Broadband communication signal detection and identification method based on deep learning

Abstract

The invention provides a broadband communication signal detection and identification method based on deep learning, and relates to the technical field of wireless communication and artificial intelligence crossing, and the method comprises the steps: obtaining in-phase orthogonal data, and carrying out the preprocessing; a DSA-RT-DETR model is constructed through an adaptive time-frequency feature fusion module, a dynamic spectrum attention module and a scale difference module; the DSA-RT-DETR model is trained and tested through the in-phase orthogonal data, and intermediate frequency data is obtained; designing a target band-pass filter according to the intermediate frequency data, and filtering the in-phase orthogonal data to obtain a pure signal; training an OfficientNetB4 model through the pure signal; and obtaining in-phase orthogonal data to be classified, and inputting the in-phase orthogonal data to be classified into the trained OfficientNetB4 model to obtain a classification identification result. According to the technical scheme, end-to-end intelligent processing from broadband signal detection to modulation identification is realized.

Description

Technical Field

[0001] This application relates to the field of wireless communication and artificial intelligence, and in particular to a method for detecting and recognizing broadband communication signals based on deep learning. Background Technology

[0002] With the rapid development of 5G / 6G communications, the Internet of Things (IoT), and smart cities, the demand for efficient and intelligent spectrum monitoring and management is growing dramatically. In the civilian market, this technology can provide telecom operators and equipment manufacturers with dynamic spectrum access solutions, optimizing spectrum resource utilization. In defense, public safety, and other fields, its excellent real-time signal detection and identification capabilities can meet the high reliability requirements of complex electromagnetic environments, making it suitable for critical tasks such as signal intelligence and electronic warfare. Furthermore, in industries such as civil aviation, railways, and intelligent transportation, this technology also provides an effective tool for spectrum regulation and interference investigation. The overall market continues to demand high-precision, low-complexity spectrum sensing technologies.

[0003] However, traditional signal detection and recognition methods suffer from low sensitivity and high false alarm rate in low signal-to-noise ratio, multi-signal coexistence, and dynamic spectrum scenarios, making it difficult to achieve high-precision real-time perception. Secondly, existing deep learning models struggle to balance detection speed and accuracy, failing to meet the requirements of low computational complexity and high recognition accuracy. Furthermore, there is a lack of end-to-end recognition systems that can uniformly process signals of multiple modulation types and scales and possess strong generalization capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide a broadband communication signal detection and recognition method based on deep learning, in order to solve the problems of poor real-time performance, low accuracy, and weak adaptability in signal detection and recognition in complex electromagnetic environments.

[0005] The above-mentioned objective of this application is achieved through the following technical solution: S1: Acquire in-phase orthogonal data and perform preprocessing; S2: Construct the DSA-RT-DETR model through an adaptive time-frequency feature fusion module, a dynamic spectrum attention module, and a scale difference module; S3: Train and test the DSA-RT-DETR model using in-phase orthogonal data to obtain intermediate frequency data; S4: Based on the intermediate frequency data, design a target bandpass filter to filter the in-phase orthogonal data and obtain a clean signal; train the EfficientNetB4 model using the clean signal. S5: Obtain the in-phase orthogonal data to be classified and input it into the trained EfficientNetB4 model to obtain the classification and recognition results.

[0006] Optionally, step S1 includes: The PlutoSDR device was used to collect in-phase and quadrature data of 53 broadband signal modulations, and the data preprocessing was performed to obtain the preliminary time-frequency distribution results of the broadband signals.

[0007] Optionally, step S3 includes: S31: The backbone network of the DSA-RT-DETR model adopts HGNetV2; an adaptive time-frequency feature fusion module is integrated in the S5 stage of the backbone network HGNetV2; through the adaptive time-frequency feature fusion module, the time domain and frequency domain deformation shifts are adaptively learned from the input signal to obtain shallow high-resolution local features, mid-level semantic features and deep abstract global features; S32: The integrated scale difference module includes a multi-scale convolution module and a dilated convolution module; the multi-scale convolution module is used to enhance the detection performance of the adaptive time-frequency feature fusion module for signals of different shapes and sizes by fusing multi-scale features with local details; the adaptive time-frequency feature fusion module is used to expand the receptive field of the model and improve feature resolution and local localization accuracy; The scale difference module is used to perform multi-scale fusion of shallow high-resolution local features, mid-level semantic features and deep abstract global features to obtain mid-frequency data. 7×7, 5×5, and 3×3 convolutional kernels were used to process shallow high-resolution local features, mid-level semantic features, and deep abstract global features, respectively. Dynamic void ratios of 3x, 5x, and 7x are applied to shallow high-resolution local features, mid-level semantic features, and deep abstract global features, respectively.

[0008] Optionally, step S31 includes: Employing a 7×7 deformable convolution kernel, the temporal offset is learned. With frequency domain offset Dynamically adjust the receptive field of the convolution, specifically including:

[0009]

[0010] in, and Represents the coordinates of the initial sampling points on the convolution kernel; Indicates the step size or magnitude of the offset; and Indicates the coordinates after the offset; Through iterative offset strategy and This allows the convolution kernel to fit the time-frequency shape of the distorted signal; at the same time, it combines 1×1 convolution and 3×3 convolution to fuse global time-frequency trends and local subtle features; the global time-frequency trend is the broadband signal profile; the local subtle features are instantaneous amplitude spikes. The backbone network HGNetV2 outputs feature maps for the last three stages. Feature map and feature map These correspond to shallow high-resolution local features, mid-level semantic features, and deep abstract global features, respectively.

[0011] Optionally, the intermediate frequency data includes: carrier frequency, signal bandwidth, and power level.

[0012] Optionally, step S5 includes: The EfficientNetB4 model includes: a scaling module and a squeezing excitation module; The scaling module is used to adaptively adjust the feature scale; the squeezing excitation module is used to enhance the expression of key features. Remove the top classification layer of the EfficientNetB4 model and add global average pooling; By adding three cascaded fully connected layers to the EfficientNetB4 model, nonlinear features are introduced; The output layer of the EfficientNetB4 model has 53 neurons, and the Softmax activation function is used to output the classification probability.

[0013] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform a broadband communication signal detection and recognition method based on deep learning.

[0014] A computer-readable storage medium storing instructions that, when executed, perform a deep learning-based broadband communication signal detection and recognition method.

[0015] The beneficial effects of the technical solution provided in this application are: By deeply integrating DSA-RT-DETR with the EfficientNetB4 model, a two-stage processing architecture is constructed: the first stage achieves real-time detection of spectrum occupancy regions through the DSA mechanism and spatiotemporal cross-attention, significantly improving the sensitivity of broadband signal detection; the second stage utilizes the EfficientNetB4 model to complete high-precision classification of signal modulation types. This two-stage processing architecture, integrating DSA-RT-DETR and EfficientNetB4 models, enables efficient collaborative processing from spectrum detection to modulation identification. Attached Figure Description

[0016] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of an embodiment of this application; Figure 2 This is a diagram showing the ASK-2 simulation signal detection results in the embodiments of this application; Figure 3 This is a diagram showing the PSK-2 simulation signal detection results in the embodiments of this application; Figure 4 This is a diagram showing the QAM-16 simulation signal detection results in the embodiments of this application; Figure 5 This is a diagram of the first signal detection result in an embodiment of this application; Figure 6 This is a diagram showing the second signal detection result in an embodiment of this application; Figure 7 This is a diagram showing the confusion matrix results of 53 types of modulation signal identification and classification in the embodiments of this application; Figure 8 This is a schematic diagram of the electronic device structure in the embodiments of this application. Detailed Implementation

[0017] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0018] The embodiments of this application provide a method for detecting and recognizing broadband communication signals based on deep learning.

[0019] Please refer to Figure 1 , Figure 1 This is a flowchart of a broadband communication signal detection and recognition method based on deep learning, as described in an embodiment of this application, including: S1: Acquire in-phase orthogonal data and perform preprocessing; S2: Construct the DSA-RT-DETR model through an adaptive time-frequency feature fusion module, a dynamic spectrum attention module, and a scale difference module; S3: Train and test the DSA-RT-DETR model using in-phase orthogonal data to obtain intermediate frequency data; S4: Based on the intermediate frequency data, design a target bandpass filter to filter the in-phase orthogonal data and obtain a clean signal; train the EfficientNetB4 model using the clean signal. S5: Obtain the in-phase orthogonal data to be classified and input it into the trained EfficientNetB4 model to obtain the classification and recognition results.

[0020] As one example, in the classification stage, EfficientNetB4 is used as the backbone network and trained on a noise-enhanced dataset to achieve high-precision recognition of 53 modulation types with a classification accuracy of 96.33%.

[0021] This application addresses three core challenges in broadband signal detection and recognition under complex electromagnetic environments by employing the aforementioned technical solutions: First, traditional methods suffer from low sensitivity and high false alarm rates in low signal-to-noise ratio, multi-signal coexistence, and dynamic spectrum scenarios, making high-precision real-time sensing difficult to achieve. Second, existing deep learning models struggle to balance detection speed and accuracy, failing to meet the requirements of low computational complexity and high recognition accuracy. Furthermore, there is a lack of end-to-end recognition systems capable of uniformly processing multi-modulation types and multi-scale signals with strong generalization capabilities. This invention deeply integrates DSA-RT-DETR with the EfficientNetB4 model, constructing a two-stage processing architecture: the first stage utilizes the DSA mechanism and spatiotemporal cross-attention to achieve real-time detection of spectrum occupancy regions, significantly improving broadband signal detection sensitivity; the second stage leverages the EfficientNetB4 model to complete high-precision classification of signal modulation types. This invention demonstrates methodological innovation and effectiveness, meeting the needs of scientific research and providing valuable reference for real-time detection of broadband communication signals under complex electromagnetic environments.

[0022] Step S1 includes: The PlutoSDR device was used to collect in-phase and quadrature data of 53 broadband signal modulations, and the data preprocessing was performed to obtain the preliminary time-frequency distribution results of the broadband signals.

[0023] As one embodiment, this invention uses PlutoSDR to collect in-phase quadrature (IQ) data of 53 broadband signal modulations and achieves automated annotation based on traditional communication algorithms to construct a high-quality, multi-type signal dataset. It innovatively integrates an Adaptive Time-Frequency Feature Fusion Module (ATFFM) into the backbone network, which can adaptively learn time-domain and frequency-domain distortion shifts from input features, effectively focusing on diverse signal features to improve feature representation capabilities. Furthermore, it introduces a scale-wise differential module (SDM) composed of Multi-Scale Convolution Modules (MSCM) and Atrous Convolution Modules (ACM). The module (ACM) expands the receptive field of the model and enhances feature resolution and local localization accuracy. The fusion of multi-scale and local details further enhances the detection performance of ATFFM for signals of different shapes and sizes. In addition, filters are designed based on intermediate frequency data obtained from spectrum detection, and a synthetic dataset containing 53 modulation types is constructed. After adjusting the signal-to-noise ratio to simulate real channel distortion, a signal classification accuracy of up to 96.33% is achieved based on the EfficientNetB4 model.

[0024] Step S3 includes: S31: The backbone network of the DSA-RT-DETR model adopts HGNetV2; an adaptive time-frequency feature fusion module is integrated in the S5 stage of the backbone network HGNetV2; through the adaptive time-frequency feature fusion module, the time domain and frequency domain deformation shifts are adaptively learned from the input signal to obtain shallow high-resolution local features, mid-level semantic features and deep abstract global features; In one specific embodiment of this application, these IQ samples can be easily extracted from broadband records using the time-frequency localization coordinates predicted by the target detector. To adapt to the time-frequency deformation of the signal, the DSA-RT-DETR model integrates an Adaptive Time-Frequency Feature Fusion (ATFFM) module in the S5 stage of the backbone network (HGNetV2). This module adaptively learns the time-domain and frequency-domain deformation shifts from the input signal, focusing on diverse signal features.

[0025] S32: The integrated scale difference module includes a multi-scale convolution module and a dilated convolution module; the multi-scale convolution module is used to enhance the detection performance of the adaptive time-frequency feature fusion module for signals of different shapes and sizes by fusing multi-scale features with local details; the adaptive time-frequency feature fusion module is used to expand the receptive field of the model and improve feature resolution and local localization accuracy; The scale difference module is used to perform multi-scale fusion of shallow high-resolution local features, mid-level semantic features and deep abstract global features to obtain mid-frequency data. 7×7, 5×5, and 3×3 convolutional kernels were used to process shallow high-resolution local features, mid-level semantic features, and deep abstract global features, respectively. Dynamic hole rates of 3x, 5x, and 7x are applied to shallow high-resolution local features, mid-level semantic features, and deep abstract global features, respectively. In one specific embodiment of this application, 7×7, 5×5, and 3×3 convolutional kernels are used respectively to process the signals, adapting to the scale difference from 0.0036MHz (narrowband weak signal) to 0.252MHz (wideband strong signal). Dilated convolution (ACM) applies a dynamic dilation rate of 3 / 5 / 7 to the multi-scale features, expanding the receptive field without increasing computation (e.g., a 7x dilation rate can cover the frequency span of a wideband signal) while preserving feature resolution, thus solving the problem of "receptive field and resolution being mutually exclusive" in traditional convolution. A scale difference module optimizes the fusion block within the CCFF. This fusion block aims to expand the receptive domain of adjacent features and fuse them into new features. The Transformer encoder focuses on key time-frequency regions through a Dynamic Spectrum Attention (DSA) module, initially enhancing signal features.

[0026] In one specific embodiment of this application, an integrated scale-wise differential module (SDM) is constructed, which consists of a multi-scale convolution module (MSCM) and an atrous convolution module (ACM). The ACM expands the receptive field of the model and improves feature resolution and local localization accuracy, while the MSCM further enhances the detection performance of ATFFM for signals of different shapes and sizes by fusing multi-scale features and local details.

[0027] Step S31 includes: Employing a 7×7 deformable convolution kernel, the temporal offset is learned. With frequency domain offset Dynamically adjust the receptive field of the convolution, specifically including:

[0028]

[0029] in, and Represents the coordinates of the initial sampling points on the convolution kernel; Indicates the step size or magnitude of the offset; and Indicates the coordinates after the offset; Through iterative offset strategy and This allows the convolution kernel to fit the time-frequency shape of the distorted signal; at the same time, it combines 1×1 convolution and 3×3 convolution to fuse global time-frequency trends and local subtle features; the global time-frequency trend is the broadband signal profile; the local subtle features are instantaneous amplitude spikes. The backbone network HGNetV2 outputs feature maps for the last three stages. Feature map and feature map These correspond to shallow high-resolution local features, mid-level semantic features, and deep abstract global features, respectively.

[0030] The intermediate frequency data includes: carrier frequency, signal bandwidth, and power level.

[0031] Step S5 includes: The EfficientNetB4 model includes: a scaling module and a squeezing excitation module; The scaling module is used to adaptively adjust the feature scale; the squeezing excitation module is used to enhance the expression of key features. Remove the top classification layer of the EfficientNetB4 model and add global average pooling; By adding three cascaded fully connected layers to the EfficientNetB4 model, nonlinear features are introduced; The output layer of the EfficientNetB4 model has 53 neurons, and the Softmax activation function is used to output the classification probability.

[0032] As one embodiment, based on the intermediate frequency (IF) parameters (center frequency, bandwidth) output during the detection phase, a target bandpass filter is designed to extract clean signals from the original IQ data. The extracted signals are divided into training and validation sets in an 8:2 ratio. EfficientNetB4 is used as the classification backbone network, and the feature scale is adaptively adjusted through a "scale module" and the key feature expression is enhanced through a "squeeze excitation (SE) module". The network structure is specifically optimized by removing the top classification layer and adding global average pooling (GAP) to compress the spatial dimension and reduce redundant information. Three fully connected layers (1024 / 1024 / 512 neurons, ReLU activation) are connected in series to introduce nonlinear features. The output layer has 53 neurons (corresponding to 53 modulation types) and uses the Softmax activation function to output the classification probability.

[0033] In one embodiment, addressing the issues of poor real-time performance, low accuracy, and weak adaptability in signal detection and recognition in complex electromagnetic environments, a two-stage processing architecture integrating DSA-RT-DETR and the EfficientNetB4 model achieves efficient collaborative processing from spectrum detection to modulation recognition. This system achieves an mAP of 85.15% and a classification accuracy of 96.33% in dynamic scenes, significantly outperforming existing mainstream solutions, and possesses advantages such as fewer parameters, high computational efficiency, and strong real-time performance. This invention can be widely applied in cognitive radio, military communications, public safety, and other fields, effectively improving the perception and recognition capabilities in complex spectrum environments, and has significant practical value and broad industrialization prospects.

[0034] In one embodiment, the present invention presents a real-time detection Transformer framework (DSA-RT-DETR) based on a dynamic spectrum attention mechanism for spectrum detection and identification in cognitive radio networks. This invention innovatively integrates DSA-RT-DETR with the EfficientNetB4 model, constructing a two-stage processing architecture: the first stage achieves real-time detection of spectrum occupancy regions through the DSA mechanism and spatiotemporal cross-attention, significantly improving the sensitivity of broadband signal detection; the second stage utilizes the EfficientNetB4 model to complete high-precision classification of signal modulation types. Experiments based on the PlutoSDR RF sensor synthetic dataset show that the framework achieves an average accuracy (mAP) of 86.7% and a modulation classification accuracy of 96.33% in dynamic scenarios, representing mAP improvements of 3.5% and 2.7% respectively compared to RT-DETR and YOLOv11m schemes. It provides an efficient real-time monitoring solution for complex RF environments with 22.6M parameters, 55.8 GFLOPs of computation, and a frame rate of 60.4 FPS.

[0035] The signal detection simulation verification of this invention is based on an experimental platform built on the CentOS7 operating system. The hardware configuration uses a server equipped with an NVIDIA GeForce RTX 4060 GPU to meet the computing power requirements for deep learning model training and inference. On the software side, the DSA-RT-DETR model is built using the PyTorch deep learning framework, the SciPy library is called to complete the signal time-frequency analysis (such as STFT transformation), OpenCV is used to process the time-frequency graph data, and the TensorRT tool is used to accelerate the model inference process to ensure the accuracy of real-time detection performance verification.

[0036] The model training process uses the following parameter configuration: 1) Initial learning rate: set to The cosine annealing scheduler is used to dynamically adjust and balance the convergence speed and parameter stability (to avoid loss oscillations caused by an excessively large learning rate or slow convergence caused by an excessively small learning rate). 2) Training rounds: 120 epochs to ensure that the loss function converges fully (experiments show that the model mAP value tends to stabilize after 120 epochs, and there is no significant performance improvement when the number of epochs is increased further). 3) Optimizer: The AdamW optimizer is used, with the weight decay coefficient set to... The momentum parameter was set to 0.9 to suppress model overfitting; 4) Input preprocessing: The frequency domain of the time-frequency plot is normalized to (-0.5MHz, 0.5MHz) to eliminate the influence of frequency scaling on anchor point features; a sliding window with a step size of 0.00262s is used in the time domain to divide the time-frequency domain into 512 bins to ensure the consistency of features of signals of different durations.

[0037] To verify the superiority of the DSA-RT-DETR model, mainstream signal detection methods were selected as comparison schemes, specifically YOLOv11mh and RT-DETR. The modulation schemes used for the simulated signals were Amplitude Shift Keying (ASK), Frequency Shift Keying (PSK), and Quadrature Amplitude Modulation (QAM). In addition, 7dB of additive white Gaussian noise was added to the simulated signals. The detection results of different models are shown below. Figure 2 , 3 As shown in Figure 4.

[0038] In simulated signal detection, the DSA-RT-DETR demonstrated superior performance. For ASK-2 signals, the red bounding box detected by the DSA-RT-DETR provided more complete and accurate signal coverage, delineating the time-frequency characteristic region of the signal more meticulously compared to YOLOv11m and RT-DETR. In PSK-2 signal detection, its detection box more accurately matched the actual signal distribution range, reducing false selection of non-signal areas and omissions of signal areas. For signals with more complex modulation schemes such as QAM-16, the DSA-RT-DETR also exhibited excellent detection capabilities, with a higher degree of fit between the detection box and the signal, more clearly and accurately identifying the time-frequency position of the signal. Its overall detection performance was superior to YOLOv11m and RT-DETR, indicating that the DSA-RT-DETR, by introducing mechanisms such as dynamic spectral attention, effectively improves the detection accuracy and localization accuracy of signals with different modulation types in broadband communication signal detection tasks.

[0039] The model performance is comprehensively evaluated by combining standard metrics in the field of target detection with metrics specific to signal detection. 1) Mean Average Precision (mAP): The average value of AP across all modulation categories, reflecting the model's ability to generalize to multiple signal types; 2) Number of parameters: measures model complexity, measured in millions (M); 3) Floating-point operations (FLOPs): Measures computational overhead, measured in billions (G); 4) Frame rate (FPS): Measures the real-time detection speed, measured in frames per second (inference speed when bs=1).

[0040] Table 1. Comparison of the simulation signal detection performance of the three models.

[0041] As shown in Table 1, which compares the simulation signal detection performance of different models, the DSA-RT-DETR model proposed in this invention performs better than the model in the previous version. The performance index reached 86.7%, which is a significant improvement compared to YOLOv11m's 84% ​​and RT-DETR's 83.2%. At the same time, the number of parameters is only 22.6M and the GFLOPs are 55.8G, which is better in terms of model lightweighting and computational overhead control. The frame rate of 60.4FPS can also meet the real-time detection requirements, and the overall performance is better.

[0042] As one example, the method was validated by collecting IQ data of 53 broadband signal modulations using a PlutoSDR RF sensor, estimating the power spectrum using the Welch method, estimating the noise floor using the minimum statistics method, generating labels by setting thresholds based on the energy detection method, and constructing a SIGMF format dataset. This dataset was then divided into training and validation sets in an 8:2 ratio. After training with the DSA-RT-DETR model, two data segments were selected from 20% of the validation set, and the DSA-RT-DETR model was used for detection testing.

[0043] (1) The data “west-wideband-modrec-ex1-tmpl2-20.04” from the validation set was used. This data contains 8-phase phase shift keying (8PSK), 4-phase phase shift keying (4-PSK), orthogonal frequency division multiplexing (OFDM), and 64-QAM signals. Detection was performed using the DSA-RT-DETR model, and the results were obtained... Figure 5 The signal detection results are shown in the figure and the intermediate frequency (IF) data are shown in Table 2.

[0044] Table 2 West-wideband-modrec-ex1-tmpl2-20.04 Intermediate Frequency Data

[0045] (2) The data “west-wideband-modrec-ex4-tmpl10-20.04” from the validation set was used. This data contains frequency modulation (FM), two-phase phase shift keying (2-PSK), single-sideband amplitude modulation (AM_SSB), and on / off keying (OOK) signals. Detection was performed using the DSA-RT-DETR model, and the results were obtained... Figure 6 The signal detection results are shown in the figure and the intermediate frequency (IF) data are shown in Table 3.

[0046] Table 3 West-wideband-modrec-ex1-tmpl2-20.04 Intermediate Frequency Data

[0047] The detection graph shows the signal modulation type and confidence level (representing the cross-parallel ratio threshold). Tables 2 and 3 show the intermediate frequency (IF) data obtained from these two data segments, including bandwidth, center frequency, and modulation scheme information.

[0048] Targeted bandpass filters were designed using intermediate frequency (IF) data (including key parameters such as center frequency and bandwidth) obtained from the DSA-RT-DETR model to filter out 53 signal classes, which were then synthesized into a standardized dataset. The EfficientNet model was then used to classify the dataset according to modulation scheme. Figure 7 As shown, the confusion matrix for modulation mode identification and classification can be obtained.

[0049] This matrix covers 53 signal types, including 2-phase phase shift keying (2-PSK), 4-phase phase shift keying (4-PSK), 16-QAM (16-QAM), 64-QAM (64-QAM), Orthogonal Frequency Division Multiplexing (OFDM), Gaussian Minimum Shift Keying (GMSK), On / Off Keying (OOK), Amplitude Modulation (AM), and Frequency Modulation (FM). Figure 7 The classification accuracy of this sample is shown to be over 96.33%.

[0050] This application also discloses an electronic device. (See reference...) Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0051] The communication bus 502 is used to enable communication between these components.

[0052] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.

[0053] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0054] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the aforementioned deep learning-based broadband communication signal detection and recognition method.

[0055] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.

[0056] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for detecting and recognizing broadband communication signals based on deep learning, characterized in that, The method includes the following steps: S1: Acquire in-phase orthogonal data and perform preprocessing; S2: Construct the DSA-RT-DETR model through an adaptive time-frequency feature fusion module, a dynamic spectrum attention module, and a scale difference module; S3: Train and test the DSA-RT-DETR model using in-phase orthogonal data to obtain intermediate frequency data; S4: Based on the intermediate frequency data, design a target bandpass filter to filter the in-phase orthogonal data and obtain a clean signal; train the EfficientNetB4 model using the clean signal. S5: Obtain the in-phase orthogonal data to be classified and input it into the trained EfficientNetB4 model to obtain the classification and recognition results.

2. The broadband communication signal detection and recognition method based on deep learning as described in claim 1, characterized in that, Step S1 includes: The PlutoSDR device was used to collect in-phase and quadrature data of 53 broadband signal modulations, and the data preprocessing was performed to obtain the preliminary time-frequency distribution results of the broadband signals.

3. The broadband communication signal detection and recognition method based on deep learning as described in claim 1, characterized in that, Step S3 includes: S31: The backbone network of the DSA-RT-DETR model adopts HGNetV2; an adaptive time-frequency feature fusion module is integrated in the S5 stage of the backbone network HGNetV2; through the adaptive time-frequency feature fusion module, the time domain and frequency domain deformation shifts are adaptively learned from the input signal to obtain shallow high-resolution local features, mid-level semantic features and deep abstract global features; S32: The integrated scale difference module includes a multi-scale convolution module and a dilated convolution module; the multi-scale convolution module is used to enhance the detection performance of the adaptive time-frequency feature fusion module for signals of different shapes and sizes by fusing multi-scale features with local details; the adaptive time-frequency feature fusion module is used to expand the receptive field of the model and improve feature resolution and local localization accuracy; The scale difference module is used to perform multi-scale fusion of shallow high-resolution local features, mid-level semantic features and deep abstract global features to obtain mid-frequency data. 7×7, 5×5, and 3×3 convolutional kernels were used to process shallow high-resolution local features, mid-level semantic features, and deep abstract global features, respectively. Dynamic void ratios of 3x, 5x, and 7x are applied to shallow high-resolution local features, mid-level semantic features, and deep abstract global features, respectively.

4. The broadband communication signal detection and recognition method based on deep learning as described in claim 3, characterized in that, Step S31 includes: Employing a 7×7 deformable convolution kernel, the temporal offset is learned. With frequency domain offset Dynamically adjust the receptive field of the convolution, specifically including: in, and Represents the coordinates of the initial sampling points on the convolution kernel; Indicates the step size or magnitude of the offset; and Indicates the coordinates after the offset; Through iterative offset strategy and This allows the convolution kernel to fit the time-frequency shape of the distorted signal; at the same time, it combines 1×1 convolution and 3×3 convolution to fuse global time-frequency trends and local subtle features; the global time-frequency trend is the broadband signal profile; the local subtle features are instantaneous amplitude spikes. The backbone network HGNetV2 outputs feature maps for the last three stages. Feature map and feature map These correspond to shallow high-resolution local features, mid-level semantic features, and deep abstract global features, respectively.

5. The broadband communication signal detection and recognition method based on deep learning as described in claim 1, characterized in that, The intermediate frequency data includes: carrier frequency, signal bandwidth, and power level.

6. The broadband communication signal detection and recognition method based on deep learning as described in claim 1, characterized in that, Step S5 includes: The EfficientNetB4 model includes: a scaling module and a squeezing excitation module; The scaling module is used to adaptively adjust the feature scale; the squeezing excitation module is used to enhance the expression of key features. Remove the top classification layer of the EfficientNetB4 model and add global average pooling; By adding three cascaded fully connected layers to the EfficientNetB4 model, nonlinear features are introduced; The output layer of the EfficientNetB4 model has 53 neurons, and the Softmax activation function is used to output the classification probability.

7. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the method as described in any one of claims 1-6.

Images