will be introduced in turn below:

[0006] Base class and new class: The algorithm of this invention uses a training dataset during training. The training dataset contains different types of interference. The types of interference are divided into two types: base class and new class. The base class interference participates in the first stage of training of the algorithm, while the new class is the unknown type that the algorithm needs to learn in the second stage.

[0007] Base class model and base class model training set: The base class model training set is used for the first stage of base class model training in the algorithm, and it consists of base class interference data. The base class model is the model obtained from the first stage of training in the algorithm.

[0008] Fine-tuning the model and its training set: The fine-tuning model training set is used for the second stage of the algorithm's fine-tuning training. It takes the form Nway Kshot, indicating that the dataset has N classes of data, with K training data points for each class. It consists of base class interference data and new class interference data. The fine-tuned model is the model obtained from the second stage of the algorithm's training. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to overcome the shortcomings and defects of the prior art and provide an interference signal identification method based on two-stage fine-tuning, which solves the problems of insufficient training data and poor generalization of existing interference signal identification algorithms.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] An interference signal identification method based on two-stage fine-tuning includes the following steps:

[0012] Step 1: Generate a base class model training set, a fine-tuning model training set, and a test dataset. The base class model training set includes time-frequency plots of interference signals of basic types, the fine-tuning model training set includes time-frequency plots of interference signals of basic types and newly added types, and the test dataset includes time-frequency plots of interference signals generated in an environment with a signal-to-interference-plus-noise ratio of 0-10dB.

[0013] Step 2: Construct an interference signal identification network model. The model obtains the frequency range of the interference signal by predicting the location of the interference signal in the interference time-frequency map, and at the same time obtains the interference category corresponding to the interference location.

[0014] Step 3: Input the base class model training set into the constructed interference signal recognition network model to perform the first stage of base class model training; and use the test dataset to test the converged model. After passing the test, the trained interference signal recognition network base class model is obtained.

[0015] Step 4: Input the fine-tuning model training set into the base class model of the interference signal recognition network to perform the second stage of fine-tuning model training; and use the test dataset to test the converged model. After passing the test, the trained fine-tuned model of the interference signal recognition network is obtained.

[0016] Step 5: Use the interference signal identification network fine-tuning model to identify interference signals.

[0017] Furthermore, the specific process of step 1 is as follows:

[0018] At the receiving end, a short-time Fourier transform is performed on the received signal to obtain the time-frequency map of the interference signal. Then, image preprocessing operations are performed on the interference time-frequency map, including mean filtering and exponentiation, to generate time-frequency maps of 6 typical interference signals, including sweep interference, single-tone interference, multi-tone interference, comb interference, partial bandwidth interference, and narrowband interference. The labels of the interference signal time-frequency map are the interference category and the interference location. The interference signal time-frequency map and the labels are combined to form a dataset.

[0019] Four basic types of interference signal time-frequency maps are randomly selected from the dataset as the base class model training set, and the remaining two interference types are used as new types. The interference signal time-frequency maps of the basic types and the new types form the fine-tuning model training set. In the fine-tuning model training set, each type of interference signal time-frequency map has the same specified number of training data, and the number of basic types is greater than the number of new types.

[0020] Interference signal time-frequency diagrams were generated under signal-to-interference-plus-noise ratios of 0dB, 2dB, 4dB, 6dB, 8dB, and 10dB, respectively, and used as test sets.

[0021] Furthermore, the interference signal identification network model includes a backbone network, a region proposal network, an ROI pooling layer, an ROI feature extraction layer, an interference classifier, and an interference location regressor.

[0022] The input data for the backbone network is the time-frequency map of the interference signal. After the input data passes through the backbone network, a feature map is extracted. The feature map is divided into W×H grid cells, where W and H represent the width and height of the feature map, respectively. In each grid cell, a set of anchor boxes are used as prior boxes for interference location regression. Each anchor box is represented by a 4-dimensional vector (x, y, w, h), which represents the center coordinates of the interference location rectangle and the width and height of the rectangle, respectively. In each grid cell, the interference location regression variable is output with an offset of 4×k relative to the anchor box position, where k represents the number of anchor boxes in each grid cell. The feature map is then input into the region proposal network and the ROI pooling layer.

[0023] The classifier of the region proposal network predicts 2×k confidence scores for each grid cell of the feature map, indicating the probability of k anchor boxes having interference and not having interference, respectively; the detector of the region proposal network regresses W×H×k interference locations, and then retains the candidate location boxes that meet the requirements through the non-maximum suppression algorithm; the output of the region proposal network is input to the ROI pooling layer.

[0024] The ROI pooling layer uses ROIAlign for pooling, and the pooled features are then input into the ROI feature extraction layer.

[0025] The ROI feature extraction layer outputs a feature vector after passing through two 3×3 convolutional layers and two fully connected layers. This feature vector is then input into the interference signal classifier and the interference location regressor.

[0026] The feature vectors are processed by an interference signal classifier and an interference location regressor to obtain the type and location of the interference signal.

[0027] Furthermore, in the second stage of fine-tuning the model training in step 4, only the parameters of the interference classification detector and the interference location regressor of the base class model are fine-tuned, while other parameters remain unchanged.

[0028] Furthermore, in steps 3 and 4, the offline training optimization method is Adam, where the loss function selected for the interference signal category is the sum of the cross-entropy loss of the predicted value of the output interference signal category and the true value of the input interference signal category; the loss function selected for the interference signal frequency range is the sum of the Smooth L1 loss of the predicted value of the output interference signal location and the true value of the input interference signal location.

[0029] The beneficial effects of this invention are as follows:

[0030] (1) Compared with traditional interference identification methods, the TSFTJR-Net of this invention is an interference identification network built based on deep learning ideas. Compared with manual feature extraction techniques, it can make fuller use of interference signal data.

[0031] (2) Compared with existing deep learning-based interference identification methods, the TSFTJR-Net of this invention enables the model to be trained on a small amount of new interference data through two-stage fine-tuning training, thereby improving the interference identification accuracy and generalization of the algorithm. Attached Figure Description

[0032] Figure 1 This is a diagram of the first-stage interference signal identification structure of the TSFTJR-Net of the present invention.

[0033] Figure 2 This is a diagram of the second-stage interference signal identification structure of the TSFTJR-Net of the present invention.

[0034] Figure 3 This is a line graph showing the accuracy of new interference recognition on test datasets with signal-to-interference-plus-noise ratios (SNRs) of 0–10 dB, obtained by training the fine-tuned model of TSFTJR-Net of the present invention through 1–15 shots.

[0035] Figure 4 The graph shows the average accuracy of basic interference recognition, the average accuracy of new interference recognition, and the average accuracy of all interference recognition on a test dataset with a signal-to-interference-plus-noise ratio (SNR) of 0-10 dB, obtained by training the fine-tuned model of TSFTJR-Net of this invention through 1-15 shots.

[0036] Figure 5 This is a line graph showing the average interference recognition accuracy of the fine-tuned model of TSFTJR-Net trained with 1 to 15 shots and several reference algorithms on datasets with a signal-to-interference-plus-noise ratio of 0 to 10 dB. Detailed Implementation

[0037] The following explanation, in conjunction with the accompanying drawings and specific embodiments, provides further details:

[0038] An interference signal identification method based on two-stage fine-tuning includes the following steps:

[0039] Step 1: Generate a base class model training set, a fine-tuning model training set, and a test dataset. The base class model training set includes time-frequency plots of interference signals of basic types, the fine-tuning model training set includes time-frequency plots of interference signals of basic types and newly added types, and the test dataset includes time-frequency plots of interference signals generated in an environment with a signal-to-interference-plus-noise ratio of 0-10dB.

[0040] The specific process is as follows:

[0041] At the receiving end, a short-time Fourier transform is performed on the signal to obtain the time-frequency map of the interference signal. Mean filtering is applied to the interference time-frequency map with a 3×3 rectangular convolution kernel. Then, a power-law operation is performed on the filtered interference time-frequency map to enhance its contrast. This preprocessing smooths out noise points and enhances the contrast of the time-frequency map. Under a random signal-to-interference-plus-noise ratio (SNR) of 0–10 dB, six types of interference signal time-frequency maps are generated for common typical interferences, including sweep interference, single-tone interference, multi-tone interference, comb interference, partial bandwidth interference, and narrowband interference. The labels on the interference time-frequency maps are the interference type and location. The interference time-frequency maps and labels together constitute the dataset.

[0042] Four basic types of interference signal time-frequency maps are randomly selected from the dataset as the base class model training set. The remaining two interference types are used as new types. The time-frequency maps of the basic types and the new types of interference signals form the fine-tuning model training set. In the fine-tuning model training set, each type of interference signal time-frequency map has the same specified number of training data, and the number of basic types is greater than the number of new types. In this example, the base class model training set includes 4 base classes, and the fine-tuning model training set has 6 forms. The N of each fine-tuning model training set is 6, and the K are 1, 2, 3, 5, 10, and 15, respectively.

[0043] Interference signal time-frequency diagrams were generated under signal-to-interference-plus-noise ratios of 0dB, 2dB, 4dB, 6dB, 8dB, and 10dB, respectively, and used as test sets.

[0044] Step 2: Construct the TSFTJR-Net interference signal recognition network model. This model is based on a deep learning target detection network. The network obtains the frequency range of the interference signal by predicting the location of the interference signal in the interference time-frequency map.

[0045] The TSFTJR-Net network structure includes a backbone network, a region proposal network, ROI (Region of Interest) pooling layers, ROI feature extraction layers, an interference classifier, and an interference location regressor. The backbone network uses the ResNet50 residual network structure as the basic feature extraction structure for the time-frequency map of the interference signal. The region proposal network consists of three convolutional layers: an intermediate convolutional layer, an ROI regressor, and an ROI classifier. The ROI pooling layer uses the ROIAlign algorithm for pooling. The ROI feature extraction layer extracts the depth features of the ROI. The last two layers are the interference signal classifier and the interference signal location regressor, composed of convolutional layers and fully connected layers, respectively, which output the type and location of the interference signal.

[0046] Step 3: Input the basic model training set generated in Step 1 into the TSFTJR-Net constructed in Step 2 for training the base class model of the first-stage interference signal recognition network. The training process is an end-to-end offline training process with 15 epochs. The batch size of data read each time is 16, and the initial learning rate is 0.01. The Adam optimization algorithm is used to optimize the training process to minimize the loss function. During the base class model training process, the four interference types in the dataset are used as base classes, and three sets of base class models are trained sequentially. Three different TSFTJR-Net base class models are obtained in this stage and participate in the training of the next stage of fine-tuning the model. The results are shown in Table 1. Table 1 shows the average prediction accuracy of the TSFTJR-Net model in the base class model training stage of this invention. The average detection accuracy of the four base classes is 92.67%.

[0047] Table 1

[0048] NwayKshot Category settings AP50 (%) bAP50 (%) nAP50 (%) - 4 base classes 92.67 92.67 -

[0049] Step 4: Input the fine-tuned model training set generated from the training dataset in Step 1 into the TSFTJR-Net base class model obtained in Step 3 for the second-stage fine-tuning training of the interference signal recognition network. Use N×K data points from the fine-tuned model training set to train the basic model parameters. The training process is an end-to-end offline training process with 20 epochs. The batch size of data read each time is 16, and the initial learning rate is 0.001. The Adam optimization algorithm is used to optimize the training process to minimize the loss function. During training, only the parameters of the interference classification detector and interference location regressor of the base class model are fine-tuned, while other parameters remain unchanged, resulting in the trained fine-tuned model of the interference signal recognition network, as shown below. Figure 2 As shown.

[0050] Furthermore, step 2 specifically involves constructing, as follows: Figure 1The TSFTJR-Net model shown takes a time-frequency image of the interference signal as input data. After passing through the backbone network, a feature map is extracted. The feature map is divided into W×H grid cells (W and H are both set to 13), where W and H represent the width and height of the feature map, respectively. Within each grid cell, a set of anchor boxes, which are actually a set of horizontally angled rectangles, are used as prior boxes for interference location regression. k represents the number of anchor boxes per grid cell; in network training, k = 9. Each location box is represented by a 4-dimensional vector (x, y, w, h), representing the center coordinates of the interference location rectangle and the width and height of the rectangle, respectively. In each grid cell, the interference location regression variable outputs a 4×k offset relative to the anchor box position. The region proposal network's classifier predicts 2×k confidence scores for each grid cell of the feature map, indicating the probability of the k anchor boxes containing interference and the probability of not containing interference. For each ROI, the region proposal network's detector regresses W×H×k interference locations, and then uses a non-maximum suppression algorithm to retain suitable ROIs and candidate location boxes.

[0051] The feature map and the output of the region proposal network are input into the ROI pooling layer, using ROI Align for pooling, outputting a feature vector W×H, where W and H are both 7. The pooled feature is then input into the ROI feature extraction layer, which passes through two 3×3 convolutional layers and two fully connected layers to output a feature vector. Finally, the feature vector is passed through an interference signal classifier and an interference signal location regressor to output the interference signal type and location. The interference signal type is 512×7, where 7 represents the confidence level of each type predicted for this ROI, and the interference signal location is 512×7×4, where 7×4 represents the predicted interference location parameters for each type within this ROI.

[0052] Furthermore, the loss function for steps 3 and 4 is defined as follows:

[0053] L = L cls +λL pos (1)

[0054] In the formula, L cls For the classification cross-entropy loss of interference types, L pos To mitigate the interference between the regression position and the true value, the Smooth-L1 loss is used, and λ is used to adjust the weight of the loss. The specific definition is as follows:

[0055]

[0056] Interference type loss part L cls for:

[0057]

[0058] In the formula, p i This indicates the probability that the proposed region is predicted as an interference region, and the label indicates the region is true when it is true in the label. It is 1 if it is true, otherwise it is 0.

[0059] Interference location loss L pos for:

[0060]

[0061] In the formula, t i The coordinates (x, y, w, h) represent the four parameters of the bounding box detected by the network. (x, y) represent the center point of the predicted object bounding box, (w, h) represent the width and height of the object bounding box, and t represents the height of the object bounding box. i T This is the truth label of the object rectangle corresponding to that area.

[0062] Furthermore, in the fine-tuning model training in step 3, two classes are used as new classes to be learned by the base model. Fine-tuning model training is performed on a training set of 1–15 shots, and the resulting fine-tuned models are tested on a test set with a signal-to-interference-plus-noise ratio (SNR) of 0–10 dB. Three evaluation metrics were used for the algorithm: AP50 (Average Precision), representing the average accuracy with an IOU (Intersection over Union) greater than 50% between the predicted bounding box and the ground truth box of the object detection; bAP50 (base class Average Precision), representing the average accuracy with an IOU greater than 50% between the predicted bounding box and the ground truth box of the base class object detection; and nAP50 (novel class Average Precision), representing the average accuracy with an IOU greater than 50% between the predicted bounding box and the ground truth box of the novel class object detection.

[0063] like Figure 3 and Figure 4As shown, the interference detection accuracy is similar on interference signal datasets with a signal-to-interference-plus-noise ratio (SNR) of 2–10 dB, while the accuracy on the 0 dB interference signal dataset differs by about 1%. However, with 15 shots, the interference signal detection accuracy for 0–10 dB reaches over 90%. Table 2 shows the average accuracy of basic interference recognition, new interference recognition, and all interference recognition classes on test datasets with an SNR of 0–10 dB, obtained by training the fine-tuned model of TSFTJR-Net of this invention with 1–15 shots. Table 2 calculates the average prediction accuracy of interference detection at 0–10 dB. It can be seen that as the number of shots gradually increases, bAP50 and nAP50 also gradually improve. When fine-tuning training is performed using a small sample size of 15 shots, the target detection accuracy for both new classes reaches over 90%, and bAP0 also reaches over 90%. The average prediction accuracy for target detection of all six classes is greater than 91%, which is a good result.

[0064] Table 2

[0065] NwayKshot AP50 (%) bAP50 (%) nAP50 (%) 6way1shot 78.59 83.86 68.05 6way2shot 86.19 88.98 79.87 6way3shot 88.63 89.86 85.56 6way5shot 89.43 91.04 86.21 6way10shot 90.88 91.80 88.71 6way15shot 91.55 92.40 90.22

[0066] To verify the effectiveness of the algorithm, the TSFTJR algorithm was compared with the SSD algorithm, Faster R-CNN algorithm, YOLO v4 algorithm, and YOLO v5 algorithm. The results are shown in Table 3. Figure 5 As shown in Table 3, the average interference recognition accuracy of the models generated by the TSFTJR-Net of this invention and the comparison algorithms under 1 to 15 shots is presented. The data in the table shows that the detection accuracy of all algorithms improves with the increase of the number of K-shots. In particular, the TSFTJR algorithm significantly outperforms other reference algorithms in the environment with a signal-to-interference-plus-noise ratio (SNR) of 0 to 10 dB, indicating that the TSFTJR algorithm has a performance advantage over other deep learning-based object detection algorithms with the same small number of samples.

[0067] Table 3

[0068] NwayKshot TSFTJD SSD FasterRcnn YOLOv4 YOLOv5 6way1shot 78.59 37.19 60.86 16.37 16.50 6way2shot 86.19 37.51 76.77 17.12 16.41 6way3shot 88.63 36.60 77.25 30.11 37.33 6way5shot 89.43 35.51 77.27 32.54 44.59 6way10shot 90.88 37.27 79.75 51.92 77.98 6way15shot 91.55 38.05 87.63 80.55 86.88

[0069] In summary, addressing the issues of insufficient feature extraction in traditional interference identification methods and the need for large amounts of training data in existing deep learning-based interference identification methods, this invention designs and constructs the TSFTJR-Net interference detection network model. This model combines interference detection with target detection, transforming interference bandwidth detection into the detection of interference locations on the interference time-frequency map, achieving a significant performance improvement in interference detection accuracy. The model training employs a two-stage approach, including basic model training and fine-tuning model training, enabling the model to maintain good generalization performance even with limited new data types. Simulation results demonstrate that this invention has significant advantages over existing algorithms.

Claims

1. A method for identifying interference signals based on two-stage fine-tuning, characterized in that, Includes the following steps: Step 1: Generate a base class model training set, a fine-tuning model training set, and a test dataset. The base class model training set includes time-frequency plots of interference signals of basic types, the fine-tuning model training set includes time-frequency plots of interference signals of basic types and newly added types, and the test dataset includes time-frequency plots of interference signals generated in an environment with a signal-to-interference-plus-noise ratio of 0-10dB. Step 2: Construct an interference signal identification network model. The model obtains the frequency range of the interference signal by predicting the location of the interference signal in the interference time-frequency map, and at the same time obtains the interference category corresponding to the interference location. Step 3: Input the base class model training set into the constructed interference signal recognition network model to perform the first stage of base class model training; and use the test dataset to test the converged model. After passing the test, the trained interference signal recognition network base class model is obtained. Step 4: Input the fine-tuning model training set into the base class model of the interference signal recognition network to perform the second stage of fine-tuning model training; The converged model was tested using a test dataset. Once the test was passed, the fine-tuned model of the interference signal recognition network was obtained. Step 5: Use the interference signal identification network fine-tuning model to identify interference signals.

2. The interference signal identification method based on two-stage fine-tuning according to claim 1, characterized in that, The specific process of step 1 is as follows: At the receiving end, a short-time Fourier transform is performed on the received signal to obtain the time-frequency map of the interference signal. Then, image preprocessing operations are performed on the interference time-frequency map, including mean filtering and exponentiation, to generate time-frequency maps of 6 typical interference signals, including sweep interference, single-tone interference, multi-tone interference, comb interference, partial bandwidth interference, and narrowband interference. The labels of the interference signal time-frequency map are the interference category and the interference location. The interference signal time-frequency map and the labels are combined to form a dataset. Four basic types of interference signal time-frequency maps are randomly selected from the dataset as the base class model training set, and the remaining two interference types are used as new types. The interference signal time-frequency maps of the basic types and the new types form the fine-tuning model training set. In the fine-tuning model training set, each type of interference signal time-frequency map has the same specified number of training data, and the number of basic types is greater than the number of new types. Interference signal time-frequency diagrams were generated under signal-to-interference-plus-noise ratios of 0dB, 2dB, 4dB, 6dB, 8dB, and 10dB, respectively, and used as test sets.

3. The interference signal identification method based on two-stage fine-tuning according to claim 2, characterized in that, The interference signal identification network model consists of a backbone network, a region proposal network, a region of interest (ROI) pooling layer, a region of interest (ROI) feature extraction layer, an interference classifier, and an interference location regressor. The input data for the backbone network is the time-frequency map of the interference signal. After the input data passes through the backbone network, a feature map is extracted. The feature map is divided into W×H grid cells, where W and H represent the width and height of the feature map, respectively. In each grid cell, a set of anchor boxes are used as prior boxes for interference location regression. Each anchor box is represented by a 4-dimensional vector (x, y, w, h), which represents the center coordinates of the interference location rectangle and the width and height of the rectangle, respectively. In each grid cell, the interference location regression variable is output with an offset of 4×k relative to the anchor box position, where k represents the number of anchor boxes in each grid cell. The feature map is then input into the region proposal network and the ROI pooling layer. The classifier of the region proposal network predicts 2×k confidence scores for each grid cell of the feature map, indicating the probability of k anchor boxes having interference and not having interference, respectively; the detector of the region proposal network regresses W×H×k interference locations, and then retains the candidate location boxes that meet the requirements through the non-maximum suppression algorithm; the output of the region proposal network is input to the ROI pooling layer. The ROI pooling layer uses ROIAlign for pooling, and the pooled features are then input into the ROI feature extraction layer. The ROI feature extraction layer outputs a feature vector after passing through two 3×3 convolutional layers and two fully connected layers. This feature vector is then input into the interference signal classifier and the interference location regressor. The feature vectors are processed by an interference signal classifier and an interference location regressor to obtain the type and location of the interference signal.

4. The interference signal identification method based on two-stage fine-tuning according to claim 3, characterized in that, In the second stage of model training in step 4, only the parameters of the interference classification detector and the interference location regressor of the base class model are fine-tuned, while other parameters remain unchanged.

5. The interference signal identification method based on two-stage fine-tuning according to claim 1, characterized in that, In steps 3 and 4, the offline training optimization method is Adam. The loss function selected for the interference signal category is the sum of the cross-entropy loss of the predicted value of the output interference signal category and the true value of the input interference signal category. The loss function selected for the interference signal frequency range is the sum of the Smooth L1 loss of the predicted value of the output interference signal location and the true value of the input interference signal location.

Images