Frequency hopping signal parameter blind detection method and system based on deep learning

Abstract

The invention provides a frequency hopping signal parameter blind detection method and system based on deep learning, and the method comprises the steps: carrying out the S transformation time-frequency analysis of a received mixed signal, and generating a time-frequency graph for suppressing a very low frequency interference signal and Gaussian white noise; a hybrid encoder architecture is constructed based on an RMT backbone network, and multi-scale time-frequency features are extracted; inputting the time-frequency diagram into a hybrid encoder architecture, extracting multi-level features through an RMT backbone network, generating a target query through a hybrid encoder, and outputting bounding box coordinates of a frequency hopping signal by a decoder; and according to the physical parameters of the bounding box coordinate mapping frequency hopping signal, calculating the frequency hopping frequency and the hopping time, and completing robust estimation of the frequency hopping frequency and the hopping time. According to the invention, direct estimation of frequency hopping frequency and time is realized by using a reverse mapping mechanism of the TF graph.

Description

Technical Field

[0001] This invention belongs to the field of frequency hopping signal detection technology, and specifically relates to a blind detection method and system for frequency hopping signal parameters based on deep learning. Background Technology

[0002] Extremely low frequency (VLF, 3-30 kHz) communication is widely used in underwater communication, biomimetic robot control, and other scenarios due to its low signal attenuation, strong penetration, and resistance to multipath interference. For example, underwater vehicles need to achieve reliable long-distance communication via the VLF band, but the VLF environment is subject to complex electromagnetic interference, including fixed-frequency signals (such as navigation signals), linear frequency modulated signals (radar interference), and burst noise (lightning pulses). These interferences pose a significant challenge to the detection and parameter estimation of frequency hopping (FH) signals in VLF communication links.

[0003] Frequency-hopping spread spectrum (FHSS) technology achieves anti-interference and low interception probability by dynamically switching carrier frequencies, and is one of the core technologies of VLF communication. However, the application of FHSS in VLF environments needs to address the following issues: Feature extraction under low signal-to-noise ratio (SNR): When SNR < 0 dB, the FH signal is submerged in noise, making it difficult to extract time-frequency features using traditional methods; Differentiation of complex interference: The time-frequency characteristics of VLF band interference signals overlap with those of FH signals (such as the short-term similarity between burst signals and frequency-hopping signals), leading to an increased false detection rate; Insufficient model generalization: Existing algorithms require parameter adjustments for different SNR conditions, lacking a unified end-to-end solution.

[0004] Existing time-frequency analysis methods, such as the Short-Time Fourier Transform (STFT), are widely used to generate time-frequency graphs by performing Fourier transforms on signal segments captured through a sliding window. However, they have significant drawbacks: fixed time-frequency resolution – the fixed window length leads to low time resolution for high-frequency signals and insufficient frequency resolution for low-frequency signals (e.g., ...). Figure 2 As shown, the time-frequency plot of the FH signal is blurred when SNR = 6dB; spectral leakage and cross-terms: asynchronous truncation leads to spectral energy diffusion, and interference overlaps with the FH signal on the time-frequency plot. S-transform (ST): Combining the advantages of STFT and continuous wavelet transform, it improves time-frequency resolution through a frequency adaptive window (high time resolution for high-frequency signals, high frequency resolution for low-frequency signals). Literature attempts to use ST for FH signal analysis, but it has not solved the noise suppression problem at low SNR (<-10dB).

[0005] Deep learning model CNN / RNN hybrid model: The literature proposes a hybrid model based on multiple spectral maps (STFT, WVD), but does not consider the interference of noise on the convolution kernel under low SNR, resulting in insufficient robustness (detection accuracy drops by more than 40% when SNR<-6 dB).

[0006] Semantic segmentation model: The literature uses Deeplabv3+ network to segment the FH signal region in the time-frequency plot, but does not model the correlation of signal jump time series, and is sensitive to sudden interference (false detection rate >25%).

[0007] ResNet combined with generalized S-transform: The literature extracts time-frequency map features through ResNet, but the model has a large number of parameters (>50M), poor real-time performance (inference speed <30 FPS), and the AP50-95 (average precision) drops below 0.3 under low SNR.

[0008] RT-DETR: A Transformer-based real-time detection model that achieves end-to-end detection through an encoder-decoder architecture, eliminating the need for non-maximum suppression (NMS) post-processing. However, its backbone network (ResNet-50) lacks the ability to model priors in the time-frequency graph space. However, the STFT of time-frequency analysis methods has limitations: when SNR < -10 dB, the time-frequency plot of the FH signal cannot be distinguished from noise; the calculation of the overlap window (e.g., 50% overlap rate) increases the algorithm complexity by more than 30%, making it difficult to meet the requirements of real-time detection. Existing research has not fully utilized the frequency adaptive characteristics of ST, and has not combined noise suppression algorithms (e.g., wavelet thresholding denoising) at low SNR, resulting in limited improvement in the signal-to-noise ratio of the time-frequency plot.

[0009] The problem of local feature dependence in deep learning models: CNN-based models (such as YOLOv8 / YOLOv9) extract local features through convolutional kernels, but lack the ability to model the global semantics of time-frequency graphs. For example, when SNR < -6 dB, YOLOv8 mistakenly identifies sudden interference as FH signal because it ignores the temporal continuity of frequency hopping signals (false detection rate > 35%).

[0010] The ResNet-50 backbone network of RT-DETR does not incorporate spatial prior knowledge of the time-frequency map (such as the preference of the FH signal in the center region of the time-frequency map), resulting in large deviations in the detection box localization.

[0011] Traditional methods have a root mean square error (RMSE) of up to 60 Hz (when SNR=0 dB) for frequency hopping estimation under complex interference, while the time parameter estimation error exceeds 2 ms, which cannot meet the requirements of high-precision communication synchronization. Summary of the Invention

[0012] The purpose of this invention is to provide a blind detection method and system for frequency hopping signal parameters based on deep learning, so as to solve the above-mentioned problems.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a blind detection method for frequency hopping signal parameters based on deep learning, comprising: S-transform time-frequency analysis is performed on the received mixed signal to generate a time-frequency plot to suppress very low frequency interference signals and Gaussian white noise; A hybrid encoder architecture is constructed based on the RMT backbone network to extract multi-scale time-frequency features; The time-frequency map is input into the hybrid encoder architecture, multi-level features are extracted through the RMT backbone network, the target query is generated by the hybrid encoder, and the bounding box coordinates of the frequency hopping signal are output by the decoder. Based on the physical parameters of the frequency-hopping signal mapped from the bounding box coordinates, the frequency hopping frequency and hopping time are calculated, and a robust estimation of the frequency hopping frequency and hopping time is completed.

[0014] Furthermore, the step of performing S-transform time-frequency analysis on the received mixed signal to generate a time-frequency diagram to suppress very low frequency interference signals and Gaussian white noise includes: The received signal is denoised and normalized. S-transform calculation: Using Gaussian window function , where the scale factor To achieve frequency-adaptive time-frequency resolution; Through formula Generate a time-frequency graph and suppress cross-term interference.

[0015] Furthermore, the hybrid encoder architecture built on the RMT backbone network includes: RMT backbone network replacement: The original RT-DETR ResNet-50 backbone is replaced with an RMT network, whose core is the Manhattan self-attention MaSA mechanism. Hybrid encoder design: Extracting multi-scale features from RMT backbone output ; A hybrid encoder is constructed using the attention-based cross-scale feature interaction module AIFI and the CNN cross-scale feature fusion module CCFF; An initial target query is generated by minimizing uncertainty in the query selection strategy, thereby optimizing detection efficiency. The decoder uses a Transformer architecture to generate bounding boxes (Bboxes) and class predictions through iterative optimization of the target query; the prediction head outputs the coordinates and confidence scores of the frequency-hopping signals Bboxes.

[0016] Furthermore, a spatial decay matrix based on Manhattan distance is introduced. Construct a two-dimensional spatial prior; calculate the self-attention weights as follows: ,in For querying key vectors, For dimensions.

[0017] Furthermore, the model is trained and improved using a time-frequency graph dataset to achieve frequency hopping signal detection and parameter estimation: Input image size, batch size, and number of training epochs; The optimizer used is AdamW, with an initial learning rate of 1e-3, and the learning rate is adjusted using a cosine annealing strategy. The loss function is a weighted combination of GIoU loss and cross-entropy loss; The input time-frequency map is used to extract features through the RMT backbone, and a target query is generated through a hybrid encoder. The decoder refines the query layer by layer and outputs the frequency hopping signal bounding box and confidence level. A confidence level threshold is set to filter out low-confidence detection results.

[0018] Furthermore, the step of inputting the time-frequency map into the hybrid encoder architecture, extracting multi-level features through the RMT backbone network, generating a target query through the hybrid encoder, and outputting the bounding box coordinates of the frequency hopping signal by the decoder includes: Model output bounding box coordinates ,in The coordinates of the top left corner The coordinates are the bottom right corner.

[0019] Furthermore, the step of calculating the frequency hopping frequency and transition time based on the physical parameters of the frequency hopping signal mapped from the bounding box coordinates, and completing the robust estimation of the frequency hopping frequency and transition time, includes: Based on the mapping relationship of the vertical axis of the time-frequency diagram, the frequency hopping frequency... The calculation is as follows:

[0020] in The maximum frequency in the spectrum. The ordinate of the bounding box center. The height of the time-frequency graph: Jump time estimation: Jump cycle Obtained by mapping the horizontal span of the bounding box:

[0021] in The total duration of signal observation. This represents the width of the time-frequency graph.

[0022] Secondly, the present invention provides a blind detection system for frequency hopping signal parameters based on deep learning, comprising: The data acquisition module is used to perform S-transform time-frequency analysis on the received mixed signal, generate a time-frequency diagram to suppress very low frequency interference signals and Gaussian white noise; A hybrid encoder architecture building module is used to build a hybrid encoder architecture based on the RMT backbone network and extract multi-scale time-frequency features. The training module is used to input the time-frequency map into the hybrid encoder architecture, extract multi-level features through the RMT backbone network, generate target queries through the hybrid encoder, and output the bounding box coordinates of the frequency hopping signal by the decoder. The output module is used to calculate the frequency hopping frequency and transition time based on the physical parameters of the frequency hopping signal mapped from the bounding box coordinates, and to complete the robust estimation of the frequency hopping frequency and transition time.

[0023] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the blind detection method for frequency hopping signal parameters based on deep learning.

[0024] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the blind detection method for frequency hopping signal parameters based on deep learning.

[0025] Compared with the prior art, the present invention has the following technical effects: This invention uses the S-transform (ST) instead of the traditional Short Time Fourier Transform (STFT) and introduces the RMT backbone network (a self-attention mechanism based on Manhattan distance).

[0026] ST's time-frequency focusing capability: ST effectively suppresses cross-term interference through variable resolution analysis, and can still clearly separate the short-time spectral characteristics of FH signals even at low SNR. Compared with STFT, ST has a stronger ability to capture weak signals, and its performance is significantly better than STFT when SNR < -10 dB.

[0027] Spatial Prior Enhancement of RMT: RMT's Manhattan self-attention mechanism strengthens local feature associations through a spatial attenuation matrix, making the model pay more attention to the frequency-hopping signal in the central region of the TF diagram, reducing the influence of background noise and interference. Experiments show that the improved RT-DETR has an AP50-95 value that stably exceeds 0.73 when SNR>-12 dB, far surpassing YoloV8 / V9 and the original RT-DETR.

[0028] This invention treats the FH signal as a "short horizontal target" in the TF diagram and directly locates frequency hopping events through an end-to-end detection architecture of RT-DETR, avoiding the sensitive dependence of traditional parameter estimation algorithms on interference type. Experimental data show that the model maintains high detection accuracy even in scenarios with multiple complex interferences.

[0029] This invention utilizes the reverse mapping mechanism of TF diagrams to achieve direct estimation of frequency hopping frequency and time.

[0030] Coordinate-driven parameter decoupling: By using the center coordinates and size of the detection box, the frequency hopping frequency and period are inversely calculated using formulas, avoiding the accumulated errors of traditional methods. Experiments show that the RMSE of the improved RT-DETR is below 30Hz when SNR>-12 dB, significantly outperforming other methods.

[0031] This invention combines the physical characteristics of ST with the global perception capability of Transformer.

[0032] Multi-interference collaborative suppression: The time-frequency joint analysis of ST can distinguish different types of signal features, while the attention mechanism of RMT further suppresses the interference of background noise.

[0033] Dynamic SNR Adaptability: The improved RT-DETR maintains a smooth performance curve over a wide range of SNR, verifying its strong adaptability to complex VLF environments. Attached Figure Description

[0034] Figure 1. Flowchart of frequency hopping signal detection according to the present invention.

[0035] Figure 2. Schematic diagram of the improved RT-DETR model structure.

[0036] Figure 3. Performance comparison of deep learning models under different signal-to-noise ratios (SNR): (a) Performance of each model Indicators; (b) for each model index.

[0037] Figure 4. Performance comparison of this scheme and RT-DETR-ST in frequency hopping signal parameter estimation: (a) RMSE of frequency hopping frequency under different signal-to-noise ratios; (b) RMSE of frequency hopping time under different signal-to-noise ratios. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings: Example 1: This invention provides a blind detection method for frequency hopping signal parameters based on deep learning, comprising: S-transform time-frequency analysis is performed on the received mixed signal to generate a time-frequency plot to suppress very low frequency interference signals and Gaussian white noise; A hybrid encoder architecture is constructed based on the RMT backbone network to extract multi-scale time-frequency features; The time-frequency map is input into the hybrid encoder architecture, multi-level features are extracted through the RMT backbone network, the target query is generated by the hybrid encoder, and the bounding box coordinates of the frequency hopping signal are output by the decoder. Based on the physical parameters of the frequency-hopping signal mapped from the bounding box coordinates, the frequency hopping frequency and hopping time are calculated, and a robust estimation of the frequency hopping frequency and hopping time is completed.

[0039] Example 2: This invention provides a blind detection method for frequency hopping signal parameters based on deep learning, comprising: The core of this invention is to achieve robust detection and blind parameter estimation of FH signals under VLF interference by combining S-transform (ST) time-frequency analysis with an improved RT-DETR model. The specific technical solution is as follows: Time-frequency feature extraction based on S-transform Time-frequency analysis process: Input signal preprocessing: The received signal $c(t)$ is denoised and normalized.

[0040] S-transform calculation: Using Gaussian window function , where the scale factor This enables frequency-adaptive time-frequency resolution.

[0041] Through formula Generate a time-frequency diagram and suppress cross-term interference.

[0042] The S-transform provides high time resolution in the high-frequency region and high frequency resolution in the low-frequency region by dynamically adjusting the window width, thereby more clearly separating the FH signal from interference signals (such as the diagonal stripes of LFM and the short horizontal lines of burst signals) under low SNR conditions.

[0043] Improved RT-DETR model architecture Model structure improvement: RMT backbone network replacement: The original RT-DETR ResNet-50 backbone is replaced with the RMT (Retentive Networks Meet VisionTransformers) network, whose core is the Manhattan Self-Attention (MaSA) mechanism.

[0044] MaSA mechanism design: Introducing a spatial decay matrix based on Manhattan distance Construct a two-dimensional spatial prior. The self-attention weights are calculated as follows: ,in For querying key vectors, For dimensions.

[0045] Multi-stage structure: The first three stages employ Decomposed MaSA, which divides the feature map into local windows for local attention calculation, reducing computational complexity.

[0046] The fourth stage uses global MaSA to capture long-range dependencies.

[0047] Hybrid encoder design: AIFI module (Attention-based Intra-scale Feature Interaction): fuses multi-level features through cross-scale self-attention. This enhances the consistency of semantic information.

[0048] The CCFF module (CNN-based Cross-scale Feature Fusion) uses 3×3 convolutions and upsampling operations to achieve feature map scale alignment and preserve local details.

[0049] Query selection that minimizes uncertainty: The Top-K features output by the static encoder are selected as the initial object query, eliminating the dependence of the traditional DETR model on learnable queries and improving detection stability.

[0050] Mechanism of action: The RMT backbone explicitly introduces spatial priors through the MaSA mechanism, making the model focus more on the central region of the FH signal in the time-frequency graph, while expanding the receptive field to capture the global context; the hybrid encoder improves the localization accuracy of small FH signals through multi-scale feature fusion.

[0051] Frequency hopping parameter estimation method Bounding box decoding and parameter mapping: Model output bounding box coordinates ,in The coordinates of the top left corner The coordinates are the bottom right corner.

[0052] Frequency hopping frequency estimation: Based on the mapping relationship of the vertical axis of the time-frequency diagram, the frequency hopping frequency... The calculation is as follows:

[0053] in The maximum frequency in the spectrum. The ordinate of the bounding box center. The height of the time-frequency graph.

[0054] Jump time estimation: Jump cycle Obtained by mapping the horizontal span of the bounding box:

[0055] in The total duration of signal observation. This represents the width of the time-frequency graph.

[0056] Mechanism of action: By using the coordinate-physical quantity mapping relationship of the time-frequency diagram, the geometric information of the detection box is converted into the frequency and time parameters of the frequency hopping signal, thus achieving end-to-end blind estimation.

[0057] Example 3: This invention provides a blind detection method for frequency hopping signal parameters based on deep learning, comprising: The specific implementation process includes the following steps: Time-frequency analysis preprocessing: The S-transform is used to generate the time-frequency diagram of the frequency hopping signal to suppress VLF environmental noise interference; Improved RT-DETR model construction: A hybrid encoder architecture based on the RMT backbone network is constructed to optimize target detection capabilities; Model training and inference: The model is trained and improved using time-frequency graph datasets to achieve frequency hopping signal detection and parameter estimation; Inverse parameter mapping: The frequency and time parameters of the frequency hopping signal are analyzed based on the bounding box coordinates output by the model.

[0058] II. Specific Implementation Steps Step 1: Time-Frequency Graph Generation and Preprocessing Signal Acquisition and Modeling: The input signal is modeled as a superposition of frequency hopping signal, fixed frequency signal, LFM signal, burst signal and AWGN; After the signal is sampled in the time domain, a time-frequency diagram is generated through S-transform, and its mathematical expression is:

[0059] The comparative experiment used STFT to generate time-frequency diagrams (128-point Hamming window) to verify the advantages of S-transform under low signal-to-noise ratio. Time-frequency graph characteristic analysis: Frequency hopping signals appear as short-duration horizontal stripes, LFM signals as diagonal stripes, fixed-frequency signals as continuous horizontal bands, burst signals as short horizontal lines, and noise as discrete points. The time-frequency images were normalized to 640×640 pixels and divided into training and validation sets (512 images for each signal-to-noise ratio, for a total of 7306 images).

[0060] Step 2: Improve the construction of the RT-DETR model Backbone network improvements: The original RT-DETR ResNet-50 was replaced with the RMT (Retentive Network with Manhattan Self-Attention) backbone network; The RMT backbone consists of a four-stage structure: the first three stages use decomposed MaSA (Manhattan Self-Attention), and the fourth stage uses the original MaSA. The MaSA mechanism introduces a two-dimensional spatial prior through the Manhattan distance spatial attenuation matrix, enhancing its focus on local features of the time-frequency graph.

[0061] Hybrid encoder design: Extracting multi-scale features from RMT backbone output ; A hybrid encoder was constructed using the AIFI (Attention-Based Cross-Scale Feature Interaction) module and the CCFF (CNN Cross-Scale Feature Fusion) module; An initial target query is generated by minimizing uncertainty in the query selection strategy, thereby optimizing detection efficiency.

[0062] Decoder and prediction head: The decoder uses a Transformer architecture to iteratively optimize the target query to generate bounding boxes (Bboxes) and class predictions. The prediction head outputs the coordinates (center point x, y, width and height w, h) and confidence level of the frequency hopping signal Bbox without the need for non-maximum suppression (NMS).

[0063] Step 3: Model Training and Inference Training parameter settings: Input image size 640×640, batch size 8, training epochs 150; The optimizer used is AdamW, with an initial learning rate of 1e-3, and the learning rate is adjusted using a cosine annealing strategy. The loss function is a weighted combination of GIoU loss and cross-entropy loss.

[0064] Reasoning process: The input time-frequency graph is used to extract features through the RMT backbone, and the target query is generated through a hybrid encoder; The decoder refines the query layer by layer, outputting the frequency hopping signal Bbox and confidence level; The confidence threshold is set to 0.5 to filter out low-confidence detection results.

[0065] Step 4: Frequency Hopping Parameter Estimation Frequency and time parameter calculation: The center point coordinate y of the Bbox is mapped to the frequency hopping frequency:

[0066] ( For the time-frequency plot height, (for maximum frequency); Bbox height h is mapped to frequency hopping duration:

[0067] ( (Total signal duration).

[0068] Error correction: The sliding window method is used to align adjacent frequency hopping bands, eliminating Bbox breakage or overlap errors caused by noise.

[0069] III. Explanation of Key Improvement Points Advantages of S-transform: Compared to STFT, the time-frequency resolution adaptive characteristics of S-transform (high frequency resolution at low frequencies and high time resolution at high frequencies) are better suited to the low signal-to-noise ratio conditions of VLF environments.

[0070] RMT backbone network: The MaSA mechanism enhances local feature extraction and expands the receptive field by using the Manhattan distance spatial decay matrix; Decompositional MaSA reduces computational complexity and balances the ability to model global and local information.

[0071] End-to-end detection architecture: Hybrid encoders fuse multi-scale features to improve the detection accuracy of small targets (short-time frequency hopping band); Uncertainty minimization query selection strategies reduce redundant detections and improve model robustness.

[0072] IV. Implementation Results Verification Simulation experiments were conducted to verify: S-conversion performance: When SNR < -10dB, the AP50 of the S-conversion is 15%-20% higher than that of the STFT; Model comparison: The improved RT-DETR AP50-95 achieves 0.73 at SNR=-12dB, which is better than YOLOv8 / YOLOv9 (which drops below 0.5). • Parameter estimation accuracy: frequency hopping frequency RMSE < 30Hz (SNR ≥ -6dB), time error < 0.5ms.

[0073] In another embodiment of the present invention, a blind detection system for frequency hopping signal parameters based on deep learning is provided, which can be used to implement the above-mentioned blind detection method for frequency hopping signal parameters based on deep learning. Specifically, the system includes: The data acquisition module is used to perform S-transform time-frequency analysis on the received mixed signal, generate a time-frequency diagram to suppress very low frequency interference signals and Gaussian white noise; A hybrid encoder architecture building module is used to build a hybrid encoder architecture based on the RMT backbone network and extract multi-scale time-frequency features. The training module is used to input the time-frequency map into the hybrid encoder architecture, extract multi-level features through the RMT backbone network, generate target queries through the hybrid encoder, and output the bounding box coordinates of the frequency hopping signal by the decoder. The output module is used to calculate the frequency hopping frequency and transition time based on the physical parameters of the frequency hopping signal mapped from the bounding box coordinates, and to complete the robust estimation of the frequency hopping frequency and transition time.

[0074] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0075] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a blind detection method for frequency hopping signal parameters based on deep learning.

[0076] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the deep learning-based blind detection method for frequency hopping signal parameters in the above embodiments.

[0077] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0078] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0079] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0080] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A blind detection method for frequency hopping signal parameters based on deep learning, characterized in that, include: S-transform time-frequency analysis is performed on the received mixed signal to generate a time-frequency plot to suppress very low frequency interference signals and Gaussian white noise; A hybrid encoder architecture is constructed based on the RMT backbone network to extract multi-scale time-frequency features; The time-frequency map is input into the hybrid encoder architecture, multi-level features are extracted through the RMT backbone network, the target query is generated by the hybrid encoder, and the bounding box coordinates of the frequency hopping signal are output by the decoder. Based on the physical parameters of the frequency-hopping signal mapped from the bounding box coordinates, the frequency hopping frequency and hopping time are calculated, and a robust estimation of the frequency hopping frequency and hopping time is completed.

2. The method for blind detection of frequency hopping signal parameters based on deep learning according to claim 1, characterized in that, The step of performing S-transform time-frequency analysis on the received mixed signal to generate a time-frequency plot to suppress very low frequency interference signals and Gaussian white noise includes: The received signal is denoised and normalized. S-transform calculation: Using Gaussian window function , where the scale factor To achieve frequency-adaptive time-frequency resolution, where For the original signal input, For time variables, For frequency, The S-transform representation of the signal; Through formula Generate a time-frequency graph and suppress cross-term interference.

3. The method for blind detection of frequency hopping signal parameters based on deep learning according to claim 1, characterized in that, The hybrid encoder architecture built on the RMT backbone network includes: RMT backbone network replacement: The original RT-DETR ResNet-50 backbone is replaced with an RMT network, whose core is the Manhattan self-attention MaSA mechanism. Hybrid encoder design: Extracting multi-scale features from RMT backbone output ; A hybrid encoder is constructed using the attention-based cross-scale feature interaction module AIFI and the CNN cross-scale feature fusion module CCFF; An initial target query is generated by minimizing uncertainty in the query selection strategy, thereby optimizing detection efficiency. The decoder uses a Transformer architecture to generate bounding boxes (Bboxes) and class predictions through iterative optimization of the target query; the prediction head outputs the coordinates and confidence scores of the frequency-hopping signals Bboxes.

4. The method for blind detection of frequency hopping signal parameters based on deep learning according to claim 3, characterized in that, Introducing a spatial decay matrix based on Manhattan distance Construct a priori knowledge in two-dimensional space; The self-attention weight is calculated as follows: ,in For querying key vectors, For dimensions.

5. The method for blind detection of frequency hopping signal parameters based on deep learning according to claim 4, characterized in that, An improved model is trained using a time-frequency graph dataset to achieve frequency hopping signal detection and parameter estimation. Input image size, batch size, and number of training epochs; The optimizer used is AdamW, with an initial learning rate of 1e-3, and the learning rate is adjusted using a cosine annealing strategy. The loss function is a weighted combination of GIoU loss and cross-entropy loss; The input time-frequency graph is used to extract features through the RMT backbone, and the target query is generated through a hybrid encoder; The decoder refines the query layer by layer, outputting the frequency hopping signal Bbox and confidence level; Set a confidence threshold to filter out low-confidence detection results.

6. The method for blind detection of frequency hopping signal parameters based on deep learning according to claim 1, characterized in that, The process of inputting the time-frequency map into the hybrid encoder architecture, extracting multi-level features through the RMT backbone network, generating a target query through the hybrid encoder, and outputting the bounding box coordinates of the frequency-hopping signal by the decoder includes: Model output bounding box coordinates ,in The coordinates of the top left corner The coordinates are the bottom right corner.

7. The method for blind detection of frequency hopping signal parameters based on deep learning according to claim 1, characterized in that, The step of calculating the frequency hopping frequency and transition time based on the physical parameters of the frequency hopping signal mapped from the bounding box coordinates, and completing the robust estimation of the frequency hopping frequency and transition time, includes: Based on the mapping relationship of the vertical axis of the time-frequency diagram, the frequency hopping frequency... The calculation is as follows: in The maximum frequency in the spectrum. The ordinate of the bounding box center. The height of the time-frequency graph: Jump time estimation: Jump cycle Obtained by mapping the horizontal span of the bounding box: in The total duration of signal observation. This represents the width of the time-frequency graph.

8. A blind detection system for frequency hopping signal parameters based on deep learning, characterized in that, include: The data acquisition module is used to perform S-transform time-frequency analysis on the received mixed signal, generate a time-frequency diagram to suppress very low frequency interference signals and Gaussian white noise; A hybrid encoder architecture building module is used to build a hybrid encoder architecture based on the RMT backbone network and extract multi-scale time-frequency features. The training module is used to input the time-frequency map into the hybrid encoder architecture, extract multi-level features through the RMT backbone network, generate target queries through the hybrid encoder, and output the bounding box coordinates of the frequency hopping signal by the decoder. The output module is used to calculate the frequency hopping frequency and transition time based on the physical parameters of the frequency hopping signal mapped from the bounding box coordinates, and to complete the robust estimation of the frequency hopping frequency and transition time.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the blind detection method for frequency hopping signal parameters based on deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the blind detection method for frequency hopping signal parameters based on deep learning as described in any one of claims 1 to 7.

Images