Ultra-short wave broadband satellite signal detection method based on superimposed spectrum and deep learning

Abstract

This invention discloses a method for detecting ultra-shortwave broadband satellite signals based on superimposed spectrum and deep learning. Firstly, this invention enhances the distinction between ultra-shortwave signals and noise floor through spectrum superposition, while suppressing the influence of short burst interference, effectively improving signal characteristics under low signal-to-noise ratio conditions. Simultaneously, it proposes an ultra-shortwave signal feature extraction backbone network, SST-Backbone (SC-ST-Backbone), which integrates multi-layer convolutional layers and a four-layer SC (shortcut)-ST (Swin Transoformer) layer. The SC-ST module in the backbone network incorporates the global focus advantage of the Transformer, combined with a residual multi-layer convolutional module focusing on local features to increase the network depth and receptive field. This enables the network model to more accurately detect broadband ultra-shortwave signals in the target frequency band, efficiently removing noise floor interference while reducing focus on anomalous signal features, improving the detection accuracy of broadband ultra-shortwave target signals in complex environments, and reducing the false alarm rate.

Description

Technical Field

[0001] This invention relates to the application of deep learning in the field of signal detection, specifically a method for detecting ultra-shortwave satellite signals based on superimposed spectrum and deep learning networks. Background Technology

[0002] Ultra-short wave (UHF) signals are a frequency band widely used in communication and broadcasting. Due to their excellent transmission performance, UHF plays a crucial role in the field of communication. The unique properties of this shortwave signal make it an ideal choice for applications such as radar, communication, navigation, and remote control. However, with the increasing complexity of the electromagnetic environment, interference can occur in UHF communication, potentially being exploited for network intrusion, electronic warfare, and other purposes. Therefore, signal detection of targets is of paramount importance.

[0003] Spectrum detection of broadband UHF signals requires processing a wider spectral range, and is characterized by dense signal distribution across the spectrum, as well as the influence of short burst interference signals. Currently, the main traditional methods include:

[0004] Spectral energy detection methods typically involve integrating the energy of the received signal and setting an energy threshold. If the integrated energy exceeds the threshold, the system determines that a signal has been detected. Energy detection is a relatively simple technique, often used to determine the presence of a signal in a single channel. It is applicable to various signal types and has low computational complexity. However, in broadband UHF signals, the bandwidth is wide, various channel types exist, and different channels interfere with each other in the spectrum. Furthermore, environmental noise fluctuates, making energy detection highly sensitive to noise and interference. This can easily lead to false detections, where noise or interference is mistaken for a signal, sometimes resulting in reduced detection performance. It is therefore unsuitable for broadband signal detection. Another method in broadband UHF signals is matched filter detection. This method usually requires prior knowledge or an accurate reference signal, meaning the characteristics of the expected signal must be understood before detection. This is difficult to implement in the case of dense broadband UHF signals, especially when the signal has uncertain or dynamically changing characteristics. Moreover, the broadband UHF band contains a large number of interference signals, making its detection performance unsuitable.

[0005] In summary, traditional signal detection methods primarily focus on determining the presence or absence of signals in narrowband environments. Furthermore, traditional signal detection typically requires substantial prior knowledge of the target signal and noise characteristics, including waveform, frequency, and time delay information, while also demanding high accuracy in parameter calculations. This leads to performance degradation in unknown or highly complex signal environments, resulting in difficulties in fully acquiring or even losing prior information, and insufficient computational precision. Consequently, these methods are ill-suited for situations involving a large number of densely packed signals in broadband UHF / UHF signals.

[0006] With the development of deep learning, deep learning-based object detection technology has achieved remarkable results in various fields, including computer vision and signal processing. Applying this technology to the detection of broadband UHF signal spectrum can provide a novel and efficient method for identifying and analyzing signal activity, anomalies, or potential threats in the spectrum. Summary of the Invention

[0007] This method proposes an ultra-shortwave broadband satellite signal detection method based on superimposed spectrum and deep learning, which transforms the traditional broadband satellite ultra-shortwave detection problem into an image detection and processing problem, achieving a higher detection rate while reducing the false alarm rate and improving classification and recognition performance.

[0008] In the complex electromagnetic environment of the 230-270MHz UHF band, traditional energy detection methods suffer from missed detections and high false alarm rates for broadband satellite signals. This invention proposes a method for broadband UHF signal detection based on a ShortcutSwin Transoformer YOLOV5s (SST-YOLOV5s) network with spectrum superposition. This method effectively solves the problems of difficulty in detecting broadband UHF satellite signals under low signal-to-noise ratios and high false alarm rates under abnormal interference conditions in traditional methods. First, spectrum superposition enhances the distinction between UHF signals and background noise, suppresses the influence of short burst interference, and effectively improves the characteristics of target signals under low signal-to-noise ratios. Simultaneously, a four-layer SC (shortcut)-ST (Swin-Transoformer) and multi-layer convolutional cascaded UHF signal feature extraction backbone network SST-Backbone (SC-ST-Backbone) is proposed. In the backbone network, the SC-ST module utilizes the global focus of Transformer to increase the network depth and receptive field by combining it with residual multilayer convolution modules that focus on local features. This enables the network model to more accurately detect broadband UHF signals in the target frequency band. It can efficiently remove the interference of background noise features, while reducing the focus on the features of abnormal signals, thus improving the detection accuracy of broadband UHF target signals in complex environments and reducing the false alarm rate.

[0009] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0010] Step 1: Acquire target satellite signal data using an ultra-shortwave antenna, and then process the acquired ultra-shortwave signal using a receiver. The receiver filters the ultra-shortwave frequency band of the acquired ultra-shortwave signal to obtain broadband ultra-shortwave signal data centered on the target frequency.

[0011] Step 2: After processing the broadband UHF signal data by the receiver, the broadband spectrum of the UHF signal is obtained by fast Fourier transform. Then, the superimposed spectrum of the broadband spectrum of the UHF signal is constructed by superimposing the spectrum. By continuously iterating and accumulating the small differences between the signal and the noise floor, the distinction between the signal and the noise floor is highlighted. The superimposed spectrum is then normalized and preprocessed.

[0012] Step 3: Construct an ultra-shortwave broadband signal detection model based on SST-YOLOV5s to improve the detection accuracy of broadband ultra-shortwave satellite signals by the deep learning network model and reduce the false alarm rate.

[0013] Step 4: Input the normalized preprocessed superimposed spectra of different patterns existing in the broadband of the target satellite band into the SST-YOLOV5s broadband ultra-shortwave satellite detection model for simulation testing of detection effect, and perform statistical comparison calculation on the detection results.

[0014] Step 1, antenna acquisition and receiver signal processing, is implemented as follows:

[0015] The target satellite's VHF / UHF signal was acquired and processed in-situ using a dual-element helical antenna and a VHF / UHF receiver. The antenna employed circular polarization with a gain of 15 dBi. The acquired VHF / UHF signal was filtered for the target frequency band by the receiver, resulting in a broadband VHF / UHF signal with a center frequency of 255 MHz and a bandwidth of 40.96 MHz. The antenna was erected at ********, specifically at 120°20′24″E, 30°18′52″N. The antenna acquired VHF / UHF signals from geostationary orbit satellites in two directions: one from the Pacific Ocean (specifically at **°E), and the other from the Indian Ocean (specifically at **°E).

[0016] The specific implementation of the superposition spectrum construction of the broadband spectrum of the ultra-shortwave signal described in step 2 is as follows:

[0017] 2-1. With a center frequency of 255MHz, the target band's ultra-shortwave signal is sampled at a sampling frequency of 40.96MHz. The resulting target signal has a starting frequency of 234.52MHz and an ending frequency of 275.48MHz.

[0018] Because broadband signals contain a large number of dense signals, traditional methods using sliding thresholds for spectrum detection are susceptible to energy diffusion from these dense signals, resulting in a high false alarm rate due to abnormal spectral energy. Furthermore, the fluctuating noise floor in the broadband spectrum makes it difficult to adapt the energy threshold, leading to a significant number of missed signals. Compared to traditional algorithms, deep learning-based broadband UHF signal detection methods offer significantly lower false alarm and false alarm rates. Therefore, this patent employs a deep learning network model to detect target satellite signals by constructing images.

[0019] 2-2. First, the target signal data with a sampling frequency of 40.96MHz is processed. By locally plotting the spectrum of the ultra-shortwave broadband target signal, the resolution of the image is effectively increased. This method uses FFT spectrum plots to be plotted every 100kHz, resulting in a total of 409 spectrum plots, covering the signal spectrum of the entire target frequency band. To obtain the signals present in the target signal, the signals present in each spectrum plot are marked and detected, and the signals and interference are marked and processed. To ensure data consistency and comparability, the image size of the local 100kHz spectrum plots is uniformly normalized to a three-channel 640×640 size.

[0020] At low signal-to-noise ratios, the observable spectral features are limited. In localized areas, signal features are scarce, and a significant portion of the signal is integrated with the noise floor, making it difficult to extract effective features. Furthermore, short bursts of interference can cause them to become exceptionally prominent. Therefore, this patent proposes constructing a superimposed spectrum through spectral overlay. By iteratively accumulating the minute differences between the signal and the noise floor, the distinguishability between the signal and the noise floor is highlighted. Simultaneously, the overlay process effectively averages the energy of short bursts of interference over a given period across multiple time segments, thereby suppressing the interference characteristics of these short bursts.

[0021] 2-3. The spectrum was simulated and plotted using 5, 10, 20 and 30 superpositions to obtain superimposed spectra with different superposition numbers. By observing the signal characteristics and required data volume of the superimposed spectra with different superposition numbers, the superimposed spectrum was finally obtained by simulating and plotting the spectrum using 30 superpositions.

[0022] Step 3: Construction of the SST-YOLOV5s ultra-shortwave broadband signal detection model:

[0023] To acquire various signal patterns present in the target satellite's broadband UHF signal, the SST-YOLOV5s broadband signal detection model was adopted. In the SST-Backbone network, the global receptive field of the Transformer is utilized to enable modules to focus on the global features of the superimposed spectrum. Simultaneously, a multi-layer convolutional module with shortcut(SC)-ST residuals is embedded in the Transformer, allowing the shortcut(SC)-ST module to focus on the global features of the superimposed spectrum. Combining the multi-layer convolutional module with residuals increases the network depth and receptive field, enabling the model to fully integrate the extracted global and local features of the UHF superimposed spectrum. The SPPF module then performs initial information exchange.

[0024] In the SST-Head target detection head, a feature pyramid is generated by multi-scale fusion of global and local features extracted from the SST-Backbone network. This pyramid, combined with the local receptive field characteristics of the C3 module (which lacks residuals) in the SST-Head target detection head, further detects and filters high-semantic-level global features and detailed local features belonging to the superimposed spectrum. After fusing these high-semantic-level global features and detailed local features, a more accurate superimposed spectral feature map is generated. After channel reduction and scaling of the superimposed spectral feature map through the Conv layer, the Detect detector performs target detection on the superimposed spectral feature map at different scales based on grid anchor points. A sigmoid activation function is used to generate more accurate and comprehensive superimposed spectral image prediction box category and location information, ultimately achieving accurate detection of the target signal spectrum.

[0025] Step 4 involves inputting the normalized preprocessed superimposed spectra of different patterns within the broadband band of the target satellite into the SST-YOLOV5s broadband ultra-shortwave satellite detection model for simulation testing of the detection effect. The detection results are then statistically compared and calculated, as follows:

[0026] In complex electromagnetic environments, the spectrum of ultra-shortwave signals is diverse, with various signals existing in the broadband spectrum. Therefore, after the dataset undergoes superimposed spectrum processing, normal signals, densely continuous signals, and target signals with different patterns and characteristics need to be labeled, while anomalous signals are left unlabeled to reduce false alarms caused by them. This method was used to create the ultra-shortwave signal detection dataset. The ultra-shortwave superimposed spectrum dataset contains 5000 images under different conditions, including normal signals, anomalous fluctuations, anomalous bandwidth, and multi-continuous signals. The first 4000 images of the ultra-shortwave superimposed spectrum dataset were used as the training set, and the last 1000 images were used as the test set for model training and testing.

[0027] 4.1 Model Performance Evaluation Indicators

[0028] This patent uses six metrics to evaluate the model during testing: detection rate, false alarm rate, precision, recall, mean average precision (mAP), and model memory usage. It also visualizes the ultra-shortwave superimposed spectral features extracted by the model. The visualized feature maps clearly show the model's focus on the image, whether it has learned the features of the ultra-shortwave superimposed spectral region, and its attention to interference and abnormal signals.

[0029] 4.2 Detection performance

[0030] The SST-YOLOV5s ultra-shortwave superimposed spectrum detection algorithm, YOLOV5s target detection algorithm, and SSD target detection algorithm proposed in the patent were compared in a broadband ultra-shortwave superimposed spectrum normal image test set, which included conventional ultra-shortwave signal images and ultra-shortwave signals with continuous spectrum. At the same time, the model proposed in the patent was used to test the false alarm effect of abnormal signals in an abnormal signal superimposed spectrum test set. Furthermore, the features extracted by the visualization model in the ultra-shortwave conventional signal image library and the ultra-shortwave continuous signal image library were visualized using the three models to observe the effect of the constructed network model on target signal detection.

[0031] The beneficial effects of this invention are as follows:

[0032] 1. It effectively solves the problem of difficulty in detecting ultra-shortwave broadband satellite channels under low signal-to-noise ratio in traditional methods.

[0033] 2. It solves the problem that deep learning models have a low detection rate of target signals and a high false alarm rate of abnormal signals when there is interference or anomalies.

[0034] 3. Compared with traditional spectrograms as input to deep learning network models, the superimposed spectrum obtained by superimposing spectrograms highlights the distinction between ultra-shortwave signals and background noise, while suppressing the influence of short burst interference, effectively improving the characteristics of target signals under low signal-to-noise ratio.

[0035] 4. The proposed SST-YOLOV5s network model constructs an SST-Backbone (SC-ST-Backbone) backbone network through a four-layer SC (shortcut)-ST (Swin-Transformer) cascaded structure with multiple convolutional layers. In the backbone network, the SC-ST module utilizes the global focus of the Transformer on global features, combined with residual multi-layer convolutional modules that focus on local features, to increase the network depth and receptive field. This enables the network model to more accurately detect broadband UHF signals in the target frequency band, effectively remove the interference of background noise features, reduce the focus on the features of abnormal signals, and improve the detection effect of broadband UHF target signals in complex environments.

[0036] Under different signal-to-noise ratio environments, compared with the YOLOV5s target detection algorithm, the SSD target detection algorithm, and the proposed SST-YOLOV5s network, the network proposed in this invention significantly improves the detection accuracy.

[0037] In summary, this invention for detecting broadband UHF satellite signals features high detection rate, low false alarm rate, and strong robustness, and maintains high detection performance even at low signal-to-noise ratios. Attached Figure Description

[0038] Figure 1 Dual-unit spiral antenna;

[0039] Figure 2 Ultra-shortwave receiver;

[0040] Figure 3 Original spectrum, symmetric spectrum, mirror-filled spectrum;

[0041] Figure 4 Different types of channel image filling spectra after capture;

[0042] Figure 5 Channel attention mechanism;

[0043] Figure 6 Network structure residual block structure diagram;

[0044] Figure 7 Visualization of feature maps for classification and recognition;

[0045] Figure 8 Average recognition rate of each method under different signal-to-noise ratios;

[0046] Figure 9 Average recognition rates of the original spectrum and mirror-filled spectrum under different network models;

[0047] Figure 10 This is a schematic diagram illustrating the detection model's focus on identifying abnormal signal types. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] This invention discloses a method for detecting ultra-shortwave broadband satellite signals based on superimposed spectrum and deep learning. Firstly, this invention enhances the distinction between ultra-shortwave signals and noise floor through spectrum superposition, while suppressing the influence of short burst interference, effectively improving signal characteristics under low signal-to-noise ratio conditions. Simultaneously, it proposes an ultra-shortwave signal feature extraction backbone network, SST-Backbone (SC-ST-Backbone), which integrates multi-layer convolutional layers and a four-layer SC (shortcut)-ST (SwinTransoformer) layer. The SC-ST module in the backbone network incorporates the global focus advantage of the Transformer, combined with a residual multi-layer convolutional module focusing on local features to increase the network depth and receptive field. This enables the network model to more accurately detect broadband ultra-shortwave signals in the target frequency band, efficiently removing noise floor interference while reducing focus on anomalous signal features, improving the detection accuracy of broadband ultra-shortwave target signals in complex environments, and reducing the false alarm rate.

[0050] The technical solution adopted in this invention:

[0051] 1.1 VHF / UHF Signal Data Processing

[0052] The target satellite's ultra-shortwave signal is transmitted through a dual-element helical antenna ( Figure 1 (as shown) and VHF receiver ( Figure 2 (As shown) Field data was collected and processed. The center frequency was 255MHz, and the target band's VHF signal was sampled at a sampling frequency of 40.96MHz, resulting in a target signal starting frequency of 234.52MHz and an ending frequency of 275.48MHz. The global broadband VHF spectrum is shown below. Figure 3 As shown.

[0053] Because broadband signals contain a large number of dense signals, traditional methods using sliding thresholds for spectrum detection are affected by the energy diffusion of these dense signals, resulting in a high false alarm rate due to abnormal spectral energy. Furthermore, the fluctuating noise floor in the broadband spectrum makes it difficult to adapt to changes in the energy threshold, leading to a significant number of missed signals. Therefore, this patented solution, compared to traditional detection algorithms, demonstrates a significantly lower false alarm rate and a lower false alarm rate compared to traditional deep learning-based methods for detecting broadband UHF signals.

[0054] First, the snapshot complex signal data with a sampling frequency of 40.96MHz was processed. By plotting the spectrum of the VHF broadband signal in local areas, the resolution of the images was effectively increased. This method plots FFT spectra in 100kHz increments, resulting in a total of 409 images. The format of the local 100kHz VHF spectrum images is as follows: Figure 4 , 5 As shown, Figure 4 The signal exists in the local frequency band of 9.36MHz-9.46MHz. Figure 5 The signal exists in the local frequency band of 14.34MHz-14.44MHz.

[0055] To acquire signals present in the VHF / UHF spectrum, signals present in each 100kHz area were labeled and detected, and the signals and interference were labeled and processed. To ensure data consistency and comparability, the 100kHz spectrum images were all normalized to a three-channel 640×640 size.

[0056] At low signal-to-noise ratios (SNR), the spectral features observable by the naked eye are limited. In localized areas of the spectrum, signal features are scarce and submerged in background noise. Furthermore, a significant portion of the signal is integrated with the background noise, making it difficult to obtain effective features. Simultaneously, short bursts of interference can cause the interference to become exceptionally prominent. Therefore, this patent proposes constructing a superimposed spectrum through spectral superposition. By iteratively accumulating the minute differences between the signal and the background noise, the distinguishability between the signal and the background noise is highlighted. Simultaneously, through superposition processing, the energy of short burst interference signals over a given period can be effectively averaged across multiple time segments, thus suppressing the interference features of short bursts. Simulations were conducted using superposition methods of 5, 10, 20, and 30 times. Based on observations of the image feature effects and the required data volume, a superposition method of 30 times was ultimately used to plot the superimposed spectrum, as shown below. Figure 6 As shown, the spectral features after 30 superpositions are more pronounced than the original spectrum. Therefore, superposition is used for processing, and the superimposed spectral signal obtained after normalization has more prominent features, providing a more effective feature map for target detection in subsequent network models.

[0057] 1.2 Construction of SST-YOLOV5s Network Model

[0058] The framework of the broadband UHF spectrum detection method based on SST-YOLOV5s is as follows: Figure 7 As shown. Leveraging the Transformer's ability to focus effectively on the global features of the superimposed spectrum, firstly, a Transformer shortcut (SC)-ST module is introduced into the SST-Backbone network. This module focuses on the global features of the superimposed spectrum. Simultaneously, the convolutional neural network's ability to focus on local superimposed spectrum features enhances the breadth and depth of the receptive field of the residual block. By fusing local and global features, the model ensures it can acquire complete and effective information of the broadband UHF superimposed spectrum. Preliminary UHF superimposed spectrum feature fusion is performed in the SPPF module. Secondly, in the SST-Head detection head, the C3 module is again fused with the global and local features extracted from the backbone network through multi-directional and multi-scale feature interaction. This further detects and filters out deeper global and local feature details belonging to the superimposed spectrum of the target UHF satellite signal. Next, the Detector performs convolutional operations to compress the number of channels and scale the feature map size. Finally, a multi-scale target detection anchor grid strategy is executed on the generated superimposed spectrum feature map to perform target signal detection. Finally, the predicted bounding boxes obtained by the Sigmoid activation function provide comprehensive information on the category and location of the superimposed spectral image, thus successfully achieving accurate detection of the target signal spectrum.

[0059] 1.2.1 SST-Backbone Network

[0060] The SST-Backbone backbone network is designed to enhance network depth and receptive field. Its core consists of convolutional layers and SC-ST modules, forming a network for extracting superimposed spectral features. By combining five Conv convolutional layers with a stride of two with four SC-ST layers, the network comprehensively and deeply captures the global features of the superimposed spectrum. The introduction of the SC-ST modules incorporates the Transformer concept, fully utilizing its advantage in global receptive field and correcting the potential for multi-layer convolutional modules to overlook global features during superimposed spectral feature extraction. To mitigate the computational burden and complexity of the introduced Transformer, the Swin-Transformer algorithm is introduced, limiting the attention mechanism's computational region to each fixed window. This effectively overcomes the high computational complexity and difficulty in achieving high-speed real-time detection caused by global self-attention computation in traditional Transformer applications on ultra-shortwave superimposed spectrum. Finally, the features extracted by the SC-ST modules are further fused through the SPPF module, achieving preliminary information synergy between global and local features, providing a more balanced and comprehensive foundation for the network's overall performance. This not only improves the network's ability to accurately abstract superimposed spectra, but also makes target detection more accurate and robust at different scales and in different contexts.

[0061] 1.2.2 SC-ST Module

[0062] The SC-ST module is a complex structure consisting of a single-layer Conv, MLP, WTA, a two-layer LN, and two residual structures. Multiple cascaded units are designed to fully leverage their role in superimposed spectral feature extraction. The module first performs convolution operations through the Conv module, fusing superimposed spectral features layer by layer. LayerNorm normalization is then applied to the feature maps to enhance model stability and convergence. Next, in the WTA, to achieve precise information exchange at both local and global scales, the window attention (W-MSA) and shifted window attention (SW-MSA) modules are used to alternately process the ultra-shortwave superimposed spectral feature maps, dividing them into non-overlapping windows. Feature fusion is achieved using the global nature of the Transformer and the local nature of the CNN, followed by an initial residual connection with the features extracted by the WTA module to optimize the model's capture of superimposed spectral features. The residually connected features are then normalized again before being input into the MLP module. Utilizing the fully connected layers of the MLP module, the model can extract global and local features of the superimposed spectrum from multiple scales. Finally, the superimposed spectral features processed by residual connections are fused with the superimposed spectral features extracted by the MLP module to improve the abstraction ability of superimposed spectral features. In summary, the SC-ST module, through clever combination, can both preserve local information and fully explore global correlations during feature extraction, thereby effectively enhancing the network's ability to extract and express superimposed spectral features.

[0063] 1.2.3 WTA Module

[0064] The structure of the WTA module is as follows: Figure 8 As shown, its core lies in controlling the alternating adjustment of the shift_size parameter (0 and 1) to flexibly select the attention calculation mechanism, achieving orderly switching between the W-MSA and SW-MSA modules, and effectively realizing the transfer and connection of feature information between different windows. In this process, the WTA module, through the alternating adjustment of the shift_size parameter, allows the model to selectively focus on global and local features. By switching the attention mechanism in a timely manner, different windows can be modeled more comprehensively and accurately during information interaction, thereby enhancing the detection model's ability to interact between global and local features of the superimposed spectrum. This ingenious design not only improves the network's flexibility but also enables the model to better adapt to feature changes at different scales and in different contexts, further optimizing the extraction and utilization of superimposed spectral features.

[0065] 1.2.4 W-MSA and SW-MSA modules

[0066] W-MSA adopts the Swin-Transformer framework and proposes a local attention mechanism. Its main steps include dividing the feature map into smaller, independent windows, followed by independent self-attention operations within each small window. This approach, utilizing local receptive fields, effectively extracts features from each window, thereby increasing the network's acquisition range of superimposed spectral features. However, if each SC-ST module uses the same W-MSA, the information interaction of superimposed spectral features will only exist within each window, failing to establish connections between global superimposed spectral features. To address this issue, a shifted window strategy is introduced to establish information interaction between windows. This window self-attention mechanism implemented through shifted windows is called SW-MSA. Specifically, SW-MSA divides the 2m×2m feature map into windows, using a shifted window strategy to move m / 2 patches in both height and width. Compared to conventional window segmentation, the shifted window segmentation method results in more significant changes to the feature map, allowing for broader and more flexible interaction of superimposed spectral features between windows. By employing a cyclic-shift strategy to improve the efficiency of SW-MSA, the superimposed spectral feature map is first windowed. The window size is set to a fixed value: the central m×m window forms a single group, while the edge windows form separate groups, resulting in nine groups in total. Then, a cyclic shift operation moves the window in the upper left corner of the feature map to the lower right corner, recombining them into three m×m windows, ultimately resulting in four m×m windows. Within each window, self-attention calculations are performed on different patches using a mask matrix. Finally, a reverse operation restores the window to its original position. To more comprehensively extract global and local features of the superimposed spectrum, W-MSA and SW-MSA operations are performed alternately, increasing the information interaction between superimposed spectral features at different locations in the feature map. This ingenious application of the strategy enables the model to more flexibly capture the complex features of the superimposed spectrum, providing the overall detection model with richer information representation and learning capabilities.

[0067] 1.2.4 SPPF Structure

[0068] SPPF is an improvement upon SPP, which stands for Spatial Pyramid Pooling, a method that maps feature maps of various sizes to a fixed-size feature vector. In YOLOv5s, SPP uses convolutional kernels of 5, 9, and 13 sizes by default for pooling. This mechanism works by passing the feature map output from the convolutional layer through these three different sized kernels for pooling, concatenating the resulting feature vectors, and then processing them through a single convolutional layer. SPPF improves upon SPP by using three 5×5 max pooling operations instead of the original 5×5, 9×9, and 13×13 max pooling. In this improvement, SPPF replaces the large convolutional kernels used in the original SPP module with multiple small-sized kernels. This design allows the data from the convolutional layer to undergo three small-sized pooling operations before being processed by a single convolutional layer, effectively enhancing the expressive power of the feature maps, enabling more efficient feature fusion, and improving running speed while maintaining computational efficiency. The overall structure of SPPF is as follows: Figure 7 As shown in the image.

[0069] 1.3SST-Head Feature Fusion Target Detection Head

[0070] The pyramid network structure comprises a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN). In this method, the SST-Head employs a feature fusion target detection head that cascades C3 and the pyramid network. The FPN integrates feature maps from top to bottom using two Conv modules with a stride of 1 and two upsampling cascaded convolutions, enhancing the semantic information of the bottom-level superimposed spectral feature maps. Conversely, the PAN integrates feature maps from bottom to top, performing convolution operations through two Conv modules and simultaneously downsampling to reduce invalid information in the bottom-level feature maps, emphasizing the positional information of the target signal in the top-level ultra-shortwave superimposed spectrum. Finally, the bottom-to-top and top-to-bottom ultra-shortwave superimposed spectral feature maps are fused and combined with the SST-Backbone network for deeper processing, forming a multi-scale feature pyramid. Finally, the C3 module is cascaded again using the characteristics of its local receptive field, more accurately detecting and filtering high-semantic global features and detailed local features in the feature pyramid, generating a more accurate superimposed spectral feature map. The multi-level fusion process makes the final feature map more representative, helping to capture key information in the superimposed spectrum more accurately. Finally, the UHF superimposed spectrum feature map is input into the Detect layer after reducing the number of channels and scaling through the Conv layer. The Detect module performs convolution operations on the superimposed spectrum feature map to generate bounding box location information for the superimposed spectrum, effectively determining the frequency band location of the target signal. Simultaneously, the Detect module performs another convolution operation to predict the category of the object within each bounding box, i.e., determining whether a target signal exists in the superimposed spectrum. The Detect module uses a confidence score to measure the model's confidence in whether an object is contained within the bounding box; a high confidence score indicates that the model believes a target signal exists in the superimposed spectrum within that bounding box. To improve detection accuracy, the Detect module applies Non-Maximum Suppression (NMS). The purpose of NMS is to eliminate multiple overlapping bounding boxes in the superimposed spectrum, ensuring that each target signal is detected only once, thereby reducing the false alarm rate. NMS filters bounding boxes based on confidence scores, removing redundant superimposed spectral boxes with high overlap to produce the final detection result.

[0071] Through the above-mentioned layer-by-layer cascaded design and operation, SST-Head can generate more accurate prediction boxes with more accurate location information, prediction categories, and confidence levels, providing a reliable foundation for the effective detection of superimposed spectral signals.

[0072] Example:

[0073] 2. Experiment and Results Analysis

[0074] Due to the lack of publicly available target satellite UHF spectrum detection image datasets, this patent study constructed an image dataset using satellite signals acquired by a laboratory antenna to evaluate the performance of the proposed algorithm. This dataset, generated through image processing of satellite signals, provides a practical basis for the experimentation and verification of the algorithm. In the complex electromagnetic environment, broadband UHF signal spectra exhibit a wide variety of characteristics. This provides our research with more challenging and realistic data scenarios.

[0075] 2.1 Experimental Dataset

[0076] Since there are currently no publicly available datasets for VHF channel classification, the dataset used in this experiment was acquired by a laboratory VHF dual-element helical antenna. The VHF superimposed spectrum dataset contains 5000 images under different conditions, including normal signals, abnormal fluctuations, abnormal bandwidth, and multi-continuous signals. The first 4000 images from the dataset were used as the training set, and the last 1000 images were used as the test set for model training and testing. The dataset format is shown in Table 1.

[0077] Table 1 Training and test signal sample data

[0078]

[0079] 2.2 Experimental Environment and Model Evaluation Indicators

[0080] The test simulation server was equipped with an NVIDIA TITAN RTX 3090 GPU and used PyTorch as the deep learning model framework. The loss function was optimized using a stochastic gradient descent (SGD) optimizer, and the input image was a 3-channel ultra-shortwave superimposed spectrum image with a resolution of 640×640. The training batch size was 32, and a total of 100 training epochs were performed. During the analysis and testing process, five metrics were used to comprehensively evaluate the model performance: precision, false alarm rate, recall, mean average precision (mAP), and detection rate. Furthermore, the ultra-shortwave superimposed spectrum features extracted by the model were visualized and analyzed. By observing the feature maps, the model's focus in the superimposed spectrum image can be intuitively understood, as well as whether the model has successfully learned the important features of the ultra-shortwave superimposed spectrum region. The degree to which different models pay attention to interference and anomalous signals is also compared, thus providing a comprehensive understanding of the model's performance. In practical applications, the size of the model needs to be weighed based on task requirements, hardware resources, and performance requirements. Therefore, the size of each network model is evaluated and compared while meeting performance requirements.

[0081] 2.3 Detection performance

[0082] In the broadband UHF superimposed spectrum normal image test set, which includes conventional UHF signal images and UHF signals with continuous spectrum, the performance of the SST-YOLOV5s UHF superimposed spectrum detection algorithm, YOLOV5s target detection algorithm and SSD target detection algorithm proposed in the test patent are compared. At the same time, the false alarm effect of abnormal signals is tested in the superimposed spectrum test set of abnormal signals using the model proposed in this paper.

[0083] 2.3.1 Feature Visualization

[0084] The datasets of conventional and continuous UHF signal images were analyzed and tested using SSD, YOLOv5s, and the algorithm presented in this paper. Visualized heatmaps were generated to illustrate the feature regions extracted by the model. In the heatmaps, the red areas represent the UHF superimposed spectral features learned by the model, i.e., the regions of interest to the model during UHF superimposed spectral image detection. Figure 9 In contrast, for conventional signal images, the SSD algorithm does not fully focus on the channel location and its area of ​​interest is scattered, only concentrating on some regions at the edges of the UHF superimposed spectrum. The YOLOv5s algorithm focuses on the central peak region of the superimposed spectrum while ignoring some edge regions. In comparison, the algorithm proposed in this paper can more comprehensively focus on the features of both the central peak region and the edge region of the UHF superimposed spectrum. In UHF continuous signal images, the SSD algorithm basically does not focus on the target superimposed spectrum region, instead activating a large number of blank areas. The YOLOv5s algorithm shows more attention to some peaks of the UHF superimposed spectrum, but does not fully activate the edge regions of the superimposed spectrum. In comparison, the algorithm proposed in this chapter more comprehensively activates both peak and edge regions, while activating only a small number of non-UHF superimposed spectrum features, and significantly activating features of the UHF superimposed spectrum region.

[0085] Comparative analysis of the heatmap reveals that, compared to SSD and YOLOv5s, the proposed ultra-shortwave superimposed spectrum detection algorithm in this chapter more accurately focuses on the ultra-shortwave superimposed spectrum portion of the image on a conventional signal test set, without focusing on the background. Furthermore, for continuous signal test sets, it can utilize global and local features to perform correlation fusion to extract target information, allowing the model to focus on all ultra-shortwave superimposed spectrum regions.

[0086] This patent also employs the SST-YOLOV5s method to detect anomalous and interference signals in broadband UHF, focusing on the types of anomalous signals and achieving the following results: Figure 10 As shown.

[0087] As can be clearly observed from the heatmap, the proposed method does not significantly focus on anomalous interference, activating only a small number of anomalous regions with a low activation level. This effectively reduces the detection rate of anomalous and interfering signals, and significantly reduces the false alarm rate for broadband UHF signals. These visualization results provide a clearer comparison of the algorithm proposed in this chapter with SSD and YOLOv5s in terms of focus points in UHF signal superimposed spectral images, thus enabling a more comprehensive evaluation of the superiority of the proposed method.

[0088] To evaluate the SST-YOLOV5s algorithm's ability to detect broadband UHF superimposed spectra in complex electromagnetic environments, different network models were trained for 100 rounds on a UHF superimposed spectrum dataset. Then, UHF signal target detection was performed on the YOLOV5s, SSD, and SST-YOLOV5s algorithms proposed in this chapter, and the detection rates were statistically analyzed. The performance of the above detection algorithms was evaluated on the entire signal test set (including interference and anomalous signals) and the normal signal test set. The recognition rate is the percentage of detected normal signals out of all signals, and the false alarm rate is the percentage of falsely identified interference and anomalous signals out of the total number of signals. The detection performance on the entire signal test set is shown in Table 2.

[0089] Table 2 shows the detection performance on the complete signal test set.

[0090] algorithm Detection rate recognition / % False alarm rate / % Traditional spectral energy detection 72.6 18.3 SSD 76.1 14.7 YOLOV5s 83.3 8.9 This article's method 85.4 7.4

[0091] The false alarm rate of this patented method is only 7.4% across the entire signal test set, which is 10.8% lower than traditional methods, 7.4% lower than the deep learning method SSD, and 1.5% lower than YOLOv5s.

[0092] The detection performance of the three algorithms on the normal signal test set is shown in Table 3:

[0093] Table 3 Detection performance on the normal signal test set

[0094] algorithm Precision rate P / % Recall rate R / % mAP / % SSD 84.2 84.0 88.4 YOLOV5s 88.0 88.6 92.9 This article's method 92.7 90.8 93.8

[0095] Statistical comparison of detection results on the superimposed spectrum normal signal test set reveals that, compared with the traditional SSD method, the method proposed in this chapter improves precision, recall, and mAP by 8.5, 6.8, and 5.4 percentage points, respectively. Compared with YOLOv5s, the patented method also achieves varying degrees of improvement in all three metrics, increasing by 4.7, 2.2, and 0.9 percentage points, respectively. These results demonstrate that, compared with the traditional SSD and YOLOv5s algorithms, the method proposed in this patent achieves higher detection precision and recall, as well as a superior mAP value, on the normal signal superimposed spectrum test set. This further validates the superior performance of this algorithm in processing superimposed spectrum signal detection tasks.

[0096] Table 4 Comparison of memory usage for each model

[0097] algorithm Model memory usage / MB SSD 92.8 YOLOV5s 14.0 This article's method 14.7

[0098] Compared to the SSD algorithm, as shown in Table 4, the method proposed in the patent has a model memory footprint of only 15.8% of that of SSD, demonstrating a significant advantage in model lightweighting. Similarly, compared to YOLOv5s, the model proposed in this patent successfully achieves model lightweighting in high-performance detection tasks, with only a slight increase in memory footprint of 0.7M, while improving accuracy. This control over model size provides support for deployment in resource-constrained environments.

[0099] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. Based on superimposed spectrum and depth Spend The method for detecting ultra-shortwave broadband satellite signals learned is characterized by, Includes the following steps: Step 1: Acquire target satellite signal data using an ultra-shortwave antenna, and then process the acquired ultra-shortwave signal using a receiver; the receiver filters the ultra-shortwave frequency band of the acquired ultra-shortwave signal to obtain broadband ultra-shortwave signal data centered on the target frequency; Step 2: After processing the broadband UHF signal data by the receiver, the broadband spectrum of the UHF signal is obtained by fast Fourier transform. Then, the superimposed spectrum of the broadband spectrum of the UHF signal is constructed by superimposing the spectrum. By continuously iterating and accumulating the small differences between the signal and the noise floor, the distinction between the signal and the noise floor is highlighted. The superimposed spectrum is then normalized and preprocessed. Step 3: Construct an ultra-shortwave broadband signal detection model based on SST-YOLOV5s to improve the detection accuracy of broadband ultra-shortwave satellite signals by the deep learning network model and reduce the false alarm rate; The construction of the SST-YOLOV5s ultra-shortwave broadband signal detection model is specifically implemented as follows: The SST-Backbone network is constructed by cascading four layers of SC-ST and multi-layer convolutional structures. In the SST-Backbone network, the global receptive field of the Transformer is utilized to enable the modules to focus on the global features of the superimposed spectrum. At the same time, multi-layer convolutional modules with shortcut(SC)-ST residuals are embedded in the Transformer, enabling the shortcut(SC)-ST modules to focus on the global features of the superimposed spectrum. The combination of multi-layer convolutional modules with residuals increases the depth and receptive field of the network, allowing the model to fully integrate the global and local features of the extracted ultra-shortwave superimposed spectrum, and the SPPF module performs the initial information interaction. Step 4: Input the normalized preprocessed superimposed spectra of different patterns existing in the broadband of the target satellite band into the SST-YOLOV5s broadband ultra-shortwave satellite detection model for simulation testing of detection effect, and perform statistical comparison calculation on the detection results.

2. The superposition spectrum and deep learning based method of claim 1. Spend The method for detecting ultra-short wave broadband satellite signals based on learning, characterized in that, The ultra-shortwave antenna described in step one adopts a high-gain circular polarization method with a gain of 15 dBi.

3. The method for detecting ultra-shortwave broadband satellite signals based on superimposed spectrum and deep learning according to claim 1, characterized in that, Step two specifically includes: 2-1. With a center frequency of 255MHz, the target band's ultra-shortwave signal is sampled at a sampling frequency of 40.96MHz. The resulting target signal has a starting frequency of 234.52MHz and an ending frequency of 275.48MHz. 2-2. First, the target signal data with a sampling frequency of 40.96MHz is processed by locally plotting the spectrum of the ultra-shortwave broadband target signal to increase the resolution of the image; 2-3. The spectrum was simulated and plotted using 5, 10, 20 and 30 superpositions to obtain superimposed spectra with different superposition numbers. By observing the signal characteristics and required data volume of the superimposed spectra with different superposition numbers, the superimposed spectrum was finally obtained by simulating and plotting the spectrum using 30 superpositions.

4. The method for detecting ultra-shortwave broadband satellite signals based on superimposed spectrum and deep learning according to claim 3, characterized in that, Step 2-2 specifically includes: drawing FFT spectrum diagrams every 100kHz, totaling 409 spectrum diagrams, including the signal spectrum diagrams of the entire target frequency band; in order to obtain the signals present in the target signal, marking and detecting the signals present in each spectrum diagram, and performing marking processing on the signals and interference; in order to ensure the consistency and comparability of the data, the image size of the local 100kHz spectrum diagrams is uniformly normalized to a three-channel 640×640 size.

5. The method for detecting ultra-shortwave broadband satellite signals based on superimposed spectrum and deep learning according to claim 3, characterized in that, The superimposed spectrum types include four categories: normal signals, abnormal fluctuation signals, abnormal bandwidth signals, and multi-continuous signals in the broadband ultra-shortwave band.

6. The method for detecting ultra-shortwave broadband satellite signals based on superimposed spectrum and deep learning according to claim 1, characterized in that, The SST-Head target detection head of the SST-YOLOV5s ultra-shortwave broadband signal detection model uses a feature pyramid obtained by multi-scale fusion of global and local features extracted by the SST-Backbone network. Combined with the local receptive field characteristics of the C3 module without residuals in the SST-Head target detection head, it detects and filters out global features with high semantic level and local features representing details belonging to the superimposed spectrum. After fusing the global features with high semantic level and local features representing details, a more accurate superimposed spectrum feature map is generated. After the superimposed spectral feature map is reduced in dimensionality by the number of channels and scaled by the Conv layer, the Detect detector performs target detection on superimposed spectral feature maps at different scales based on grid anchor points. The sigmoid activation function generates more accurate and comprehensive superimposed spectral image prediction box category and location information, ultimately achieving accurate detection of the target signal spectrum.

7. The method for detecting ultra-shortwave broadband satellite signals based on superimposed spectrum and deep learning according to claim 6, characterized in that... The construction of the SC-ST module in the SST-Backbone backbone network is specifically implemented as follows: Composed of a single-layer Conv, MLP, WTA, two-layer LN, and two residual structures, the cascaded multiple units aim to fully leverage their role in superimposed spectral feature extraction; the SC-ST module uses convolution operations and normalization to fuse and stabilize features; feature extraction is optimized through window attention mechanism and residual connections; features are extracted using fully connected layers and multi-scale methods; and finally, feature fusion enhances the ability to abstract superimposed spectral features.

8. The method for detecting ultra-shortwave broadband satellite signals based on superimposed spectrum and deep learning according to claim 6, characterized in that... The construction of the SST-Head target detection header is specifically implemented as follows: The SST-Head object detection head consists of a Feature Pyramid Network (FPN) and a Path Aggregation Network (PAN). The FPN integrates feature maps from top to bottom, enhancing the semantic information of the superimposed spectral features at the bottom and top levels. The PAN integrates feature maps from bottom to top, emphasizing the positional information of the target signal at the top level. Finally, the two are fused to form a multi-scale feature pyramid, and the accuracy of the superimposed spectral feature maps is improved through cascading. Finally, the Detect module performs convolution operations to generate the bounding box position information of the superimposed spectral location.

Images