Fast broadband signal detection and identification method and device based on time-frequency diagram

Abstract

The invention provides a rapid broadband signal detection and identification method and device based on a time-frequency graph, and the method comprises the steps: converting a broadband signal into a time-frequency graph, and inputting the time-frequency graph into a YOLO-X network for signal time-frequency positioning after the time-frequency graph is preprocessed; constructing a complete signal instance containing a Welch power spectrum based on a positioning result; outputting signal category and confidence evaluation through two-stage feature analysis; and finally, realizing automatic processing or expert secondary discrimination according to a confidence threshold. According to the method, deep learning and signal processing technologies are fused, and the signal detection speed, the recognition accuracy and the decision reliability in a complex electromagnetic environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of communication signal processing technology, and specifically to a method and apparatus for fast broadband signal detection and identification based on time-frequency graphs. Background Technology

[0002] With the rapid development of communication technology, the electromagnetic spectrum, as a crucial carrier of information transmission, is increasingly valuable. Broadband signal detection technology, as a core component of spectrum monitoring, primarily aims to rapidly identify and estimate the parameters of target signals in complex electromagnetic environments. In the civilian communications field, this technology can effectively implement dynamic allocation and monitoring of spectrum resources, achieve intelligent identification of interference signals and precise location of illegal radio stations, and powerfully maintain order in the electromagnetic space. In professional radio monitoring and public safety fields, broadband signal detection technology can also be used for interference investigation, spectrum resource management, and emergency communication support, providing crucial support for maintaining communication order and security.

[0003] In the field of broadband signal detection, traditional methods are mainly based on fixed-threshold energy detection or matched filtering principles. These methods can achieve good performance when processing narrowband signals, but they exhibit significant limitations when facing complex electromagnetic environments and large-bandwidth signals: on the one hand, as the signal bandwidth increases, the noise accumulation effect leads to a decrease in the effective signal-to-noise ratio and insufficient detection accuracy; on the other hand, fixed-threshold methods have poor adaptability and are difficult to cope with dynamically changing electromagnetic environments, easily leading to high false alarm rates and increased false negative rates, especially under low signal-to-noise ratio conditions, where the algorithm performance degrades more significantly.

[0004] In recent years, deep learning-based signal detection methods have shown great potential by converting signals into time-frequency maps and using object detection networks (such as the YOLO series) for signal localization. However, existing methods still have significant limitations: First, most methods stop at bounding box detection on the time-frequency map, lacking in-depth analysis and feature mining of the signal itself; second, the recognition results usually only give a single category, lacking a unified and quantifiable confidence measurement mechanism, making it difficult to assess the reliability of the results; third, the processing flow is rigid, unable to flexibly handle difficult signals with low confidence, and lacks intelligent decision-making capabilities through human-machine collaboration. Summary of the Invention

[0005] To address the problems in the prior art, this invention provides a fast broadband signal detection and recognition method and apparatus based on time-frequency diagrams, aiming to balance detection speed, recognition accuracy, and decision reliability in broadband signal detection.

[0006] A fast broadband signal detection and identification method based on time-frequency graphs includes the following steps:

[0007] Step 1: Acquire broadband data, perform short-time Fourier transform on the broadband data to generate a broadband signal time-frequency diagram, thereby converting the one-dimensional time-domain signal into a two-dimensional time-frequency domain representation; this is the data foundation for all subsequent processing steps.

[0008] Step 2: The three-dimensional data of the broadband signal time-frequency graph are subjected to outlier removal, smoothing filtering, and normalization to suppress noise, enhance signal features, and form a normalized time-frequency graph suitable for deep neural network processing; this step improves the robustness of subsequent signal detection.

[0009] Step 3: Input the normalized time-frequency map into the trained YOLO-X network model. By regressing the bounding boxes of the energy concentration regions in the time-frequency map, the time-frequency boundary parameters of each detected signal are output. The time-frequency boundary parameters include the start frequency, end frequency, start time, and end time, completing the initial localization of the signal in the time-frequency domain. This step establishes a basic identity for each detected signal, providing core localization information for the subsequent formation of complete signal instances. This step achieves rapid signal discovery and coarse-grained parameter extraction.

[0010] Step 4: Based on the time-frequency boundary parameters, extract the corresponding time segments from the original broadband data, perform digital down-conversion at the detected center frequency to shift it to the baseband, and obtain the separated baseband narrowband signals through bandpass filtering and decimation. Through the above processing, the signal region detected in Step 3 is transformed into a signal instance containing specific signal entities. Then, the Welch power spectrum of each baseband narrowband signal is calculated using the Welch method to obtain a smooth and stable power spectral density curve. At this point, each signal instance contains complete time-frequency boundary parameters, baseband narrowband signal, and Welch power spectrum. This step deepens the analysis from the image level to the signal level, providing richer features for fine identification.

[0011] Step 5: Based on the time-frequency boundary parameters, perform preliminary classification of signal instances and assign initial confidence scores; then extract statistical features, including peak spacing, linewidth, and out-of-band leakage, from the Welch power spectrum curve; based on the matching degree between the statistical features and the signal category classifier, output the final signal category determination and update the confidence scores; this step realizes the quantitative reliability assessment of the identification results.

[0012] Step 6: Compare the confidence level of the signal instance with a preset threshold. For signal instances that are higher than the threshold, automatically output the recognition result and store it in the frequency information table. For signal instances that are lower than the threshold, package their relevant data and submit them to the expert system for secondary judgment. This step forms a decision-making mechanism that combines automated processing with human-machine collaboration.

[0013] A fast broadband signal detection and identification device for implementing the above method includes:

[0014] The signal time-frequency characterization module is used to acquire broadband data and perform short-time Fourier transform to generate a broadband signal time-frequency diagram.

[0015] The time-frequency graph enhancement module is connected to the signal time-frequency characterization module and is used to perform outlier removal, smoothing filtering and normalization processing on the broadband signal time-frequency graph to output a normalized time-frequency graph.

[0016] The signal time-frequency localization module is connected to the time-frequency map enhancement module and has a built-in YOLO-X network model for receiving the normalized time-frequency map and outputting the time-frequency boundary parameters of the signal.

[0017] The signal instance construction module, connected to the signal time-frequency positioning module, is used to extract the corresponding time segment from the original broadband data according to the received time-frequency boundary parameters, perform digital down-conversion to obtain the baseband narrowband signal, and calculate its Welch power spectrum, thereby constructing a complete signal instance for each detected signal.

[0018] The signal recognition and confidence assessment module, connected to the signal instance construction module, is used to perform two-level feature analysis and classification for each signal instance: first, it performs rapid classification based on its time-frequency boundary parameters and generates an initial confidence score; then, it performs fine analysis based on Welch power spectrum features, outputs the final category through the signal classification model, and updates the confidence score.

[0019] The intelligent decision-making and result management module is connected to the signal recognition and confidence assessment module. It is used to compare the confidence of each signal instance with a preset threshold, automatically output the recognition result for signal instances with high confidence and store it in the frequency information table; for signal instances with low confidence, it packages their complete data packets and submits them to the expert system for secondary judgment.

[0020] The beneficial effects of this invention are as follows: By combining time-frequency image enhancement processing with the YOLO-X deep neural network, this invention effectively suppresses background noise and enhances signal features, thereby significantly improving the signal detection accuracy and robustness in low signal-to-noise ratio environments. Simultaneously, by combining digital down-conversion based on detection results with Welch power spectrum analysis, the analysis dimension is extended from the time-frequency image to the signal itself, enabling the system to acquire richer and more robust spectral features, greatly improving the accuracy and reliability of signal recognition. Finally, through the synergistic effect of multi-level confidence assessment and intelligent decision-making mechanisms, the invention combines automated processing of high-confidence results with expert judgment of low-confidence results, ultimately enabling broadband signal detection to balance detection speed, recognition accuracy, and decision reliability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention;

[0022] Figure 2 This is a structural block diagram of the device of the present invention;

[0023] Figure 3 This is a schematic diagram of the hardware connection relationship of the device of the present invention;

[0024] Figure 4 This is a block diagram of the signal time-frequency characterization module;

[0025] Figure 5 This is a block diagram of the control transmission module;

[0026] Figure 6 This is a comparison of the detection accuracy of the method of the present invention and the traditional signal detection method under the same experimental conditions.

[0027] Figure 7 The figure shows a comparison of the recognition accuracy between the signal recognition method without confidence assessment and the method of this invention. Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings. Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The directional terms such as left, center, right, top, and bottom in the embodiments of the present invention are only relative concepts or referenced to the normal use state of the product, and should not be considered restrictive.

[0029] A fast broadband signal detection and recognition method based on time-frequency graphs, such as Figure 1 As shown, it includes the following steps:

[0030] Step 1: Acquire broadband data. The analog signal received by the RF front-end is down-converted and sampled by the ADC to obtain a digital baseband signal. A short-time Fourier transform is performed on the broadband data. The window function can be a Hamming window to generate a broadband signal time-frequency diagram (time-frequency-amplitude), converting the one-dimensional time-domain signal into a two-dimensional time-frequency domain representation. This process includes down-conversion of the RF signal to the intermediate frequency signal, a second down-conversion of the intermediate frequency signal to the baseband IQ signal, and sampling through an analog-to-digital converter to complete the conversion of the analog signal to a digital signal.

[0031] In practice, the received broadband radio frequency signal is down-converted to an intermediate frequency (IF) signal; the IF signal is then down-converted a second time to obtain a baseband analog signal with zero or near-zero IF; and the baseband analog signal is then sampled and quantized at high speed using an analog-to-digital converter to obtain a discrete digital signal sequence. ,in, For sampling point index; for signal The formula for calculating the short-time Fourier transform is as follows:

[0032]

[0033] in, At a certain point in time The analysis window function at the location, The angular frequency is used to generate a three-dimensional time-frequency spectrum of the signal through this transformation, providing a refined time-frequency domain representation for subsequent detection and recognition.

[0034] Step 2: Perform outlier removal, smoothing filtering, and normalization on the three-dimensional data of the broadband signal time-frequency graph to suppress noise, enhance signal features, and form a normalized time-frequency graph suitable for deep neural network processing. Specifically, this includes: first, using a median filter to remove salt-and-pepper noise-like outliers in the time-frequency graph to avoid these anomalies affecting subsequent processing; second, using a two-dimensional Gaussian filter for smoothing to suppress random noise and local jitter, making the time-frequency graph smoother while preserving the main contours and boundaries of the signal; finally, using a minimum-maximum normalization method to linearly scale the amplitude data to the [0,1] interval, compressing the dynamic range and making signals of different intensities comparable at the same scale.

[0035] This step first removes or replaces isolated points or extreme values ​​in the time-frequency amplitude data that significantly deviate from the overall distribution. A median filter is used to effectively remove outliers such as salt-and-pepper noise while better preserving the edge details of the signal. Second, the amplitude data is smoothed within a certain range by convolving a two-dimensional Gaussian filter with the image to reduce random noise and local jitter, making the time-frequency graph smoother while preserving the main contours and boundaries of the signal. Finally, the amplitude data is mapped to [0,1] to compress the dynamic range, making signals of different intensities comparable at the same scale, which facilitates subsequent model processing and feature extraction.

[0036] Step 3: Input the normalized time-frequency map into the trained lightweight YOLO-X network model. By regressing the bounding boxes of the energy concentration regions in the time-frequency map, the model outputs the time-frequency boundary parameters of each detected signal. These parameters include the start frequency, end frequency, start time, and end time, completing the initial localization of the signal in the time-frequency domain. This step establishes a basic identity for each detected signal, providing core localization information for the subsequent formation of complete signal instances. The YOLO-X network model is a lightweight network optimized through model quantization and pruning, deployed on a Jetson TX2 embedded GPU platform for high-speed parallel inference. The normalized coordinates of the output bounding boxes can be multiplied by the time and frequency resolution of the time-frequency map to obtain the actual start and end times and frequencies of the signal. Specifically, the bounding box information is typically represented as follows: , representing the center point coordinates, width, and height of the rectangle, respectively, and the start time. can be The termination time is obtained by multiplying by the time resolution. can be Multiplying by the time resolution yields the initial frequency. can be The termination frequency is obtained by multiplying by the frequency resolution. can be Multiply by the frequency resolution to obtain;

[0037] The YOLO-X network model is a high-performance, anchor-free target detector with advantages of fast detection speed and high accuracy. It takes a pre-processed time-frequency image as input and feeds it into a pre-trained YOLO-X network model. The model performs bounding box regression on the energy-concentrated regions in the time-frequency image, directly outputting the start and end frequencies of each candidate signal on the frequency axis, and the start and end times on the time axis, thus accurately describing the time-frequency boundaries of the signal. Unlike traditional methods based on fixed thresholds, this detection process maintains high detection accuracy and robustness even in complex scenarios with low signal-to-noise ratios and overlapping signals.

[0038] Step 4: Based on the time-frequency boundary parameters, extract the corresponding time segment from the original broadband data, perform digital down-conversion at the detected center frequency to shift it to the baseband, and obtain the separated baseband narrowband signal through bandpass filtering and decimation. In specific implementation, firstly, use the detected center frequency as the local oscillation frequency to perform digital down-conversion on the signal, shifting it to the zero intermediate frequency or baseband position; then, through bandpass filtering and decimation operations, retain only the effective frequency components corresponding to the detection bandwidth; through the above processing, the signal region detected in Step 3 is transformed into a signal instance containing specific signal entities; then, the Welch power spectrum of each baseband narrowband signal is calculated using the Welch method to obtain a smooth and stable power spectral density curve; thus, each signal instance contains complete time-frequency boundary parameters, baseband narrowband signal, and Welch power spectrum; wherein, the calculation of the Welch power spectrum specifically includes: dividing the baseband narrowband signal into several potentially overlapping segments, performing Fourier transform on each segment after windowing to obtain a periodogram, and then averaging the periodograms of all segments to obtain a smooth power spectral density curve;

[0039] In practice, based on the start frequency, end frequency, and start and end times output by the YOLO-X network model, the corresponding time segments in the original broadband sampling data are extracted; and the detected center frequency is used as the basis for further analysis. As the local oscillation frequency, the center frequency The signal is digitally down-converted to shift it to baseband; this process is achieved by multiplying the signal by a complex oscillator.

[0040]

[0041] in, It is the sampling period. The imaginary unit satisfies In engineering and communication signal processing, it is commonly used Non-mathematical symbols To avoid confusion with current symbols This leads to confusion; the aforementioned complex exponential term is equivalent to generating a numerical local oscillator signal, used to perform frequency shifting of the input signal. Then, a low-pass filter is used to... Filtering is performed to remove signals of other frequencies and image frequency components, and downsampling is performed to obtain a clean baseband narrowband signal. For each obtained baseband signal, the Welch method is further used for power spectrum estimation. The signal is segmented and windowed, and the periodogram of each segment is calculated. Then, the average of multiple segments is taken to reduce the variance and obtain a smooth and stable power spectral density curve. Compared with the traditional method that relies on single-frame FFT, the power spectrum estimation based on Welch averaging can effectively improve the robustness of spectrum estimation and the distinguishability of signal features in low signal-to-noise ratio environments.

[0042] Step 5: Based on the time-frequency boundary parameters, classify and distinguish signal instances, such as continuous wave signals, amplitude-modulated signals, frequency-modulated signals, and multi-carrier signals, and generate an initial confidence level based on the classification result; then, extract statistical features, including peak spacing, linewidth, and out-of-band leakage, from the Welch power spectrum curve; based on the matching degree between the statistical features and the signal category classifier (preset signal category classifier), output the final signal category determination and update the confidence level, thus completing the quantitative reliability assessment of the signal instances;

[0043] This step involves two levels of feature analysis and classification. First, based on the start and end frequency information output by the YOLO-X network model, combined with the center frequency and bandwidth after down-conversion, the signal is quickly classified and distinguished, differentiating between continuous wave signals, amplitude-modulated signals, frequency-modulated signals, multi-carrier signals, etc., and each detected signal is assigned an initial confidence score to reflect the reliability of the recognition result. Second, the baseband signal obtained from down-conversion and its corresponding Welch power spectrum are analyzed in depth to extract more detailed feature parameters, including spectral peak spacing, spectral linewidth, out-of-band leakage, spectral roll-off rate, flatness, symmetry, and modulation-related statistical features (such as constellation diagram analysis, higher-order cumulant analysis, etc.). The extracted features are combined into a feature vector and input into a deep neural network specifically for signal classification. After training, this classifier can comprehensively judge complex feature combinations and output a more accurate signal category and corresponding confidence score. The system compares the confidence scores of all possible categories and selects the one with the highest score as the final recognition result.

[0044] Step 6: Compare the confidence level of the signal instance with a preset threshold. For signal instances with a confidence level higher than the threshold, automatically output the identification result and store it in the frequency information table. For signal instances with a confidence level lower than the threshold, package their relevant data and submit them to the expert system for secondary discrimination. The trigger condition for the secondary discrimination by the expert system is that the highest confidence level of the signal instance is lower than the preset threshold (e.g., 0.7). The system will automatically mark the signal instance as pending review and push its corresponding time-frequency slice, baseband narrowband signal, Welch power spectrum, and preliminary identification result to the expert system. For signal instances with a high confidence level, the system directly outputs its discrimination result, including key information such as the category, start and end frequencies, and start and end times of the signal instance, to the frequency information table, thus completing the automated processing flow of the signal instance.

[0045] To ensure system reliability, a confidence threshold (e.g., 0.7) is set. If the highest confidence level output in step 5 is lower than this threshold, it indicates that the model's judgment of the signal is uncertain, possibly due to poor signal quality, rare signal type, or strong interference. In this case, the system will package all relevant information of the signal (time-frequency slices, baseband data, power spectrum, preliminary identification results, etc.) and mark it as "pending expert review," then push it to professionals for manual analysis and judgment. For signals with high confidence, the system will directly output their discrimination results, including key information such as signal category, start and end frequencies, and start and end times. Subsequently, this information will be stored in the frequency information table for recording and management. The frequency information table, as a centralized database, stores detailed frequency data for each signal, facilitating subsequent review, analysis, and monitoring, while providing a reliable basis for subsequent signal management, interference investigation, and other intelligent decision-making.

[0046] A fast broadband signal detection and identification device for implementing the above method, such as Figure 2 and Figure 3 As shown, it includes:

[0047] The signal time-frequency characterization module is used to acquire broadband data and perform short-time Fourier transform to generate a broadband signal time-frequency diagram; combined with... Figure 4 As shown, this module implements software radio functionality based on the AD9361 chip, providing... The transceiver channel performs functions such as up-conversion, down-conversion, filtering, and amplification of signals. This module serves as the analog front-end of the device and the interface with the physical electromagnetic environment, responsible for high-fidelity capture of airborne signals. Based on the AD9361 broadband transceiver chip, it supports a wide operating frequency range from 100MHz to 6GHz, and has two independent receiving channels and two independent transmitting channels, providing a flexible configuration. MIMO transceiver link; the module integrates a programmable low-noise amplifier (LNA), automatic gain control (AGC) loop, high dynamic range mixer, and adjustable bandwidth analog filter; via the SPI bus, the control transmission module can precisely digitally control its internal parameters, including real-time adjustment of receive gain, selection of up / down conversion operating modes, and precise setting of local oscillator frequency; the module down-converts, filters, and amplifies the received RF signal, and outputs an analog I / Q baseband signal to the subsequent signal processing module;

[0048] The time-frequency graph enhancement module, connected to the signal time-frequency characterization module, is used to perform outlier removal, smoothing filtering, and normalization on the broadband signal time-frequency graph, outputting a normalized time-frequency graph. This module uses a Zynq-7000 series chip (such as XC7Z100) as its core processor to complete the device's digital signal processing. This module is the core of the device's digital signal processing, responsible for performing processing tasks with extremely high real-time and throughput requirements. The core hardware platform is Xilinx. The Zynq-7000 series chips (such as the XC7Z100) have a heterogeneous architecture that meets the requirements of this invention. The internal high-performance FPGA programmable logic is used to implement fixed, large-scale parallel signal processing algorithms, including high-speed sampling of analog I / Q signals from the RF transceiver module via a high-speed analog-to-digital converter interface, and subsequent data preprocessing operations such as digital down-conversion, FIR filter-based channel filtering, and wavelet transform-based noise reduction. Simultaneously, it efficiently performs power spectral density estimation and short-time Fourier transform calculations in spectrum analysis. Based on the requirements of the signal processing module, and considering device cost and future upgrades, a resource requirement and chip resource ratio of 70% is recommended. Therefore, the XC7Z100 chip is selected, with a surface-mount FFG900 package, a size of 31mm × 31mm, and up to 360 Select I / O resources and a maximum of 130 PS I / O resources.

[0049] The signal time-frequency localization module, connected to the time-frequency map enhancement module, incorporates a YOLO-X network model to receive the normalized time-frequency map and output the signal's time-frequency boundary parameters. This module is implemented on a Jetson TX2 embedded GPU platform, with its core being an optimized, lightweight YOLO-X network model deployed on this platform, specifically responsible for parallel inference of the time-frequency map and signal localization. The module integrates 256 NVIDIA CUDA cores to achieve high-speed parallel inference of the deep learning model. This module serves as the intelligent computing and acceleration unit of the device, completing the deep learning inference task. It uses an NVIDIA Jetson TX2 processor, integrating 256 NVIDIA CUDA cores, a multi-core 64-bit ARM v8 processor. The Jetson TX2 processor integrates a 128-bit DDR4 memory controller, supporting ECC functionality to ensure reliability while meeting storage capacity requirements. Additionally, the module integrates 32GB (256Gbit) of eMMC. The FLASH module meets the system's operational requirements. It connects to the signal processing module via a high-speed PCIe bus to ensure low-latency data transmission. A lightweight YOLO-X model, optimized and quantized using tools such as NVIDIA TensorRT, runs on it. After receiving a time-spectrum slice from the signal processing module, the GPU module uses its CUDA cores to perform high-speed parallel inference, outputting the precise bounding box coordinates of the signal on the time-spectrum and resolving the pixel coordinates into high-precision bandwidth and time parameters.

[0050] A signal instance construction module, connected to the signal time-frequency localization module, is used to extract the corresponding time segment from the original broadband data based on the received time-frequency boundary parameters, perform digital down-conversion to obtain the baseband narrowband signal, and calculate its Welch power spectrum, thereby constructing a complete signal instance for each detected signal. This module implements digital down-conversion and Welch power spectrum calculation on an FPGA, utilizing the FPGA to handle high-throughput, low-latency real-time digital signal processing tasks. Based on the time-frequency boundary parameters output by the YOLO-X model, this module extracts the corresponding time segment from the original broadband data, performs digital down-conversion at the detected center frequency to shift it to the baseband, and obtains the separated baseband narrowband signal through bandpass filtering and decimation. Subsequently, the Welch method is used to calculate the power spectrum of each baseband narrowband signal to obtain a smooth and stable power spectral density curve. These operations are implemented on the FPGA, utilizing the FPGA's high parallel processing capabilities to ensure real-time performance and high throughput.

[0051] The signal recognition and confidence assessment module, connected to the signal instance construction module, performs two-level feature analysis and classification for each signal instance: first, it performs rapid classification based on its time-frequency boundary parameters and generates an initial confidence score; then, it performs fine analysis based on Welch power spectrum features, outputs the final category through a signal classification model, and updates the confidence score. This module first compares the time-frequency boundary parameters (such as center frequency and bandwidth) with a preset signal parameter library to perform rapid classification and generate an initial confidence score; subsequently, it performs in-depth analysis of the baseband signal and its Welch power spectrum, extracting features including the number of spectral peaks, spectral line positions, spectral flatness, and power percentage bandwidth, and combines these with modulation mode features (such as constellation diagram analysis and higher-order cumulant analysis) to form a feature vector, which is then input into a deep neural network classifier for fine classification, outputting the final signal category and updating the confidence score.

[0052] The intelligent decision-making and result management module, connected to the signal recognition and confidence assessment module, compares the confidence level of each signal instance with a preset threshold. For high-confidence signal instances, it automatically outputs the recognition result and stores it in the frequency information table; for low-confidence signal instances, it packages their complete data packets and submits them to the expert system for secondary judgment. This module receives the decision result and confidence level output by the signal recognition module, compares the final confidence level with a preset threshold (e.g., 0.7); formats and outputs high-confidence results and stores them in the frequency information table (stored in memory); marks low-confidence results and packages their related data (time-frequency plot slices, baseband data, power spectrum, preliminary recognition results, etc.), submitting them to experts for manual judgment through a human-computer interaction interface; the frequency information table serves as a centralized database, storing detailed frequency data for each signal for easy subsequent review, analysis, and monitoring.

[0053] The fast broadband signal detection and identification device also includes a control and transmission module for unified device configuration and status management, data scheduling, and external transmission; combined with Figure 5As shown, this module configures and manages onboard devices such as the RF chip AD9361 via the ZYNQ chip's SPI bus, and establishes communication with the PC via an Ethernet switch and a high-speed data transmission bus, enabling centralized control and data exchange for the entire device. The control and transmission module's functions are primarily handled by the ARM processor system in the signal processing module and the embedded operating system running on it. This ARM processor, acting as the main control unit, configures and monitors the RF transceiver module (such as AD9361) via the SPI bus, manages the signal processing IP cores within the FPGA via the AXI bus, and distributes tasks and retrieves results from the GPU module via the PCIe bus. It maintains a first-in-first-out (FIFO) asynchronous task queue, enabling pipelined operation of signal processing and GPU inference. Simultaneously, this processor runs the gigabit Ethernet interface communication protocol stack, packaging the final detection and identification results and transmitting them in real-time to the higher-level control center via the network or storing and displaying them locally. For transmission, the PC sends commands via the RJ-45 interface, which are then transmitted to the baseband processing chip via the high-speed data transmission bus between the Ethernet switch and the ZYNQ chip, thus enabling management and control of the device.

[0054] Through collaborative work, the various modules form a complete processing chain from signal reception, time-frequency analysis, signal instance construction, intelligent recognition to decision output, enabling parallel processing and intelligent management of multiple signal instances in a broadband environment. The device adopts a heterogeneous architecture of FPGA+GPU+CPU, giving full play to the unique advantages of various processors: using FPGA to handle high-throughput, low-latency real-time digital signal processing tasks; using GPU as a massively parallel computing engine to accelerate deep learning model inference; and using CPU for top-level scheduling, state management, and external communication, thereby simultaneously meeting the stringent requirements of real-time performance and high-precision intelligent processing.

[0055] An electronic device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the above-described method; the electronic device integrates data acquisition hardware to form a complete end-to-end signal detection and recognition system, and through the coordinated work of its internal functional modules, realizes an automated processing flow from data acquisition, time-frequency analysis, intelligent detection to fine recognition and result storage.

[0056] A computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method; the storage medium is a tangible, non-transient storage medium that carries the technical core of the method of the present invention, enabling it to be distributed, installed, and deployed in the form of a software product, thereby replicating the technical effects of the present invention on various general-purpose or dedicated hardware platforms.

[0057] like Figure 6 As shown, traditional signal detection methods employ a typical detection strategy based on energy detection (ED), which is computationally simple and widely used in engineering. Therefore, as a comparative method, the method of this invention uses a deep learning method based on YOLO-X to perform deep representation of the input signal before completing the detection decision. This study uses detection accuracy as the main performance evaluation index, defined as follows:

[0058]

[0059] Among them, TP, TN, FP, and FN represent the number of true positives, true negatives, false positives, and false negatives detected, respectively;

[0060] from Figure 6 As can be seen, the detection accuracy of traditional energy detection methods is 88.4%, while the method proposed in this invention achieves a detection accuracy of 98.5%, representing a performance improvement of approximately 10.1%. Traditional methods are prone to decision instability under complex environments and noise interference, while this scheme significantly improves signal separability and robustness through deep feature extraction; therefore, this scheme demonstrates significantly better performance than traditional methods in the detection stage.

[0061] like Figure 7 As shown, the signal recognition method without confidence assessment employs a simple method based on amplitude spectrum statistical features combined with manual threshold decision-making. This method first extracts basic statistical quantities such as the mean, energy, and kurtosis of the amplitude spectrum from the received signal, and then completes the classification decision through a preset threshold rule; therefore, the recognition accuracy is limited under complex channel conditions. This invention introduces a deep feature fusion and confidence discrimination mechanism into the recognition model, thereby improving the overall classification performance; the recognition performance is evaluated using recognition accuracy, which is defined as:

[0062]

[0063] That is, the proportion of correctly identified samples out of all test samples;

[0064] from Figure 7 As can be seen, the recognition accuracy of traditional recognition methods is 86.3%, while the method of the present invention reaches 97.9%, improving the recognition performance by 11.6%. Especially in the recognition task of complex signals of multiple categories, the method of the present invention can more effectively distinguish subtle differences between categories, improving the reliability and stability of recognition.

[0065] Figure 6 and Figure 7The experimental results fully demonstrate that the detection and recognition methods proposed in this invention are significantly superior to traditional methods in terms of performance. Under the same experimental settings, the method of this invention improves the detection accuracy by about 10% and the recognition accuracy by more than 11%. This shows that by using deep feature extraction, fusion mechanism and confidence decision strategy, the robustness of signal detection and recognition can be significantly enhanced, thus effectively proving the superiority and practical value of the solution of this invention.

[0066] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A fast broadband signal detection and identification method based on time-frequency graphs, characterized in that, Includes the following steps: Step 1: Collect broadband data, perform short-time Fourier transform on the broadband data, and generate a broadband signal time-frequency diagram; Step 2: Perform outlier removal, smoothing filtering, and normalization on the three-dimensional data of the broadband signal time-frequency graph to suppress noise, enhance signal features, and form a normalized time-frequency graph suitable for deep neural network processing; Step 3: Input the normalized time-frequency map into the trained YOLO-X network model. By regressing the bounding boxes of the energy concentration regions in the time-frequency map, output the time-frequency boundary parameters of each detected signal. The time-frequency boundary parameters include the start frequency, end frequency, start time, and end time, thus completing the initial localization of the signal in the time-frequency domain. Step 4: Based on the time-frequency boundary parameters, extract the corresponding time segment from the original broadband data, perform digital down-conversion at the detected center frequency to shift it to the baseband, and obtain the separated baseband narrowband signal through bandpass filtering and decimation; transform the signal region detected in Step 3 into a signal instance containing a specific signal entity; then calculate the Welch power spectrum of each baseband narrowband signal using the Welch method to obtain a smooth and stable power spectral density curve; thus, each signal instance contains complete time-frequency boundary parameters, baseband narrowband signal, and Welch power spectrum. Step 5: Based on the time-frequency boundary parameters, perform preliminary classification of signal instances and assign initial confidence scores; then extract statistical features, including peak spacing, linewidth, and out-of-band leakage, from the Welch power spectrum curve; based on the matching degree between the statistical features and the signal classifier, output the final signal class determination and update the confidence scores. Step 6: Compare the confidence level of the signal instance with a preset threshold, and automatically output the recognition result for signal instances that are higher than the threshold and store them in the frequency information table; For signal instances below the threshold, their relevant data are packaged and submitted to the expert system for secondary judgment.

2. The fast broadband signal detection and identification method based on time-frequency diagrams according to claim 1, characterized in that: The YOLO-X network model is a lightweight network that has undergone model quantization and pruning optimization, and it is deployed on a GPU to achieve parallel inference.

3. The fast broadband signal detection and identification method based on time-frequency diagrams according to claim 1, characterized in that: The calculation of the Welch power spectrum specifically includes: dividing the baseband narrowband signal into several overlapping segments, windowing each segment and performing a Fourier transform to obtain a periodogram, and then averaging the periodograms of all segments to obtain a smooth power spectral density curve.

4. The fast broadband signal detection and identification method based on time-frequency diagrams according to claim 1, characterized in that: The trigger condition for the expert system's secondary judgment is that the highest confidence level of the signal instance is lower than a preset threshold. The system will automatically mark the signal instance as pending review and push its corresponding time-frequency plot slice, baseband narrowband signal, Welch power spectrum, and preliminary identification results to the expert system.

5. A fast broadband signal detection and identification device for implementing the above method, characterized in that, include: The signal time-frequency characterization module is used to acquire broadband data and perform short-time Fourier transform to generate a broadband signal time-frequency diagram. The time-frequency graph enhancement module is connected to the signal time-frequency characterization module and is used to perform outlier removal, smoothing filtering and normalization processing on the broadband signal time-frequency graph to output a normalized time-frequency graph. The signal time-frequency localization module is connected to the time-frequency map enhancement module and has a built-in YOLO-X network model for receiving the normalized time-frequency map and outputting the time-frequency boundary parameters of the signal. The signal instance construction module, connected to the signal time-frequency positioning module, is used to extract the corresponding time segment from the original broadband data according to the received time-frequency boundary parameters, perform digital down-conversion to obtain the baseband narrowband signal, and calculate its Welch power spectrum, thereby constructing a complete signal instance for each detected signal. The signal recognition and confidence assessment module, connected to the signal instance construction module, is used to perform two-level feature analysis and classification for each signal instance: first, it performs rapid classification based on its time-frequency boundary parameters and generates an initial confidence score; then, it performs fine analysis based on Welch power spectrum features, outputs the final category through the signal classification model, and updates the confidence score. The intelligent decision-making and result management module is connected to the signal recognition and confidence assessment module. It is used to compare the confidence of each signal instance with a preset threshold, automatically output the recognition result for signal instances with high confidence and store it in the frequency information table; for signal instances with low confidence, it packages their complete data packets and submits them to the expert system for secondary judgment.

6. The fast broadband signal detection and identification device for implementing the above method according to claim 5, characterized in that: The signal time-frequency localization module adopts a GPU-accelerated YOLO-X network model, and the signal instance construction module implements digital down-conversion and Welch power spectrum calculation on the FPGA.

7. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-4.

Images