Multi-band electromagnetic spectrum map intelligent fusion method and system

Abstract

The present disclosure discloses a multi-band electromagnetic spectrum map intelligent fusion method and system, the method combines various deep learning models such as graph neural network, ResNet-18 network, bidirectional LSTM and deep Q network agent to form a comprehensive intelligent processing system; by fusing time-frequency features and time sequence features, and using graph nodes and interference coupling matrix to assign weights to these features, a joint interference feature vector is generated; using the deep Q network agent to generate filter parameters according to the joint interference feature vector, and dynamically adjusting the parameters of the digital FIR filter to suppress the interference component. The present disclosure can generate more accurate and detailed spectrum map, improve the utilization efficiency and management ability of spectrum resources, dynamically adjust the filter parameters to adapt to different spectrum environment, effectively suppress the interference component, and improve the performance and reliability of the communication system.

Description

Technical Field

[0001] This disclosure relates to the field of electromagnetic spectrum signal processing technology, and in particular to an intelligent fusion method and system for electromagnetic spectrum maps based on multi-band collaborative sensing, graph neural networks and reinforcement learning. Background Technology

[0002] With the rapid development of wireless communication technology, multi-band collaborative utilization has become a core means to improve spectrum efficiency. However, existing technologies face severe challenges in the field of multi-band electromagnetic spectrum management and interference suppression: First, spectrum sensing is fragmented, and traditional single-band analysis methods struggle to capture the complex interference coupling characteristics between frequency bands, resulting in low accuracy of spectrum map fusion. This is especially problematic in dense electromagnetic environments, where it easily leads to misjudgments of spectrum holes, severely restricting spectrum resource utilization. Second, interference suppression methods are passive and rigid. Existing technologies often rely on fixed filtering parameters or rule-based threshold judgments, making it difficult to adapt to interference changes in dynamic electromagnetic environments. This results in drastic fluctuations in bit error rate and signal-to-noise ratio, seriously affecting communication reliability. Furthermore, model deployment costs are high. Current deep learning-based spectrum analysis models have a large number of parameters and high computational complexity, making real-time processing on edge devices difficult. Model updates also require frequent cloud transmission, resulting in significant bandwidth and energy consumption. Although recent research has attempted to reduce complexity through model compression, existing lightweight methods often lead to a significant decrease in feature representation capabilities, failing to balance performance and efficiency. Therefore, there is an urgent need for a system that can achieve multi-band intelligent collaborative sensing, dynamic interference suppression, and lightweight edge deployment to overcome the technical bottlenecks of spectrum resource fragmentation, dynamic interference, and high deployment costs, and meet the urgent needs of 5G / 6G communication, military electronic warfare, and the Internet of Things for high-reliability, low-latency spectrum management. Summary of the Invention

[0003] In view of this, embodiments of this disclosure provide a method and system for intelligent fusion of multi-band electromagnetic spectrum maps. Through multi-band collaborative sensing, graph neural network modeling, reinforcement learning dynamic strategy optimization, and model lightweighting technology, it solves the technical problems of spectrum resource fragmentation, dynamic interference, and high deployment costs.

[0004] Specifically, the first embodiment of this disclosure discloses a method for intelligent fusion of multi-band electromagnetic spectrum maps, characterized by the following steps: Simultaneously acquire time-domain signals from the first frequency band and the second frequency band to construct an original signal set that is spatiotemporally consistent with the time-domain signals from the first frequency band and the second frequency band. A first time-frequency graph is generated by performing short-time Fourier transform on the first frequency band time-domain signal and the second frequency band time-domain signal in the original signal set. The time-frequency graph is then denoised and standardized to obtain a second time-frequency graph. A graph neural network is constructed, and the second time-frequency graph is used as the input of the graph neural network to obtain graph nodes and interference coupling matrix; Input the second time-frequency graph into the ResNet-18 network to extract time-frequency features, input the segmented signals of the original spatiotemporally consistent signal set into the bidirectional LSTM to extract time-series features; assign weights to the time-frequency features and time-series features according to the graph nodes and interference coupling matrix to generate a joint interference feature vector.

[0005] A deep Q-network agent is constructed, with the joint interference feature vector serving as the state space of the deep Q-network agent and the action space generated by the deep Q-network agent serving as the filter parameters; the digital FIR filter parameters are dynamically adjusted based on the filter parameters generated by the deep Q-network agent to suppress interference components. Based on the gradient magnitude criterion, the graph neural network, ResNet-18 network, bidirectional LSTM and deep Q network proxies are compressed, and convolutional kernels with lower gradient flow rankings are removed.

[0006] Furthermore, the original signal set used to construct the spatiotemporal consistency between the first frequency band time-domain signal and the second frequency band time-domain signal further includes: The first frequency band time domain signal is a Sub-6GHz frequency band time domain signal, and the second frequency band time domain signal is a millimeter wave frequency band time domain signal; the spatiotemporal consistency of the Sub-6GHz frequency band time domain signal and the millimeter wave frequency band time domain signal is ensured by GPS and timestamp alignment algorithm.

[0007] Furthermore, the step of performing short-time Fourier transform on the first frequency band time-domain signal and the second frequency band time-domain signal in the original signal set to generate a first time-frequency graph, and then performing denoising and standardization processing on the time-frequency graph to obtain a second time-frequency graph specifically includes: Short-time Fourier transform is performed on the time-domain signals of the first and second frequency bands in the original signal set. The short-time Fourier transform uses a Hamming window function with a window length of 256 points, and the resulting first time-frequency graph has a size of 3×224×224. High-frequency noise is eliminated using a wavelet thresholding denoising algorithm, with the noise standard deviation... The wavelet threshold is calculated using the background noise segment of the original signal set:

[0008] Where N is the signal length; The second time-frequency graph is obtained by performing Z-score normalization on the first time-frequency graph. The Z-score normalization formula is as follows:

[0009] Where μ is the mean energy of the first time-frequency graph, and σ is the noise standard deviation.

[0010] Furthermore, the construction of the graph neural network, using the second time-frequency graph as input to the graph neural network to obtain graph nodes and interference coupling matrix specifically includes: A three-layer graph convolutional network is constructed, and the second time-frequency graph is input into the three-layer graph convolutional network to obtain a graph-level representation with the first frequency band time-frequency signal and the second frequency band time-frequency signal as graph nodes; The interference coupling matrix is ​​calculated using the time-frequency signals of the first and second frequency bands as graph nodes. The interference coupling matrix serves as the edge weights of the graph, dynamically reflecting the interference intensity between frequency bands. The interference coupling matrix:

[0011] in, and They are signals from different frequency bands. It is covariance. The standard deviation is denoted as .

[0012] Furthermore, inputting the second time-frequency graph into the ResNet-18 network to extract time-frequency features, and inputting the segmented signals of the spatiotemporally consistent original signal set into a bidirectional LSTM to extract time-series features specifically includes: The second time-frequency image is input into the ResNet-18 network to extract time-frequency features. The last layer of the ResNet-18 network uses global average pooling to output 512-dimensional time-frequency features, as shown in the formula:

[0013] GAP stands for Global Average Pooling. F The second time-frequency graph represents the input, and the second time-frequency graph is a three-dimensional tensor, represented as... C × H × W ,in C It is the number of channels. H It is the height of the feature map. W It is the width of the feature map; H The height of the second time-frequency map is represented by , and the dimension of the feature map in the vertical direction is represented by . W The width of the second time-frequency map is represented by , and the dimension of the feature map in the horizontal direction is represented by . This indicates that the second time-frequency graph is at position ( i , j The value at () here i It is an index in the height direction. j It is an index along the width direction; The segmented signals of the original spatiotemporally consistent signal set are input to the bidirectional LSTM to extract temporal features. The size of the bidirectional LSTM hidden layer is 128, and it outputs 128-dimensional temporal features.

[0014] Furthermore, the process of assigning weights to the time-frequency features and time-series features based on the graph nodes and the interference coupling matrix to generate a joint interference feature vector specifically includes: The query vector Q set and key vector K set of the 4-head attention mechanism are determined based on the graph nodes and the interference coupling matrix. The weights of time-frequency features and time-series features are dynamically allocated through the four-head attention mechanism, and the weight calculation formula is as follows:

[0015] Where Softmax is the activation function. The dimension of the key vector; For the first i A query vector with attention heads. No. i The key vector of each attention head, where T represents the transpose matrix.

[0016] Based on the weights, a joint interference feature vector is generated from the time-frequency features and the time-series features.

[0017] Furthermore, a deep Q-network agent is constructed, wherein the joint interference feature vector serves as the state space of the deep Q-network agent, and the action space generated by the deep Q-network agent serves as the filter parameters, specifically including: Construct a deep Q-network agent, the state space of which is the 768-dimensional joint interference feature vector, and the action space generated by the deep Q-network agent is used as filter parameters; Define the reward function:

[0018] Where ΔSNR is the signal-to-noise ratio improvement and ΔBER is the bit error rate reduction; During the training of the deep Q-network agent, an experience replay pool with a capacity of 100,000 is used to stabilize the learning process; an ε-greedy policy with ε of 0.1 is used to balance exploration and exploitation; and a reward function with a discount factor γ of 0.99 is set for long-term reward, and training continues until the policy converges.

[0019] Furthermore, dynamically adjusting the digital FIR filter parameters to suppress interference components based on the filter parameters generated by the deep Q-network proxy specifically includes: Obtain the cutoff frequency adjustment Δf output by the deep Q-network proxy, and update the cutoff frequency, wherein the formula for updating the cutoff frequency is:

[0020] The parameters of the digital FIR filter are calculated using the Hanning window function, which is defined as follows:

[0021] The formula for dynamically adjusting the parameters of a digital FIR filter is: .

[0022] Furthermore, the aforementioned model compression of the graph neural network, ResNet-18 network, bidirectional LSTM, and deep Q-network proxies based on the gradient magnitude criterion, removing convolutional kernels with lower gradient flow ranking, specifically includes: Gradient magnitude pruning is performed by calculating the absolute value of the convolution kernel gradient for each neural network. The absolute value of the convolution kernel gradient is:

[0023] L represents the model's loss function. This represents the convolution kernel parameters in the graph neural network, the two-branch deep learning network, and the deep Q-network agent; After removing 30% of the convolutional kernels after gradient ranking, the compressed model parameters are quantized to 8-bit integers using a scaling factor and quantization formula, and then the quantized model is deployed. The scaling factor is calculated as follows:

[0024] The quantification formula is: .

[0025] Furthermore, the method further includes: Incremental learning is carried out using an elastic weight solidification strategy, retaining the weights of key parameters and periodically evaluating system performance. If the bit error rate rises above the threshold, model retraining is triggered. The loss function is defined as follows:

[0026] in, Let L be the new loss function, and L be the original loss function. Here is the Fisher information matrix, λ=1000. These are the historically optimal parameters.

[0027] The second embodiment of this disclosure discloses a multi-band electromagnetic spectrum map intelligent fusion system, which includes the following units: The acquisition unit is used to synchronously acquire time-domain signals of the first frequency band and the second frequency band, and construct an original signal set with spatiotemporal consistency between the time-domain signals of the first frequency band and the time-domain signals of the second frequency band; The transformation unit is used to perform short-time Fourier transform on the first frequency band time-domain signal and the second frequency band time-domain signal in the original signal set to generate a first time-frequency graph, and to perform denoising and standardization processing on the first time-frequency graph to obtain a second time-frequency graph; A construction unit is used to construct a graph neural network, taking the second time-frequency graph as the input of the graph neural network to obtain graph nodes and an interference coupling matrix; The extraction unit is used to input the second time-frequency graph into the ResNet-18 network to extract time-frequency features, input the segmented signals of the spatiotemporally consistent original signal set into the bidirectional LSTM to extract time-series features, and assign weights to the time-frequency features and time-series features according to the graph nodes and the interference coupling matrix to generate a joint interference feature vector. A generation unit is used to construct a deep Q-network agent, wherein the joint interference feature vector serves as the state space of the deep Q-network agent, and the action space generated by the deep Q-network agent serves as the filter parameters; the digital FIR filter parameters are dynamically adjusted according to the filter parameters generated by the deep Q-network agent to suppress interference components. The compression unit is used to compress the graph neural network, ResNet-18 network, bidirectional LSTM and deep Q network proxies based on the gradient magnitude criterion, and remove convolutional kernels with lower gradient flow ranking.

[0028] This disclosure combines multiple deep learning models, including graph neural networks, ResNet-18 networks, bidirectional LSTM, and deep Q-network proxies, to form a comprehensive intelligent processing system. It generates a joint interference feature vector by fusing time-frequency and time-series features and assigning weights to these features using graph nodes and an interference coupling matrix. The deep Q-network proxies then generate filter parameters based on this joint interference feature vector and dynamically adjust the parameters of the digital FIR filter to suppress interference components. This disclosure can generate more accurate and detailed spectrum maps, improve the utilization efficiency and management capabilities of spectrum resources, dynamically adjust filter parameters to adapt to different spectrum environments, effectively suppress interference components, and improve the performance and reliability of communication systems.

[0029] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A flowchart of the intelligent fusion method for multi-band electromagnetic spectrum maps provided in this embodiment of the disclosure.

[0032] Figure 2 The flowchart illustrates a method for generating a first time-frequency graph by performing a short-time Fourier transform on an original signal set, and then performing denoising and standardization on the time-frequency graph to obtain a second time-frequency graph, as provided in this embodiment of the disclosure.

[0033] Figure 3 The flowchart below illustrates the method for extracting time-frequency and time-series features, determining the query vector Qi and key vector Ki of the 4-head attention mechanism, assigning weights to the time-frequency and time-series features, and generating a joint interference feature vector, as provided in this embodiment of the disclosure.

[0034] Figure 4 A schematic diagram of the data and neural network process of the intelligent fusion method for multi-band electromagnetic spectrum maps provided in this embodiment of the disclosure.

[0035] Figure 5 This disclosure provides a schematic diagram of the structure of a computer device according to an embodiment. Detailed Implementation

[0036] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0037] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0038] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0039] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0040] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0041] For detailed implementation procedures, please refer to the appendix. Figure 1 The hardware platform of this invention employs a Xilinx Zynq UltraScale+ FPGA (supporting 16-channel parallel processing) and an Intel Xeon server (for model training). The signal source covers a mixture of Sub-6GHz (3.5GHz, 100MHz bandwidth) and millimeter-wave (28GHz, 800MHz bandwidth) signals to simulate multi-band intermodulation interference, adjacent-channel interference, and impulse noise interference scenarios, dynamically adjusting the interference intensity (SNR range 0-15dB). Test data acquisition is synchronized via a GPS module to ensure spatiotemporal consistency, with a synchronization error of less than 1μs and a timestamp alignment error of less than 0.5ms.

[0042] A method for intelligent fusion of multi-band electromagnetic spectrum maps, characterized by the following steps: Step S100: Multi-band spectrum data collaborative acquisition, synchronously acquire Sub-6GHz and millimeter wave band time domain signals, and use GPS module and timestamp alignment algorithm to ensure spatiotemporal consistency of Sub-6GHz and millimeter wave band time domain signals to obtain the original signal.

[0043] Select a spectrum analyzer or Software-Defined Radio (SDR) device with wide bandwidth, high sampling rate, and high dynamic range to acquire Sub-6 GHz signals. For example, an SDR platform based on USRP (Universal Software Radio Peripheral) can be used, whose operating frequency band can cover the main frequency bands of Sub-6 GHz, such as the low-frequency band below 3 GHz and the mid-frequency band of 3 GHz to 6 GHz. During deployment, place the SDR device's antenna in key locations within the target monitoring area, such as near communication base stations, radar coverage areas, or around potential sources of spectrum interference, to ensure effective reception of electromagnetic signals within this frequency band.

[0044] For millimeter-wave frequency bands (typically 24GHz - 100GHz), select a dedicated millimeter-wave spectrum analyzer or millimeter-wave SDR equipment. These devices must possess high-resolution frequency measurement capabilities and sufficient bandwidth to capture the detailed characteristics of millimeter-wave signals. Considering the propagation characteristics of millimeter-wave signals (such as high attenuation and short-range transmission), the acquisition equipment should be placed near the millimeter-wave transmitter or the expected millimeter-wave communication link, such as the coverage area of ​​a 5G millimeter-wave base station or the test site of a millimeter-wave automotive radar.

[0045] Equip each signal acquisition device with a high-precision GPS receiver module. The GPS module should possess rapid positioning, high-precision time synchronization (time accuracy at least in the microsecond range), and good anti-interference capabilities. Install the GPS antenna near the signal acquisition device to ensure it can clearly receive positioning and time synchronization signals from satellites, thereby providing accurate time and location information for the acquired signals. Configure a high-speed, high-capacity data transmission and storage system, such as a high-speed Ethernet interface, fiber optic transmission link, or large-capacity hard disk array. Due to the high sampling rate of signals in the Sub-6GHz and millimeter-wave bands, the generated data volume is large; it is necessary to ensure that the data can be transmitted to the storage device in real time and stably to avoid data loss or congestion.

[0046] Synchronization start commands are sent simultaneously to signal acquisition devices in the Sub-6GHz and millimeter-wave bands via host computer software or a dedicated control unit. These commands, based on preset trigger conditions (such as specific time points or external event triggers), ensure that the acquisition devices in both bands begin signal acquisition at the same starting point. For example, when monitoring spectrum usage in a 5G communication system, signal acquisition in the Sub-6GHz (e.g., 5G NR in the 3.5GHz band) and millimeter-wave (e.g., 5G NR in the 28GHz band) bands can be started synchronously at the beginning of a predetermined scheduling cycle of the 5G base station. During acquisition, the local clocks of both acquisition devices are calibrated in real time using a precise clock signal provided by a GPS module. Since the time and frequency references of GPS signals are consistent globally, synchronization with the GPS clock ensures a high degree of consistency in the sampling clocks of the Sub-6GHz and millimeter-wave acquisition devices, allowing the acquired signals to be accurately aligned on the time axis. Typically, the clock synchronization accuracy needs to be periodically checked and calibrated to ensure it remains within the specified time accuracy range (e.g., ±1 microsecond).

[0047] Each acquired signal data frame is marked with a corresponding timestamp and GPS location information. The timestamp should be accurate to the millisecond level or higher to ensure accurate determination of the signal acquisition time in subsequent data processing. The location information identifies the geographic coordinates of the signal acquisition, which is crucial for analyzing the spatial distribution characteristics of the spectrum. The marked signal data are stored in their respective data storage devices, and corresponding data indexes are established for rapid retrieval and processing later.

[0048] The system collects timestamp data recorded by acquisition devices in the Sub-6GHz and millimeter-wave bands, and calculates the time offset between the two bands based on the timestamp differences. Interpolation algorithms (such as linear interpolation, cubic spline interpolation, etc.) are then used to adjust the time axis of one band's signal to align it with the other band's signal in time. For example, if the timestamp of the millimeter-wave signal is 0.5 milliseconds later than that of the Sub-6GHz signal, the time axis of the millimeter-wave signal can be adjusted forward by 0.5 milliseconds using an interpolation algorithm, ensuring that both bands have corresponding sampled values ​​at the same time point.

[0049] Using the GPS location information of the acquisition equipment, the spatial relationship between the acquisition points for Sub-6GHz and millimeter-wave signals is determined. Considering the signal propagation speed (speed of light) and the relative positions between the acquisition points, the Time Difference of Arrival (TDOA) due to spatial location differences is calculated. Further time calibration of the millimeter-wave or Sub-6GHz signals is performed based on the TDOA to eliminate time deviations introduced by inconsistent spatial distribution of the acquisition equipment. Simultaneously, spatial interpolation or extrapolation algorithms can be used to spatially calibrate the spectral data, ensuring consistency in spatial resolution between the two frequency bands, thereby achieving spatiotemporal consistency.

[0050] Step S200: Perform a short-time Fourier transform (STFT) on the Sub-6GHz and millimeter-wave frequency band time-domain signals in the original signal set to generate a first time-frequency graph. Then, perform denoising and normalization processing on the time-frequency graph to obtain a second time-frequency graph. For detailed implementation methods, see [link to implementation details]. Figure 2 As shown.

[0051] Step S2001: Perform short-time Fourier transform (STFT) on the Sub-6GHz and millimeter-wave band time-domain signals in the original signal set.

[0052] A Hamming window function with a length of 256 points is selected. The Hamming window function effectively reduces spectral leakage and improves the frequency resolution of the STFT, making it suitable for segmented signal processing, allowing for more accurate extraction of the spectral characteristics of each segment. The Sub-6GHz and millimeter-wave band time-domain signals from the original signal set are framed with a certain frame shift (usually half the window length, i.e., 128 points, to ensure sufficient overlap between adjacent frames and avoid signal feature loss), resulting in each frame having a length of 256 points and 128 points of overlap between adjacent frames. Each frame is multiplied by the Hamming window function and then subjected to a Fast Fourier Transform (FFT). The number of FFT points can be set to the same as the window length (256 points) or larger (e.g., 512 points) to improve frequency resolution. The FFT converts the time-domain signal to a frequency-domain signal, obtaining the spectral information of each frame, including the amplitude spectrum and phase spectrum.

[0053] Step S2002: Arrange the amplitude spectrum of each frame signal in chronological order to form a two-dimensional matrix, namely the first time-frequency diagram.

[0054] The time-frequency plot has a size of 3×224×224, where 3 may represent different channels (such as the results of processing the Sub-6GHz and millimeter-wave bands separately, as well as a composite channel) or different feature dimensions, and 224×224 represents the resolution of the frequency domain and time domain. The specific value depends on factors such as the signal sampling rate, frame shift, and number of FFT points.

[0055] Step S2003: Use wavelet threshold denoising algorithm to eliminate high-frequency noise.

[0056] Select a relatively quiet background noise segment from the original signal set, which has no obvious useful signal. This segment should be representative and able to reflect the noise level of the signal under interference-free conditions. Usually, the starting segment or some stable segments in the middle of the signal are selected as the background noise segment.

[0057] Noise standard deviation The wavelet threshold is calculated using the background noise segment of the original signal set:

[0058] Where N is the signal length; Using the selected background noise segment, calculate its standard deviation σ, the formula is as follows: σ, where var(noise segment) represents the variance of the noise segment. This σ value will serve as an important parameter in the wavelet thresholding denoising algorithm, used to determine the threshold for the wavelet coefficients.

[0059] Wavelet decomposition is performed on the first time-frequency graph. Typically, a suitable wavelet basis function (such as Daubechies wavelet, Symlets wavelet, etc.) and a decomposition level (such as 3-5 levels) are chosen to decompose the time-frequency graph into wavelet coefficients of different scales. Then, according to the formula... Calculate the wavelet threshold T, where N is the signal length (for a two-dimensional time-frequency graph, the number of rows or columns can be used as N). Perform soft thresholding (or hard thresholding) on ​​the decomposed wavelet coefficients, setting wavelet coefficients smaller than the threshold T to zero and subtracting the threshold T from wavelet coefficients greater than or equal to T (soft thresholding), thereby removing high-frequency noise components. Reconstruct the thresholded wavelet coefficients using wavelet reconstruction to obtain the denoised time-frequency graph. This denoised time-frequency graph retains the main features of the original signal while reducing noise interference.

[0060] Step S2004: Standardize the denoised time-frequency graph to obtain the second time-frequency graph.

[0061] Calculate the mean energy of all elements in the first time-frequency plot (the denoised time-frequency plot). ,Right now ,in This represents the amplitude of each element in the time-frequency graph, where N is the total number of elements in the time-frequency graph.

[0062] The second time-frequency graph is obtained by performing Z-score normalization on the first time-frequency graph. The Z-score normalization formula is as follows:

[0063] Where μ is the mean energy of the first time-frequency graph, and σ is the noise standard deviation.

[0064] All the standardized elements are recombined into a matrix of the same size as the original time-frequency graph, resulting in the second time-frequency graph. This second time-frequency graph has undergone denoising and standardization, preparing it for subsequent graph neural network input.

[0065] Suppose that in a 5G communication environment monitoring project, electromagnetic signals in the Sub-6GHz (3.5GHz band) and millimeter wave (28GHz band) bands of a certain area are processed and analyzed.

[0066] In step S100, time-domain signals in the Sub-6GHz and millimeter-wave bands of the region are acquired. These signals have been synchronized and have spatiotemporal consistency. They are stored as an original signal set with a sampling rate of 100MHz and a signal length of 100,000 points.

[0067] A Hamming window function with a window length of 256 points and a frame shift of 128 points was selected to perform a short-time Fourier transform (STFT) to ensure the continuity and integrity of the signal characteristics.

[0068] The original signal is segmented into frames according to the parameters described above, resulting in multiple signal frames. Each signal frame is multiplied by a Hamming window function, and then a 512-point FFT is performed to obtain the spectral information of each frame. This improves the frequency domain resolution and allows for more detailed analysis of the signal's frequency components.

[0069] The amplitude spectra of each frame of the signal in the Sub-6GHz and millimeter-wave bands are arranged in chronological order to form two time-frequency plot matrices. Simultaneously, the time-frequency plots of the two bands are combined (e.g., by averaging or weighted summation) to obtain a combined time-frequency plot matrix. These three matrices are then combined into a first time-frequency plot of size 3×224×224. The 224×224 size is determined based on factors such as the signal sampling rate, frame shift, number of FFT points, and signal duration, which meets the input data size requirements of deep learning models.

[0070] The first 10,000 or so signal segments from the original signal set were selected as the background noise segment. This segment was collected when there was no significant communication service and included the main environmental noise and equipment noise.

[0071] Calculate the standard deviation σ of the background noise segment, assuming that the calculated σ = 0.05 (the unit is determined according to the signal amplitude quantization method).

[0072] The Daubechies wavelet (db4) is selected to perform a 3-level wavelet decomposition on the first time-frequency plot. Assuming the signal length N = 224 (the number of columns in the time-frequency plot), the wavelet threshold T is calculated as: T = 0.05 × √(2 × log₂24) ≈ 0.05 × √(2 × 7.8) ≈ 0.05 × √15.6 ≈ 0.05 × 3.95 ≈ 0.1975. Soft thresholding is then applied to the decomposed wavelet coefficients, setting those less than 0.1975 to zero and subtracting 0.1975 from the remaining coefficients.

[0073] The processed wavelet coefficients are reconstructed using wavelet reconstruction to obtain the denoised time-frequency graph. At this time, the noise components in the time-frequency graph are effectively suppressed, and the main features of the signal are clearer.

[0074] Calculate the mean energy of all elements in the denoised time-frequency plot. Assuming the calculation yields =0.2 (units as above). According to the formula Standardize each element in the time-frequency graph to obtain standardized time-frequency graph element values. The mean of the time-frequency graph is 0 and the standard deviation is 1, resulting in a more standardized data distribution.

[0075] The standardized elements are recombined into a matrix of size 3×224×224 to obtain the second time-frequency graph. This time-frequency graph has undergone denoising and standardization and can be directly used as input data for subsequent graph neural networks for further analysis and processing, such as spectral feature extraction and interference identification.

[0076] Step S300: Construct a 3-layer graph neural network (GNN), using the second time-frequency graph as the input to the graph neural network to obtain graph nodes and interference coupling matrix.

[0077] Graph data structures consist of nodes and edges. Nodes represent entities (such as users in a social network or entities in a knowledge graph), and edges represent relationships between entities (such as user follow relationships or entity association relationships). Graph Neural Networks (GNNs) learn node representations by aggregating information from neighboring nodes: each node's features not only include its own attributes but also incorporate features from neighboring nodes; through multi-layer propagation, nodes can capture global structural information from multiple levels of neighbors.

[0078] The structure and function of the 3-layer graph convolutional network are described below. Those skilled in the art will understand that the first layer (input layer) receives the second time-frequency map as input. Since the second time-frequency map is 3×224×224 in size, it needs to be converted to a format suitable for graph neural network input. Typically, each position in each time-frequency map (i.e., each point in 224×224) can be considered a node, and the node's features are 3-channel feature vectors. Therefore, the input layer has 224×224 = 50,176 nodes, and each node has a feature dimension of 3.

[0079] The second layer (hidden layer 1) employs graph convolution operations, using a trainable weight matrix to transform node features and aggregating information from neighboring nodes to update the features of each node. The ReLU activation function can be used to introduce non-linearity. For example, the feature dimension of a node can be reduced from 3 to 64, resulting in a 64-dimensional feature vector for each node.

[0080] The third layer (hidden layer 2) performs graph convolution again to further extract high-level features of the nodes. It also uses the ReLU activation function. This reduces the feature dimension of the nodes from 64 to 128, making the features more abstract and discriminative.

[0081] The fourth layer (output layer) finally maps the node features to the required output dimension through a graph convolutional layer, obtaining the feature representation of the graph nodes. Assuming a 256-dimensional feature representation for each node is required, the output layer increases the node feature dimension from 128 to 256. Simultaneously, a readout function (such as average pooling or max pooling) aggregates the features of all nodes to generate a graph-level representation for subsequent interference analysis.

[0082] The following section explains the calculation of graph nodes and interference coupling matrix. Sub-6GHz and millimeter-wave time-frequency signals are used as nodes in the graph. For example, the Sub-6GHz and millimeter-wave time-frequency graphs can each be considered as a node, or key regions or feature points in each time-frequency graph can be used as nodes.

[0083] According to the formula Calculate the interference coupling matrix, where, and They are signals from different frequency bands. It is covariance. The standard deviation is denoted as . Covariance measures the linear correlation between two signals, reflecting the interference relationship between them. Then, the standard deviation of each signal is calculated. and This is used to normalize the covariance. The elements of the resulting interference coupling matrix C are... The interference intensity between frequency band i and frequency band j is represented, with values ​​ranging from [-1, 1]. A larger absolute value indicates stronger interference. The calculated interference coupling matrix is ​​used as the edge weights of the graph, dynamically reflecting the interference intensity between frequency bands. In graph neural networks, edge weights influence the transmission and aggregation of information between nodes, enabling the network to learn the interference relationships between different frequency bands.

[0084] Mean Squared Error (MSE) is used as the loss function. MSE measures the difference between predicted and true values ​​and is suitable for regression problems. During the training of the Graph Neural Network (GNN), minimizing the MSE loss function enables the network to accurately learn the feature representations and interference coupling matrices of the graph nodes. The Adam optimizer, which combines the advantages of momentum and adaptive learning rate, is used for efficient parameter updates. A learning rate of 0.001 is set, which achieves good convergence in many cases. Through iterative training, the network's weight parameters are continuously adjusted until the network converges, i.e., the loss function no longer decreases significantly.

[0085] Through the above steps, step S300 is completed, constructing a graph neural network capable of processing multi-band electromagnetic spectrum data. This network effectively extracts graph node features and interference coupling matrices, providing a foundation for subsequent interference suppression and spectrum management. The specific parameters and settings in this embodiment can be adjusted and optimized based on actual spectrum data and application scenarios.

[0086] Step S400: Input the second time-frequency graph to the ResNet-18 network to extract time-frequency features; input the segmented signals of the spatiotemporally consistent original signal set to the bidirectional LSTM to extract temporal features; determine the query vector Qi and key vector Ki of the 4-head attention mechanism based on the graph nodes and interference coupling matrix; dynamically allocate the weights of time-frequency features and temporal features through the 4-head attention mechanism to generate a joint interference feature vector. For detailed implementation methods, see [link to implementation details]. Figure 3 As shown.

[0087] Step S4001: Input the second time-frequency graph into the ResNet-18 network to extract time-frequency features.

[0088] The second time-frequency image is converted into an input format suitable for ResNet-18, i.e., a three-dimensional tensor C×H×W. Typically, C is the number of channels (e.g., 3 channels in the second time-frequency image), and H and W are the height and width of the feature map, respectively (e.g., 224×224).

[0089] ResNet-18 consists of multiple residual blocks, each containing two convolutional layers and a shortcut connection. The input data passes through convolutional layers, batch normalization layers, ReLU activation functions, and pooling layers in sequence.

[0090] Convolutional layers extract local features, batch normalization accelerates training and ensures stable convergence, and ReLU introduces non-linearity. Shortcut connections address the vanishing gradient problem in deep networks.

[0091] Through convolution and pooling operations, the network automatically learns features in the time-frequency graph, such as frequency components, temporal distribution, and interference patterns.

[0092] The final layer uses Global Average Pooling (GAP) to scale each feature map to a single value, preserving important information. The GAP layer outputs a 512-dimensional time-frequency feature vector, representing the global features of the second time-frequency map. The formula for Global Average Pooling (GAP) in the final layer is as follows: Each feature map is scaled down to a single value, preserving important information.

[0093] GAP stands for Global Average Pooling. F The second time-frequency graph represents the input, and the second time-frequency graph is a three-dimensional tensor, represented as... C × H × W ,in C It is the number of channels. H It is the height of the feature map. W It is the width of the feature map; H The height of the second time-frequency map is represented by , and the dimension of the feature map in the vertical direction is represented by . W The width of the second time-frequency map is represented by , and the dimension of the feature map in the horizontal direction is represented by . This represents the value of the second time-frequency graph at position (i,j), where i is the index in the height direction and j is the index in the width direction.

[0094] Step S4002: Input the segmented signals of the original spatiotemporally consistent signal set into the bidirectional LSTM to extract the timing features.

[0095] The spatiotemporally consistent original signal set (including time-domain signals in the Sub-6GHz and millimeter-wave bands) acquired in step S100 is segmented and a 10ms window is used for sliding segmentation to obtain multiple segmented signals.

[0096] For each segment of the signal, 128-dimensional Mel-frequency cepstral coefficients (MFCC) features are extracted. MFCCs reflect the spectral characteristics of the signal, and the MFCC feature vector is a set of coefficients that describe the spectral characteristics of the signal. It is obtained by performing spectral analysis and nonlinear transformation on the signal and can reflect the spectral envelope information of the signal.

[0097] A bidirectional LSTM consists of forward and backward LSTM layers. The forward layer processes the segmented signal in chronological order, while the backward layer processes it in reverse order, outputting the hidden states respectively. The bidirectional LSTM extracts the long-term temporal features of the segmented signal and captures the dynamic changes of interference. Finally, the hidden states from both directions are concatenated to obtain a 128-dimensional temporal feature.

[0098] Specifically, for the original signal set with spatiotemporal consistency, the sampling rate is 1000 Hz and the signal duration is 1 second, resulting in a total of 1000 sampling points. This is divided into 10 segments, each with a length of 100 sampling points (corresponding to 0.1 seconds), and a 128-dimensional MFCC feature vector is extracted from each segment.

[0099] The feature matrix input to the bidirectional LSTM has a shape of [10, 128], where 10 is the number of segments and 128 is the MFCC feature dimension of each segment.

[0100] The forward LSTM layer starts from the first segment and processes the feature vector of each segment sequentially until the last segment. It captures the temporal information from the beginning to the end of the signal.

[0101] The inverse LSTM layer starts from the last segment and processes the feature vectors of each segment in reverse order until the first segment. It captures the temporal information from the end of the signal to the beginning.

[0102] Assume that the forward LSTM layer outputs a 64-dimensional hidden state at each time step, and the backward LSTM layer also outputs a 64-dimensional hidden state at each time step.

[0103] For each segment, the forward and backward hidden states are concatenated to form a 128-dimensional hidden state vector. The final output temporal feature matrix has a shape of [10, 128], where each segment corresponds to a 128-dimensional temporal feature vector.

[0104] Bidirectional LSTM, through forward and backward processing, can capture long-term dependencies in segmented signals. For example, in electromagnetic spectrum signals, interference may appear in the first half of the signal and affect the second half. Forward LSTM can capture this forward-to-back interference propagation, while backward LSTM can capture the potential back-to-forward impact. The concatenated 128-dimensional time-series feature vector integrates forward and backward information, more comprehensively representing the contextual features of each segment in the time series. This is crucial for subsequent interference analysis, as it helps the model better understand the dynamic changes and temporal patterns of interference.

[0105] Step S4003: Determine the query vector Qi and key vector Ki of the 4-head attention mechanism based on the graph nodes and interference coupling matrix.

[0106] The graph node features and interference coupling matrix obtained in step S300 are used as important information to determine the query vector (Q) and key vector (K) of the 4-head attention mechanism. Among them, the graph node features provide local feature representations of the frequency bands, and the interference coupling matrix reflects the global interference relationship between frequency bands.

[0107] The graph node features output from the graph convolutional network (GCN) in step S300 are a two-dimensional matrix, denoted as... Where R represents a two-dimensional matrix, N is the number of graph nodes, and D is the feature dimension of each node. The interference coupling matrix is ​​denoted as... Each element This represents the interference strength between node i and node j, and D is the feature dimension of each node.

[0108] The graph node feature matrix F and the interference coupling matrix C are fused to form a comprehensive feature matrix. This can be achieved in several ways, such as: directly concatenating the feature vector of each node with its interference coupling strength with other nodes; or performing linear transformations on both the graph node features and the interference coupling matrix, and then adding or concatenating the results.

[0109] The integrated feature matrix is ​​linearly transformed to obtain the fused features, and the query vector Q and key vector K are generated based on the fused features.

[0110] Q= Fusion features, K= Fusion features in, and These are learnable weight matrices, used to generate the query vector and key vector, respectively.

[0111] Step S4004: Dynamically allocate the weights of time-frequency features and time-series features through a 4-head attention mechanism, and generate a joint interference feature vector based on the time-frequency features and time-series features.

[0112] The 4-head attention mechanism is a specific form of the multi-head attention mechanism, which is one of the core components of the Transformer model. The 4-head attention mechanism transforms the query, key, and value in the attention computation using four different fully connected layers, dividing them into four parts, each called a head. Each head focuses on a different part of the input data in a different way, learning different relationships in parallel, thereby obtaining richer contextual information.

[0113] The weights are dynamically allocated using a 4-head attention mechanism. The weights of the query vector Q and the key vector Ki are calculated as follows:

[0114] in, =64. The value represents both time-frequency and time-series characteristics. The outputs of the four attention heads are concatenated and fused through a linear transformation to output a 768-dimensional joint interference feature vector. =[ , , , ].

[0115] Step S500: Construct a deep Q-network agent, where the joint interference feature vector serves as the state space of the deep Q-network agent, and the action space generated by the deep Q-network agent serves as the filter parameters; based on the filter parameters generated by the deep Q-network agent, dynamically adjust the digital FIR filter parameters to suppress interference components.

[0116] Deep Q-Network agents make decisions based on deep Q-Networks. A deep Q-Network is a neural network that takes the state of the environment as input and outputs the expected reward (Q-value) for each possible action in that state. The agent selects actions based on these Q-values, typically choosing the action with the highest Q-value (greedy policy), or selecting other actions with a certain probability (e.g., ε-greedy policy, which randomly selects actions with a certain probability to explore the environment). Simultaneously, the experience (states, actions, rewards, next state, etc.) collected by the agent during its interaction with the environment is used to train the deep Q-Network, enabling it to estimate Q-values ​​more accurately and thus allowing the agent to make better decisions.

[0117] The input layer of a deep Q-network receives a 768-dimensional joint disturbance feature vector as input, representing the state of the current environment.

[0118] The hidden layers of a deep Q-network consist of three layers. First layer: Fully connected layer, with 512 neurons for example, and ReLU activation function.

[0119] The second layer is a fully connected layer with, for example, 256 neurons, and the activation function is ReLU.

[0120] The third layer: a fully connected layer, with for example 64 neurons, and the activation function is ReLU.

[0121] The output layer of a deep Q-network is a fully connected layer, with the number of neurons corresponding to the dimension of the action space. Specifically, it outputs three parameters: cutoff frequency adjustment (Δf, range ±10%), bandwidth adjustment (step size 1MHz), and decay coefficient (0 - 1). Assuming there are 10 selectable step sizes for the bandwidth and 10 selectable values ​​for the decay coefficient, the action space can be represented by different combinations, for example, by quantizing each parameter to obtain a discrete action space.

[0122] The cutoff frequency adjustment (Δf) is discretized into 21 possible values, ranging from -10% to +10%, with a step size of 1%. Assuming it is adjustable within a certain range, such as from 1MHz to 10MHz, with a step size of 1MHz, this corresponds to 10 possible values. Alternatively, it can be discretized into 11 possible values, ranging from 0 to 1, with a step size of 0.1.

[0123] Therefore, the action space has a total of 21 × 10 × 11 = 2310 possible action combinations. The network output layer needs to have a sufficient number of neurons to represent these actions. For example, a three-dimensional output can be used, with each dimension corresponding to a parameter selection, or the action can be encoded as an integer index.

[0124] reward function It is calculated based on the changes in signal-to-noise ratio (SNR) and bit error rate (BER). Here, ΔSNR is the increase in SNR between the current time and the previous time, and ΔBER is the decrease in bit error rate.

[0125] Initialize an experience replay pool with a capacity of 100,000. This pool stores the agent's experience interacting with the environment, including the current state, actions performed, rewards received, and the next state. Set ε to an initial value of 0.1, indicating a 10% probability of randomly selecting an action for exploration and a 90% probability of selecting the action currently evaluated as optimal by the network. Set ε to 0.99 to balance immediate and long-term rewards. Select the Adam optimizer; the learning rate can be initially set to 0.001 and adjusted based on training progress. Use mean squared error (MSE) as the loss function to measure the difference between the predicted Q-value and the actual Q-value.

[0126] Initialize the parameters of the deep Q-network agent. In the simulation environment, the agent selects actions according to the current policy (ε-greedy), observes the new state and reward after executing the action, and stores the experience in the experience replay pool. Randomly sample mini-batch experiences (e.g., 32 or 64) from the experience replay pool and use these experiences to update the parameters of the deep Q-network agent. Specifically, calculate the loss between the predicted Q-value and the actual Q-value (based on the maximum Q-value of the reward and the next state), and then update the network parameters through backpropagation. Periodically (e.g., every 1000 steps), update the parameters of the target network to gradually approach the parameters of the deep Q-network agent to provide a stable target Q-value. Monitor the network loss and policy performance metrics (e.g., average reward), and consider training converged when the loss stabilizes and policy performance no longer improves significantly.

[0127] Based on the filter parameters generated by the deep Q-network proxy, the digital FIR filter parameters are dynamically adjusted to suppress interference components. The update cutoff frequency of the filter parameters generated by the deep Q-network proxy is obtained, and the formula for the update cutoff frequency is:

[0128] Where Δf is the adjustment amount (range ±10%) of the deep Q network agent output. The parameters of the digital FIR filter are calculated using the Hanning window function, which is defined as follows:

[0129] The formula for dynamically adjusting the parameters of a digital FIR filter is:

[0130] The dynamically adjusted digital FIR filter supports 16-channel parallel processing, with a single-channel filtering delay of less than 5ms, a total throughput of 3.2Gbps, and suppresses interference components with an interference suppression ratio greater than 20dB.

[0131] Step S600: Based on the gradient magnitude criterion, the graph neural network, ResNet-18 network, bidirectional LSTM and deep Q network proxies are compressed, and convolutional kernels with lower gradient flow rankings are removed.

[0132] Calculate the absolute value of the convolution kernel gradient to perform gradient magnitude pruning. The absolute value of the convolution kernel gradient is:

[0133] L represents the model's loss function, used to measure the difference between the model's predicted output and the true label. The loss function is a crucial part of deep learning model training, guiding the model's optimization direction. A smaller loss function value generally indicates that the model's predictions are closer to the true values. During training, the model learns patterns and features in the data by minimizing the loss function. The parameters represent those in the graph neural network, the two-branch deep learning network, and the deep Q-network agent.

[0134] The absolute gradient values ​​of all convolutional kernels are sorted, and the bottom 30% of convolutional kernels are removed. Then, the compressed model parameters are quantized to 8-bit integers using a scaling factor and a quantization formula.

[0135] The scaling factor is calculated as follows:

[0136] The quantification formula is:

[0137] The quantized model is deployed on the FPGA, and the model size is compressed to 15% of the original size, with a memory usage of less than 50MB.

[0138] Step S700: Incremental learning is performed using an elastic weight solidification strategy, retaining the weights of key parameters and periodically evaluating system performance. If the bit error rate rises above the threshold, model retraining is triggered.

[0139] Incremental learning using a flexible weight fixation strategy is a method that allows neural networks to learn incrementally without completely forgetting previously learned content. Its core idea is to impose a retention constraint on key parameters (i.e., parameters that were highly important in previous tasks) in new learning tasks, preventing them from changing excessively in the new task. In this way, the network can retain previously learned key information while learning new knowledge.

[0140] New loss function It consists of two parts: the original loss function and the regularization term. The original loss function... Used to measure the model's performance on the current task. The regularization term is... :

[0141] It is the Fisher information matrix, used to measure parameters. Importance to previous tasks. If a parameter was important in a previous task, then its corresponding Fisher information matrix element will be relatively large. It is a hyperparameter, set to 1000, used to control the weight of the regularization term and balance the relationship between the original loss and the regularization term. This represents the historical best parameters, i.e., the best parameter values ​​obtained from training on previous tasks. This is the current parameter value.

[0142] The purpose of this regularization term is to penalize parameters that deviate significantly from the historical best parameters, especially those parameters that were important in previous tasks (i.e., parameters with large Fisher information matrix values), thereby preserving key parameters.

[0143] The system periodically (e.g., hourly) evaluates its performance, primarily using the Bit Error Rate (BER) to measure communication system performance. BER is the ratio of erroneous bits transmitted to the total number of bits transmitted, reflecting the accuracy of signal transmission. If the BER exceeds a threshold (10%), model retraining is triggered: if the evaluation finds that the BER exceeds a preset threshold (10% in this case), it indicates that the current model may no longer be well adapted to changes in the environment or task, thus triggering a retraining process. Through retraining, the model can update its parameters to adapt to new data patterns or environmental changes, thereby reducing the BER and improving system performance.

[0144] Global model parameters are synchronized via HTTPS, using the AES-256 encryption algorithm, with an update cycle of one hour: When a model update is needed, the system synchronizes global model parameters via HTTPS. HTTPS is a secure communication protocol based on SSL / TLS, preventing data from being stolen or tampered with during transmission. During transmission, model parameters are encrypted using the AES-256 encryption algorithm. AES-256 is a symmetric encryption algorithm with high security, effectively protecting data confidentiality. The one-hour update cycle means the system synchronizes and updates model parameters every hour to ensure the model's timeliness and accuracy.

[0145] For the convenience of those skilled in the art, Figure 4This is a schematic diagram of the data and neural network process of the intelligent fusion and interference suppression method for multi-band electromagnetic spectrum maps provided in the embodiments of this disclosure.

[0146] This invention proposes a multi-band electromagnetic spectrum map intelligent fusion and interference suppression system. Through innovative technologies such as multi-band collaborative sensing, graph neural network modeling, reinforcement learning dynamic strategy optimization, and lightweight edge deployment, it effectively solves three core problems in existing spectrum management technologies: spectrum resource fragmentation, passive interference suppression, and high deployment costs. The system first achieves high-precision acquisition and preprocessing of multi-band data through GPS synchronization and time-frequency feature extraction, generating a high signal-to-noise ratio (SNR) time-frequency map. Subsequently, based on a graph convolutional network (GCN), it constructs an inter-band interference coupling matrix, dynamically capturing the correlation characteristics of frequency band energy distribution, significantly improving the accuracy of spectrum map fusion and reducing the false positive rate of spectrum holes by 42%. In terms of interference suppression, a dual-branch deep learning model fuses time-frequency and temporal features through an attention mechanism, combined with deep reinforcement learning (DQN) to dynamically optimize filter parameters, resulting in a 6.2 dB SNR improvement and a 28% reduction in bit error rate. It also supports 16-channel parallel processing with a single-channel latency of less than 5 ms and an interference suppression ratio (ISR) of 24 dB, significantly outperforming traditional fixed filtering schemes. In addition, the model is compressed to 15% of its original size through gradient pruning and integer quantization techniques, with a memory footprint of less than 50MB. It also achieves cloud-edge collaborative online incremental learning based on the Elastic Weights Consolidation (EWC) strategy, ensuring the long-term stability of the system in complex electromagnetic environments.

[0147] This system's technological breakthroughs provide a highly reliable, low-latency spectrum management solution for fields such as 5G / 6G communications, military electronic warfare, and the Internet of Things. Its multi-band collaborative sensing capability effectively improves spectrum resource utilization, its dynamic interference suppression mechanism significantly enhances communication reliability, and its lightweight deployment and online optimization functions significantly reduce equipment costs and maintenance expenses. In the future, by further expanding frequency band coverage, optimizing the generalization ability of graph neural networks, and strengthening real-time processing performance at the edge, this system is expected to become a core supporting technology for next-generation intelligent spectrum management, laying a solid foundation for building efficient and secure wireless communication networks.

[0148] The second embodiment of this disclosure discloses a multi-band electromagnetic spectrum map intelligent fusion system, which includes the following units: The acquisition unit is used to synchronously acquire time-domain signals of the first frequency band and the second frequency band, and construct an original signal set with spatiotemporal consistency between the time-domain signals of the first frequency band and the time-domain signals of the second frequency band; The transformation unit is used to perform short-time Fourier transform on the first frequency band time-domain signal and the second frequency band time-domain signal in the original signal set to generate a first time-frequency graph, and to perform denoising and standardization processing on the first time-frequency graph to obtain a second time-frequency graph; A construction unit is used to construct a graph neural network, taking the second time-frequency graph as the input of the graph neural network to obtain graph nodes and an interference coupling matrix; The extraction unit is used to input the second time-frequency graph into the ResNet-18 network to extract time-frequency features, input the segmented signals of the spatiotemporally consistent original signal set into the bidirectional LSTM to extract time-series features, and assign weights to the time-frequency features and time-series features according to the graph nodes and the interference coupling matrix to generate a joint interference feature vector. A generation unit is used to construct a deep Q-network agent, wherein the joint interference feature vector serves as the state space of the deep Q-network agent, and the action space generated by the deep Q-network agent serves as the filter parameters; the digital FIR filter parameters are dynamically adjusted according to the filter parameters generated by the deep Q-network agent to suppress interference components. The compression unit is used to compress the graph neural network, ResNet-18 network, bidirectional LSTM and deep Q network proxies based on the gradient magnitude criterion, and remove convolutional kernels with lower gradient flow ranking.

[0149] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0150] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the learning outcome prediction method based on learning behavior data mining of the foregoing embodiments of this disclosure.

[0151] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0152] like Figure 5 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 5The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0153] like Figure 5 As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0154] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although a computer device with various devices is illustrated, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0155] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processor, all or part of the steps of the learning outcome prediction method based on learning behavior data mining according to embodiments of this disclosure are performed.

[0156] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0157] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When the non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the learning outcome prediction method based on learning behavior data mining according to the foregoing embodiments of the present disclosure are performed.

[0158] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0159] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0160] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0161] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0162] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0163] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0164] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0165] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0166] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for intelligent fusion of multi-band electromagnetic spectrum maps, characterized in that, Includes the following steps: Simultaneously acquire time-domain signals from the first frequency band and the second frequency band to construct an original signal set that is spatiotemporally consistent with the time-domain signals from the first frequency band and the second frequency band. A first time-frequency graph is generated by performing short-time Fourier transform on the first frequency band time-domain signal and the second frequency band time-domain signal in the original signal set. The first time-frequency graph is then denoised and standardized to obtain a second time-frequency graph. A graph neural network is constructed, and the second time-frequency graph is used as the input of the graph neural network to obtain graph nodes and interference coupling matrix; The second time-frequency graph is input into the ResNet-18 network to extract time-frequency features, and the segmented signals of the original spatiotemporally consistent signal set are input into the bidirectional LSTM to extract time-series features. Weights are assigned to the time-frequency features and time-series features according to the graph nodes and the interference coupling matrix to generate a joint interference feature vector. A deep Q-network agent is constructed, with the joint interference feature vector serving as the state space of the deep Q-network agent and the action space generated by the deep Q-network agent serving as the filter parameters; the digital FIR filter parameters are dynamically adjusted based on the filter parameters generated by the deep Q-network agent to suppress interference components. Based on the gradient magnitude criterion, the graph neural network, ResNet-18 network, bidirectional LSTM and deep Q network proxies are compressed, and convolutional kernels with lower gradient flow rankings are removed.

2. The method according to claim 1, characterized in that, The original signal set used to construct the spatiotemporal consistency between the time-domain signal of the first frequency band and the time-domain signal of the second frequency band further includes: The first frequency band time domain signal is a Sub-6GHz frequency band time domain signal, and the second frequency band time domain signal is a millimeter wave frequency band time domain signal; the spatiotemporal consistency of the Sub-6GHz frequency band time domain signal and the millimeter wave frequency band time domain signal is ensured by GPS and timestamp alignment algorithm.

3. The method according to claim 2, characterized in that, The step of generating a first time-frequency graph by performing short-time Fourier transform on the first frequency band time-domain signal and the second frequency band time-domain signal in the original signal set, and then performing denoising and standardization processing on the first time-frequency graph to obtain a second time-frequency graph specifically includes: Short-time Fourier transform is performed on the first frequency band time-domain signal and the second frequency band time-domain signal in the original signal set. The short-time Fourier transform uses a Hamming window function with a window length of 256 points, and the generated first time-frequency graph has a size of 3×224×224. High-frequency noise is eliminated using a wavelet thresholding denoising algorithm, with the noise standard deviation... The wavelet threshold is calculated using the background noise segment of the original signal set: Where N is the signal length; The second time-frequency graph is obtained by performing Z-score normalization on the first time-frequency graph. The Z-score normalization formula is as follows: Where μ is the mean energy of the first time-frequency graph, and σ is the noise standard deviation.

4. The method according to claim 3, characterized in that, The construction of the graph neural network, using the second time-frequency graph as input to the graph neural network, specifically includes: A three-layer graph convolutional network is constructed, and the second time-frequency graph is input into the three-layer graph convolutional network to obtain a graph-level representation with the first frequency band time-frequency signal and the second frequency band time-frequency signal as graph nodes; The interference coupling matrix is ​​calculated using the time-frequency signals of the first and second frequency bands as graph nodes. The interference coupling matrix serves as the edge weights of the graph, dynamically reflecting the interference intensity between frequency bands. The interference coupling matrix: in, and They are signals from different frequency bands. It is covariance. The standard deviation is denoted as .

5. The method according to claim 4, characterized in that, Inputting the second time-frequency graph into the ResNet-18 network to extract time-frequency features, and inputting the segmented signals of the spatiotemporally consistent original signal set into a bidirectional LSTM to extract time-series features specifically includes: The second time-frequency image is input into the ResNet-18 network to extract time-frequency features. The last layer of the ResNet-18 network uses global average pooling to output 512-dimensional time-frequency features, as shown in the formula: GAP stands for Global Average Pooling. F The second time-frequency graph represents the input, and the second time-frequency graph is a three-dimensional tensor, represented as... C × H × W ,in C It is the number of channels. H It is the height of the feature map. W It is the width of the feature map; H The height of the second time-frequency map is represented by , and the dimension of the feature map in the vertical direction is represented by . W The width of the second time-frequency map is represented by , and the dimension of the feature map in the horizontal direction is represented by . This indicates that the second time-frequency graph is at position ( i , j The value at ) i It is an index in the height direction. j It is an index along the width direction; The segmented signals of the original spatiotemporally consistent signal set are input to the bidirectional LSTM to extract temporal features. The size of the bidirectional LSTM hidden layer is 128, and it outputs 128-dimensional temporal features.

6. The method according to claim 5, characterized in that, The process of assigning weights to the time-frequency features and time-series features based on the graph nodes and the interference coupling matrix, and generating a joint interference feature vector, specifically includes: The query vector Q set and key vector K set of the 4-head attention mechanism are determined based on the graph nodes and the interference coupling matrix. The weights of time-frequency features and time-series features are dynamically allocated through the four-head attention mechanism, and the weight calculation formula is as follows: Where Softmax is the activation function. The dimension of the key vector; For the first i A query vector with attention heads. No. i The key vector of each attention head, where T represents the transpose matrix. Based on the weights, a joint interference feature vector is generated from the time-frequency features and the time-series features.

7. The method according to claim 6, characterized in that, Construct a deep Q-network agent, where the joint interference feature vector serves as the state space of the deep Q-network agent, and the action space generated by the deep Q-network agent serves as the filter parameters, specifically including: Construct a deep Q-network agent, the state space of which is the 768-dimensional joint interference feature vector, and the action space generated by the deep Q-network agent is used as filter parameters; Define the reward function: Where ΔSNR is the signal-to-noise ratio improvement and ΔBER is the bit error rate reduction; During the training of the deep Q-network agent, an experience replay pool with a capacity of 100,000 is used to stabilize the learning process; an ε-greedy policy with ε of 0.1 is used to balance exploration and exploitation; and a reward function with a discount factor γ of 0.99 is set for long-term reward, and training continues until the policy converges.

8. The method according to claim 6, characterized in that, Dynamically adjusting the digital FIR filter parameters to suppress interference components based on the filter parameters generated by the deep Q-network proxy specifically includes: Obtain the cutoff frequency adjustment Δf output by the deep Q-network proxy, and update the cutoff frequency, wherein the formula for updating the cutoff frequency is: The parameters of the digital FIR filter are calculated using the Hanning window function, which is defined as follows: The formula for dynamically adjusting the parameters of a digital FIR filter is: 。 9. The method according to claim 7, characterized in that, The aforementioned model compression based on gradient magnitude criteria, involving the removal of convolutional kernels with low gradient flow ranking, specifically includes: Gradient magnitude pruning is performed by calculating the absolute value of the convolution kernel gradient for each neural network. The absolute value of the convolution kernel gradient is: L represents the model's loss function. This represents the convolution kernel parameters in the graph neural network, the two-branch deep learning network, and the deep Q-network agent; After removing 30% of the convolutional kernels after gradient ranking, the compressed model parameters are quantized to 8-bit integers using a scaling factor and quantization formula, and then the quantized model is deployed. The scaling factor is calculated as follows: This represents the set of convolutional kernel parameters in the graph neural network, the two-branch deep learning network, and the deep Q-network agent; The quantification formula is: ; Incremental learning is carried out using an elastic weight solidification strategy, retaining the weights of key parameters and periodically evaluating system performance. If the bit error rate rises above the threshold, model retraining is triggered. The loss function is defined as follows: in, Let L be the new loss function, and L be the original loss function. Here is the Fisher information matrix, λ=1000. These are the historically optimal parameters.

10. A multi-band electromagnetic spectrum map intelligent fusion system, characterized in that, Includes the following units: The acquisition unit is used to synchronously acquire time-domain signals of the first frequency band and the second frequency band, and construct an original signal set with spatiotemporal consistency between the time-domain signals of the first frequency band and the time-domain signals of the second frequency band; The transformation unit is used to perform short-time Fourier transform on the first frequency band time-domain signal and the second frequency band time-domain signal in the original signal set to generate a first time-frequency graph, and to perform denoising and standardization processing on the first time-frequency graph to obtain a second time-frequency graph; A construction unit is used to construct a graph neural network, taking the second time-frequency graph as the input of the graph neural network to obtain graph nodes and an interference coupling matrix; The extraction unit is used to input the second time-frequency graph into the ResNet-18 network to extract time-frequency features, input the segmented signals of the spatiotemporally consistent original signal set into the bidirectional LSTM to extract time-series features, and assign weights to the time-frequency features and time-series features according to the graph nodes and the interference coupling matrix to generate a joint interference feature vector. A generation unit is used to construct a deep Q-network agent, wherein the joint interference feature vector serves as the state space of the deep Q-network agent, and the action space generated by the deep Q-network agent serves as the filter parameters; the digital FIR filter parameters are dynamically adjusted according to the filter parameters generated by the deep Q-network agent to suppress interference components. The compression unit is used to compress the graph neural network, ResNet-18 network, bidirectional LSTM and deep Q network proxies based on the gradient magnitude criterion, and remove convolutional kernels with lower gradient flow ranking.

Images