Radio detection sensitivity improving method and system based on AI operation

Abstract

The invention discloses a radio detection sensitivity improving method and system based on AI operation. The method comprises the following steps: acquiring an original weak signal of a measurement node, and carrying out noise reduction, feature extraction and data encapsulation to generate a processed weak signal; processing signals synchronously uploaded by multiple nodes in the same time period are sorted into a signal feature set; constructing a map neural network node space based on the geographic position of the measurement node and the detection map, and inputting the signal feature set into the space to form a joint signal feature; performing signal reconstruction on the joint features through a signal reconstruction model constructed by a multi-scale memory network, wherein the model is trained by adopting a small sample-adversarial mode; when it is detected that the joint feature anomaly ratio exceeds a threshold value, the adversarial sample is updated, and model retraining is triggered. According to the method, high-precision reconstruction and abnormal self-adaptive optimization of weak signals are realized through multi-node spatial-temporal feature fusion and a dynamic adversarial training mechanism.

Description

Technical Field

[0001] The present invention relates to the field of radio detection, and in particular to a method and system for improving the sensitivity of radio detection based on AI computing. Background Art

[0002] Radio detection mostly relies on single-node digital signal processing and static filtering technology. Although it can suppress local interference, it lacks the ability to model multi-node spatiotemporal correlation features. In addition, the use of a fixed training mode makes it difficult for the model to dynamically adapt to abnormal fluctuations in complex signal environments. The accuracy of weak signal reconstruction is limited, and the improvement of sensitivity and anti-interference capabilities faces bottlenecks. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to propose a method and system for improving the sensitivity of radio detection based on AI calculations to solve the problems of low accuracy and insufficient sensitivity of weak signal reconstruction caused by the inability to dynamically adapt to abnormal fluctuations.

[0004] To achieve the above technical objectives, in a first aspect, the present application provides a method for improving the sensitivity of radio detection based on AI computing, comprising:

[0005] Obtaining the original weak signal of the measurement node and preprocessing the original weak signal to obtain a processed weak signal. The preprocessing includes noise reduction, feature extraction and data encapsulation of the original weak signal.

[0006] Upload multiple processed weak signals synchronously, and organize the processed weak signals in the same period into a signal feature set;

[0007] Obtain the geographic location information and detection map of the measurement nodes and construct the graph neural network node space;

[0008] Input the signal feature set into the graph neural network node space to form a joint signal feature;

[0009] The joint signal features are input into the trained signal reconstruction model to obtain the reconstructed original signal. The signal reconstruction model is configured as a multi-scale memory network and trained using a small sample-adversarial model.

[0010] In addition, the anomaly ratio in the joint signal feature is obtained. If the anomaly ratio exceeds the preset anomaly threshold, the sample data in the small sample-adversarial mode is updated, and after the update, the small sample-adversarial mode is triggered to retrain the signal reconstruction model.

[0011] In some embodiments, obtaining an original weak signal from a measurement node and preprocessing the original weak signal to obtain a processed weak signal includes:

[0012] An adaptive constant false alarm rate (CFAR) detection algorithm is used to eliminate background noise from the original weak signal and obtain a primary filtered signal. The adaptive CFAR detection algorithm dynamically adjusts the detection threshold to keep the false alarm probability constant and calculates the local noise power estimate within a sliding window.

[0013] Perform a time-frequency joint analysis on the primary filtered signal and separate the aliased signal components using the Wigner-Ville distribution algorithm to obtain the separated signal. The Wigner-Ville distribution algorithm calculates the energy distribution of the signal on the time-frequency plane and effectively separates the signal components through cross-term suppression processing.

[0014] The separated signal is subjected to deep noise suppression processing using a long short-term memory autoencoder to obtain a noise-reduced signal. The long short-term memory autoencoder extracts the signal timing features through the encoder and reconstructs the pure signal waveform in the decoder.

[0015] A lightweight convolutional neural network is used to extract multidimensional features from the denoised signal, generating a feature vector containing time domain features, frequency domain features, and modulation features. The lightweight convolutional neural network uses a channel pruning method to remove redundant convolution kernels and retain key feature extraction channels.

[0016] The feature vector is encapsulated with the corresponding timestamp and geographic location information to form a standard format for processing weak signals. The data encapsulation process includes feature dimension normalization and metadata verification steps.

[0017] In some embodiments, synchronously uploading multiple processed weak signals and organizing the processed weak signals in the same time period into a signal feature set includes:

[0018] Performing time alignment processing based on timestamp information contained in the processed weak signal to obtain a first processed signal;

[0019] Establish a spatial index based on the geographical location information in the weak signal processing to determine the relative position relationship of each measurement node;

[0020] Sorting the first processed signals according to the spatial index to construct a feature matrix of spatiotemporal correlation;

[0021] Perform integrity check on the feature matrix and remove missing or abnormal data entries;

[0022] The feature matrix is ​​converted into a signal feature set output, where the signal feature set includes a plurality of second processed signals.

[0023] In some embodiments, obtaining geographic location information of measurement nodes and a detection map and constructing a graph neural network node space includes:

[0024] Obtain geographic location information from weak signals and load a preset detection map containing terrain elevation, obstacle distribution, and electromagnetic environment characteristics.

[0025] Spatial registration is performed between the geographic location information and the detection map to obtain registration features, including:

[0026] Calculate the precise projection coordinates of geographic location information in the detection map;

[0027] Establishing a correlation mapping between geographic location information and map features;

[0028] Mark electromagnetic sensitive areas and signal blocking areas;

[0029] Generate registration features based on precise projection coordinates, correlation mapping, electromagnetic sensitive areas, and signal-blocked areas;

[0030] The GNN node space is constructed based on the registration features, and the GNN node space containing spatial topology and electromagnetic characteristics is obtained, including:

[0031] Each measurement node is a graph node, and the spatial relationship between measurement nodes is an edge connection relationship;

[0032] Calculate the edge weight based on the distance between the measured nodes and the line-of-sight condition;

[0033] The map features in the fused detection map are used as node attributes.

[0034] In some embodiments, inputting a signal feature set into a graph neural network node space to form a joint signal feature includes:

[0035] Mapping the second processed signal to a graph node corresponding to a measurement node in the graph neural network;

[0036] Based on the edge connection relationship in the node space of the graph neural network, the graph attention network is used to calculate the signal spatial dependency between the measurement nodes. The signal spatial dependency is configured to be represented by spatial attention weights, including:

[0037] Taking the spatial distance and visibility conditions of the measurement nodes as prior knowledge, the correlation score between the graph node features corresponding to each pair of connected measurement nodes is calculated;

[0038] Normalize the relevance scores to generate spatial attention weights;

[0039] Based on the spatial attention weight, the signal features between measurement nodes are fused to obtain spatial enhancement features, including:

[0040] performing weighted aggregation on the second processed signal features of adjacent measurement nodes according to the spatial attention weights;

[0041] retaining the second processed signal characteristics of the current measurement node itself;

[0042] The aggregated adjacent node features are concatenated with the current node features to generate spatial enhancement features;

[0043] The spatial enhancement features are processed at multiple levels to obtain joint signal features, including:

[0044] In the first graph attention layer, the signal features of directly adjacent measurement nodes are fused using spatial attention weights;

[0045] In the second graph attention layer, the signal features of indirectly connected measurement nodes are fused based on the output features of the first layer;

[0046] Maintaining the feature information of the original second processed signal through cross-layer residual connections;

[0047] The fused joint signal feature is output, and the joint signal feature includes the collaborative representation of the second processed signal of each measurement node after the spatial topological relationship is enhanced.

[0048] In some embodiments, the joint signal features are input into a trained signal reconstruction model to obtain a reconstructed original signal. The signal reconstruction model is configured as a multi-scale memory network, including:

[0049] Construct a short-term memory module to obtain short-term memory features, including:

[0050] A temporal convolutional network is used to process the joint signal features. The temporal convolutional network captures the local waveform features by dilating the convolution kernel, which is expressed by formula (1). Formula (1) is as follows:

[0051] ST = TCN(J);

[0052] In formula (1), ST is the short-term memory feature, J is the joint signal feature, and TCN(J) is the temporal convolution operation;

[0053] Construct a long-term memory module to obtain long-term memory features, including:

[0054] The Transformer encoder is used to model the dependencies across measurement nodes, which is expressed by formula (2). Formula (2) is as follows:

[0055]

[0056] In formula (2), LT is the long-term memory feature, F tr (J+F at (Q,K,V)) is the encoding function of Transformer, F at (Q, K, V) is the expression of dependency relationship, is the attention mechanism function, Q is the query matrix, K is the key matrix, V is the value matrix, d is the feature dimension, K T is the transposed matrix of the key matrix;

[0057] Build a prototype memory library and obtain retrieval prototype features, including:

[0058] Construct an updateable signal category prototype set, and retrieve the signal category prototype in the signal category prototype set through the attention mechanism, which is expressed by formula (3). Formula (3) is as follows:

[0059]

[0060] In formula (3), PT is the retrieval prototype feature, α l is the contribution weight of the lth signal category prototype in signal reconstruction, m l is the prototype feature of the prototype of the l-th signal category, sim(ST,m l ) is the cosine similarity function, exp(sim(ST,m l )) is ST and m l The matching degree, ∑ p exp(sim(ST,m p )) is normalized, p is the counting unit, m P ∈M,M={m1,m2,…,m l}, M is the prototype set of signal categories;

[0061] The short-term memory feature, long-term memory feature and retrieval prototype feature are subjected to feature fusion and signal reconstruction to obtain the reconstructed original signal, which is expressed by formula (4). Formula (4) is as follows:

[0062]

[0063] In formula (4), G is the fusion feature of short-term memory feature, long-term memory feature and retrieval prototype feature. To reconstruct the original signal, F de (G) is the convolution function.

[0064] In some embodiments, training the signal reconstruction model using the few-shot adversarial model includes:

[0065] Build a small sample training module, including:

[0066] A support set is constructed by randomly selecting several signal category prototypes from the signal category prototype set. A preset number of samples are provided for each category to form a small sample training task. The support set is used to simulate the small sample conditions in actual scenarios.

[0067] Calculate the prototype center of each signal category prototype in the support set. The prototype center is obtained by aggregating the features of the support set samples through the feature extraction function of the multi-scale memory network, and is used to establish the baseline feature representation of the signal category.

[0068] Prototype loss is calculated based on the prototype center, which minimizes the distance between the same type of signal features and the prototype center, maximizes the distance between different types of signal features, and enhances the ability to distinguish signal categories.

[0069] Build an adversarial training module, including:

[0070] Generate adversarial samples based on the original training samples. The adversarial samples are obtained by calculating the gradient of the loss function with respect to the input sample and adding a perturbation of a limited magnitude along the gradient direction.

[0071] The adversarial sample is input into the multi-scale memory network, and the adversarial loss between the reconstructed signal and the true signal is calculated. The adversarial loss is used to evaluate the robustness of the model to interference signals. By maximizing the adversarial loss, the model's anti-interference ability is improved.

[0072] Execute the joint optimization process until training is complete, including:

[0073] Alternate between small-sample training in the small-sample training module and adversarial training in the adversarial training module. Small-sample training adapts the model to small-sample conditions by optimizing the prototype loss, while adversarial training improves model robustness by optimizing the adversarial loss.

[0074] The objective function is constructed by combining prototype loss, adversarial loss and reconstruction loss. The reconstruction loss is used to ensure the fidelity of signal reconstruction.

[0075] Dynamically adjust the weight parameters corresponding to the combined prototype loss, adversarial loss, and reconstruction loss, and adaptively balance the optimization intensity of different training objectives according to the training stage.

[0076] In some embodiments, obtaining an abnormality ratio in the joint signal feature includes:

[0077] Perform time-frequency domain feature extraction on the joint signal features to obtain a time-frequency distribution feature matrix containing instantaneous frequency, spectral entropy and modulation depth features;

[0078] Calculate the Mahalanobis distance between the signal feature vector in the time-frequency distribution feature matrix and the standard signal vector to obtain the anomaly score corresponding to the signal feature vector;

[0079] Count the number of signal feature vectors whose anomaly scores exceed the dynamic anomaly threshold in the current detection cycle, and record it as the number of anomaly features;

[0080] The ratio of the number of abnormal features to the total number of signal feature vectors in the detection period is calculated as the abnormality ratio.

[0081] In some embodiments, updating sample data in the small sample-adversarial mode and triggering the small sample-adversarial mode to retrain the signal reconstruction model after the update includes:

[0082] Based on the detection results whose anomaly ratio exceeds the preset anomaly threshold, the signal feature vectors whose anomaly scores meet the requirements within the current detection cycle are selected as new sample data;

[0083] Add the newly added sample data to the training dataset of the small sample-adversarial model, and remove some of the original sample data in the training dataset to complete the update of the sample data;

[0084] The signal reconstruction model is retrained using the updated training dataset, and the updated signal reconstruction model is used to continue processing the new joint signal features.

[0085] In a second aspect, the present invention further provides a radio detection sensitivity enhancement system based on AI computing, applicable to the method of the first aspect, the system comprising:

[0086] Multiple measurement nodes are arranged in a preset area according to a preset method. The preset area is the area to be detected. The measurement nodes are used to obtain original weak signals of the measurement nodes.

[0087] Multiple edge processors, each edge processor is set at a measurement node, and the edge processor is used to pre-process the original weak signal;

[0088] Multiple communication units, each communication unit is set at a measurement node, and the communication unit is used to synchronously upload multiple processed weak signals;

[0089] The server is used to organize the processed weak signals of the same time period into a signal feature set; obtain the geographic location information and detection map of the measurement node, and construct the graph neural network node space; input the signal feature set into the graph neural network node space to form a joint signal feature; input the joint signal feature into the trained signal reconstruction model to obtain the reconstructed original signal. The signal reconstruction model is configured as a multi-scale memory network construction, and the signal reconstruction model is trained using a small sample-adversarial mode; and, obtain the anomaly ratio in the joint signal feature. If the anomaly ratio exceeds the preset anomaly threshold, the sample data in the small sample-adversarial mode is updated, and after the update, the small sample-adversarial mode is triggered to retrain the signal reconstruction model.

[0090] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0091] The present invention provides a method and system for improving the sensitivity of radio detection based on AI computing. The method includes: obtaining the original weak signal of the measurement node and performing noise reduction, feature extraction, and data encapsulation to generate a processed weak signal; organizing the processed signals of the same time period uploaded synchronously by multiple nodes into a signal feature set; constructing a graph neural network node space based on the geographic location of the measurement node and the detection map, and inputting the signal feature set into this space to form a joint signal feature; reconstructing the joint feature through a signal reconstruction model constructed by a multi-scale memory network, wherein the model adopts a small sample-adversarial mode training; when it is detected that the joint feature anomaly ratio exceeds a threshold, updating the adversarial sample and triggering model retraining. The above technical solution achieves high-precision reconstruction of weak signals and abnormal adaptive optimization through multi-node spatiotemporal feature fusion and dynamic adversarial training mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0093] Figure 1 is a method step diagram of steps S101 to S105 of the method described in the specific embodiment;

[0094] Figure 2 is a method step diagram of steps S201 to S205 of the method described in the specific embodiment;

[0095] Figure 3 It is a method step diagram of steps S301 to S305 of the method described in the specific implementation method. DETAILED DESCRIPTION

[0096] The present invention will be described in further detail below with reference to the accompanying drawings and examples. It is particularly noted that the following examples are intended only to illustrate the present invention and are not intended to limit the scope of the present invention. Similarly, the following examples are only some embodiments of the present invention and are not intended to be exhaustive. All other embodiments obtained by those of ordinary skill in the art without creative effort are intended to fall within the scope of protection of the present invention.

[0097] See also Figure 1 In a first aspect, this embodiment provides a method for improving the sensitivity of radio detection based on AI computing, comprising:

[0098] S101, obtaining an original weak signal from a measurement node, and preprocessing the original weak signal to obtain a processed weak signal, wherein the preprocessing includes performing noise reduction, feature extraction, and data encapsulation on the original weak signal;

[0099] S102, synchronously uploading multiple processed weak signals, and organizing the processed weak signals in the same time period into a signal feature set;

[0100] S103: Obtaining geographic location information of measurement nodes and detection maps, and constructing a graph neural network node space;

[0101] S104. Input the signal feature set into the graph neural network node space to form a joint signal feature;

[0102] S105. Input the joint signal features into a trained signal reconstruction model to obtain a reconstructed original signal. The signal reconstruction model is configured as a multi-scale memory network and trained using a small sample-adversarial model.

[0103] In addition, the anomaly ratio in the joint signal feature is obtained. If the anomaly ratio exceeds the preset anomaly threshold, the sample data in the small sample-adversarial mode is updated, and after the update, the small sample-adversarial mode is triggered to retrain the signal reconstruction model.

[0104] In step S101, the original weak signal refers to the low signal-to-noise ratio signal captured by the measurement node through the radio receiving device. Its preprocessing includes noise reduction, feature extraction, and data packaging. Noise reduction uses filtering algorithms to suppress interference from environmental noise and hardware noise. Feature extraction parameterizes the signal's time-frequency domain characteristics, such as spectral peaks and phase continuity. Data packaging converts the processed signal into a unified format, such as a time-frequency matrix, to facilitate subsequent transmission and joint analysis.

[0105] In step S102, synchronous uploading ensures the synchronization of multi-node signal acquisition in the same period through timestamp alignment. The signal feature set refers to integrating the processed weak signals of each node into a multi-dimensional feature matrix in time sequence, which is used to characterize the spatiotemporal distribution characteristics of multi-source signals.

[0106] In step S103, the geographic location information includes the latitude and longitude coordinates and relative position relationship of the measurement node. The detection map is an electronic map containing geographic environment features, and the geographic environment features include obstacle distribution, electromagnetic interference source marking, etc. The graph neural network node space maps the measurement nodes to vertices in the graph structure, and constructs connecting edges between vertices according to geographic location, signal propagation path and environmental characteristics, thereby modeling the spatial correlation between multiple nodes.

[0107] In step S104, the joint signal feature refers to the inter-node information transmission and aggregation of the signal feature set through the graph neural network, such as the weighted fusion of neighborhood node features, to generate a high-order feature representation of the fused spatiotemporal correlation.

[0108] In step S105, the multi-scale memory network extracts local signal details and global temporal dependencies through parallel multi-level convolutional modules and introduces memory units to store historical feature patterns. In small-sample adversarial training, the generator generates a reconstructed signal based on a small number of labeled samples, and the discriminator distinguishes the generated signal from the real signal, optimizing the model parameters through adversarial iteration. The anomaly ratio is calculated by counting the proportion of feature dimensions in the joint signal features that deviate from the historical distribution threshold. The preset anomaly threshold is set based on historical data statistics or empirical values. When retraining is triggered, updating the sample data means incorporating the current anomaly features into the generator input of adversarial training to enhance the model's adaptability to abnormal patterns.

[0109] The steps of this embodiment can be understood as: breaking through the limitations of single-node isolated processing through multi-node signal spatiotemporal joint modeling (i.e., graph neural network node space), using dynamic adversarial training mechanism (i.e., small sample-adversarial mode retraining) to achieve the model's adaptive optimization of abnormal signals, and finally improving the accuracy of weak signal reconstruction through multi-scale feature fusion (i.e., signal reconstruction model). For example, when a sudden strong interference in a certain area causes abnormal signals in local nodes, the graph neural network compensates for the characteristics of the damaged nodes through spatial correlation, and at the same time, the abnormal ratio exceeds the threshold to trigger model retraining, so that the reconstruction model can quickly adapt to the new interference pattern.

[0110] This embodiment constructs a graph neural network node space to integrate the geographic location information of multiple measurement nodes with the geographic environmental features in the detection map, mapping the scattered original weak signals into spatially correlated graph structure vertices. Connecting edges are constructed using the signal propagation paths and environmental characteristics between nodes, overcoming the limitations of the traditional single-node isolated processing mode with insufficient spatiotemporal correlation modeling capabilities. By characterizing the spatiotemporal distribution characteristics of the signal feature set and aggregating the joint signal features using a graph neural network, high-order fusion of multi-source signals is achieved. Combined with the multi-scale memory network's ability to concurrently extract local details and global temporal dependencies, the accuracy of weak signal reconstruction is effectively improved. A small sample-adversarial training mechanism is employed to iteratively optimize model parameters through dynamic adversarial iterations between the generator and the discriminator. Sample data updates and model retraining are triggered based on anomaly ratio detection, enabling the signal reconstruction model to adapt to abnormal signal fluctuations in complex electromagnetic environments. When the interference source distribution suddenly changes or local node signals become abnormal, spatial correlation is leveraged to compensate for damaged node features and rapidly adjust the reconstruction strategy, thereby maintaining high-sensitivity detection while enhancing anti-interference capabilities and system robustness.

[0111] See also Figure 2In some embodiments, obtaining an original weak signal from a measurement node and preprocessing the original weak signal to obtain a processed weak signal includes:

[0112] S201, using an adaptive constant false alarm rate detection algorithm to eliminate background noise in the original weak signal to obtain a primary filtered signal, wherein the adaptive constant false alarm rate detection algorithm dynamically adjusts the detection threshold value to keep the false alarm probability constant and calculates a local noise power estimate within a sliding window;

[0113] S202, performing a time-frequency joint analysis on the primary filtered signal, separating the aliased signal components using a Wigner-Ville distribution algorithm to obtain a separated signal, calculating the energy distribution of the signal on a time-frequency plane using the Wigner-Ville distribution algorithm, and achieving effective separation of the signal components through cross-term suppression processing;

[0114] S203, using a long short-term memory autoencoder to perform deep noise suppression processing on the separated signal to obtain a noise-reduced signal, the long short-term memory autoencoder extracts signal timing features through the encoder, and reconstructs a pure signal waveform in the decoder;

[0115] S204, performing multi-dimensional feature extraction on the noise reduction signal through a lightweight convolutional neural network to generate a feature vector including time domain features, frequency domain features, and modulation features. The lightweight convolutional neural network uses a channel pruning method to remove redundant convolution kernels and retain key feature extraction channels;

[0116] S205. Encapsulate the feature vector with the corresponding timestamp and geographic location information to form a standard format for processing weak signals. The data encapsulation process includes feature dimension normalization and metadata verification steps.

[0117] In step S201, an adaptive constant false alarm rate detection algorithm is used to eliminate background noise in the original weak signal to obtain a primary filtered signal, including:

[0118] The original weak signal is input into the sliding detection window for segmentation processing;

[0119] Calculate the signal power spectral density within the sliding detection window;

[0120] determining a local noise power estimate based on the signal power spectral density;

[0121] generating a dynamic detection threshold based on a local noise power estimate;

[0122] Use dynamic detection threshold to filter the original weak signal;

[0123] The adaptive constant false alarm rate (CFAR) detection algorithm suppresses background noise by dynamically adjusting the detection threshold, adaptively maintaining a constant false alarm probability based on the local characteristics of the signal. The sliding detection window is a fixed-length interval used to segment the original weak signal. The local noise power within the window is estimated by calculating the signal power spectral density. The dynamic detection threshold is adjusted in real time based on the noise power, thereby suppressing background noise (such as environmental electromagnetic interference and hardware noise floor) while preserving the valid signal components.

[0124] In step S202, a time-frequency joint analysis is performed on the primary filtered signal, and the aliased signal components are separated using the Wigner-Ville distribution algorithm to obtain a separated signal, including:

[0125] Converting the primary filtered signal into analytical signal form;

[0126] Calculate the instantaneous autocorrelation function of the analytical signal;

[0127] Perform a two-dimensional Fourier transform on the instantaneous autocorrelation function;

[0128] The kernel function is used to suppress cross terms of the transformation results;

[0129] Extracting signal components from the suppressed distribution map;

[0130] Joint time-frequency analysis uses the Wigner-Ville distribution algorithm to analyze the time-frequency energy distribution characteristics of the signal. The analytical signal form refers to converting the real signal into a complex signal to eliminate the negative frequency component; cross-term suppression uses a kernel function to smooth the interference components on the time-frequency plane. For example, the Choi-Williams kernel function is used to weaken the false energy distribution caused by signal aliasing, thereby separating the target signal components, such as communication signals with different modulation methods.

[0131] In step S203, a long short-term memory autoencoder is used to perform deep noise suppression processing on the separated signal to obtain a noise-reduced signal, including:

[0132] The separated signal is input into the bidirectional LSTM layer of the encoder;

[0133] Extract temporal features through bidirectional LSTM layer;

[0134] Compress the time series features to the bottleneck layer;

[0135] Use causal LSTM in the decoder to reconstruct the signal waveform;

[0136] Output the reconstructed pure signal waveform;

[0137] The encoder of the long short-term memory autoencoder captures the forward and backward temporal dependencies of the signal through a bidirectional LSTM layer. The bottleneck layer is used to compress the temporal features into a low-dimensional latent space to filter out residual noise. The causal LSTM layer of the decoder gradually reconstructs the pure signal waveform based on the latent features to ensure strict causal alignment of the output timing.

[0138] In step S204, multi-dimensional feature extraction is performed on the noise reduction signal through a lightweight convolutional neural network to generate a feature vector containing time domain features, frequency domain features and modulation features, including:

[0139] The denoised signal is fed into the depthwise separable convolution module;

[0140] Evaluate the feature extraction contribution of each convolution channel;

[0141] Remove convolution channels whose contribution is lower than a preset threshold;

[0142] Use optimized convolutional networks to extract multidimensional features;

[0143] Output feature vector containing time domain, frequency domain and modulation features;

[0144] The lightweight convolutional neural network reduces the number of parameters through the depth-wise separable convolution module. The contribution of feature extraction is quantitatively evaluated by calculating the average activation intensity or gradient importance of the convolution channel output. The preset threshold is determined based on experience or iterative pruning experiments. The time domain features include signal amplitude envelope, zero-crossing rate, etc., the frequency domain features cover the spectrum centroid, bandwidth distribution, and the modulation features involve parameters such as phase jump and symbol rate.

[0145] In step S205, the feature vector is encapsulated with the corresponding timestamp and geographic location information to form a standard format for processing weak signals, including:

[0146] Perform normalization on the feature vector;

[0147] Additional collection timestamp information;

[0148] Add measurement node position coordinates;

[0149] Perform data integrity checks;

[0150] Output the processed weak signal in standard format.

[0151] Preferably, feature dimension normalization eliminates dimensional differences through Z-score standardization, metadata verification verifies the integrity of the association between timestamp, geographic location information and feature vector through hash algorithm, and weak signals processed in standard format usually adopt binary or JSON structured encapsulation to ensure multi-node data compatibility.

[0152] The steps of this embodiment can be understood as: suppressing background noise through adaptive dynamic threshold filtering, improving the accuracy of aliased signal separation by combining time-frequency distribution cross-term suppression, achieving deep noise reduction by utilizing the time series modeling capability of the LSTM autoencoder, extracting multi-dimensional features through a lightweight network, and finally forming multi-source signal data aligned in time and space through standardized packaging. For example, when multiple communication signal bands overlap, the Wigner-Ville distribution separates the independent signal components by suppressing cross-terms, and the LSTM autoencoder further filters out residual noise. The modulation features extracted by the lightweight network can accurately distinguish signals of different communication formats, providing high-quality input for subsequent joint analysis.

[0153] This embodiment dynamically adjusts the detection threshold to suppress background noise through an adaptive constant false alarm rate detection algorithm, retains effective signal components based on the local noise power estimate within the sliding window, and solves the signal distortion problem caused by traditional fixed threshold filtering; adopts the Wigner-Ville distribution algorithm to perform time-frequency joint analysis of the primary filtered signal, eliminates negative frequency interference by analyzing the signal form, and uses the kernel function to suppress cross terms, effectively separating the independent components in the aliased signal; combines the bidirectional LSTM encoder and causal LSTM decoder of the long short-term memory autoencoder to achieve deep noise suppression in the process of time series feature compression and reconstruction, and improves the waveform fidelity of weak signals; through the channel pruning optimization of the lightweight convolutional neural network, the key feature extraction capability is retained and the computational complexity is reduced, while extracting multi-dimensional representations of time domain, frequency domain and modulation features; finally, the weak signal is processed by feature dimension normalization and metadata verification encapsulation standardization to ensure the alignment of spatiotemporal information and multi-node data compatibility. This embodiment forms a closed-loop processing chain from noise suppression, signal separation, deep noise reduction to feature encapsulation, significantly improving the feature extraction quality and preprocessing efficiency of weak signals in complex electromagnetic environments, and laying a data foundation for subsequent multi-node joint analysis and high-precision reconstruction.

[0154] See also Figure 3 In some embodiments, uploading multiple processed weak signals synchronously and organizing the processed weak signals in the same time period into a signal feature set includes:

[0155] S301, performing time alignment processing based on timestamp information contained in the processed weak signal to obtain a first processed signal;

[0156] S302: Establish a spatial index based on the geographical location information in the processed weak signal to determine the relative position relationship of each measurement node;

[0157] S303, sorting the first processed signals according to the spatial index, and constructing a spatiotemporal correlation feature matrix;

[0158] S304: Perform integrity check on the feature matrix and remove missing or abnormal data entries;

[0159] S305: Convert the feature matrix into a signal feature set output, where the signal feature set includes a plurality of second processed signals.

[0160] In step S301, time alignment processing is performed based on the timestamp information contained in the weak signal, including:

[0161] Extract GPS timestamps and local sampling timestamps from the signal data of each measurement node;

[0162] Calculate the clock deviation of each node and perform compensation calibration;

[0163] Use interpolation algorithm to time align asynchronous sampling points;

[0164] Establish a unified time reference coordinate system;

[0165] The GPS timestamp is the global time reference provided by satellite timing, while the local sampling timestamp is the sampling instant recorded by the node's internal clock. Time alignment calculates clock deviations at each node, such as the time difference between the GPS and local clocks, and performs compensation calibration. Asynchronous sampling points are aligned to a unified time reference coordinate system using interpolation algorithms (such as linear or spline interpolation), ensuring synchronization of multi-node signals on the same time scale.

[0166] In step S302, a spatial index is established based on the geographical location information in the weak signal, including:

[0167] Analyze the latitude and longitude coordinate data of each measurement node;

[0168] Calculate the relative distance and azimuth relationship between nodes;

[0169] Construct a spatial location topology map based on spherical triangulation;

[0170] Assign a unique spatial identifier to each node;

[0171] The geographic location information is obtained by parsing and processing the latitude and longitude coordinate data encapsulated in weak signals. The spatial index is established by calculating the relative distance and azimuth between nodes through spherical triangulation (such as calculating the great circle distance with the geocentric coordinate system as a reference), and constructing a spatial location topology map (a network structure with nodes as vertices and relative position relationships as edges). The unique spatial identifier is a unique code generated based on the node latitude and longitude hash, which is used to characterize the spatial ownership of the node.

[0172] In step S303, the time-aligned processed weak signals are sorted according to the spatial index to construct a spatiotemporal correlation feature matrix, including:

[0173] Resample the feature vectors according to a unified time base;

[0174] Sort the feature vectors by spatial identifier order;

[0175] Combining time domain features and spatial domain features to form a three-dimensional feature tensor;

[0176] Add spatiotemporal correlation tags to establish mapping relationships between features;

[0177] The feature matrix of spatiotemporal association resamples the feature vectors through a unified time reference, for example, aligning non-uniformly sampled signals to a fixed time interval, arranging the feature vectors in order of spatial identifiers to form a spatial dimension, combining time domain features (such as temporal changes in signal amplitude) and spatial domain features (such as differences in signal strength between nodes) to construct a three-dimensional feature tensor, and the spatiotemporal association label establishes a mapping relationship between features through coordinate encoding, where the coordinate encoding can be a combination of a time index and a spatial identifier.

[0178] In step S304, the integrity check of the feature matrix is ​​performed, including:

[0179] Check whether the data coverage at each time point meets the requirements;

[0180] Verify the dimensionality consistency of the feature vector;

[0181] Detect outliers and perform data repair;

[0182] Confirm the correctness of spatiotemporal association labels;

[0183] Meeting data coverage requirements means that each time point must contain valid data from a preset proportion of nodes. Dimension consistency verification ensures that all feature vectors have the same length and structure. Outlier detection uses statistical distribution or clustering algorithms to identify data entries that deviate from group characteristics. Data repair can be achieved by interpolating features from adjacent nodes or filling in historical data.

[0184] In step S305, a signal feature set including time synchronization and spatial correlation is output.

[0185] In the output signal feature set, the second processed signal refers to the standardized feature data unit after time-space alignment, index sorting and verification.

[0186] The steps of this embodiment can be understood as: eliminating multi-node data heterogeneity through spatiotemporal benchmark unification (i.e., time alignment and spatial indexing), realizing signal spatiotemporal correlation characterization by topological modeling (i.e., spatial position topology map) and tensor construction (i.e., spatiotemporal correlation feature matrix), and then ensuring input quality through data cleaning (i.e., integrity verification). For example, when a node is temporarily missing due to occlusion, the missing data can be repaired by interpolating the spatial correlation of adjacent nodes, and spatiotemporal labeling ensures that the repaired data is accurately mapped to the original acquisition environment, providing highly consistent input for subsequent joint analysis.

[0187] This embodiment achieves the unification of the time base of multi-node signals through deviation compensation calibration between GPS timestamps and local sampling timestamps, and constructs a spatial position topology map and assigns a unique spatial identifier in combination with spherical triangulation to solve the problem of signal correlation loss caused by heterogeneous time and space bases in distributed measurement; resamples the feature vectors based on the unified time base and constructs a three-dimensional feature tensor by sorting them according to the spatial identifier to achieve deep fusion of time domain features and spatial domain features; ensures the integrity and reliability of the time and space correlation feature matrix through data coverage compliance verification, dimensional consistency verification and outlier detection and repair mechanism; the final output standardized signal feature set accurately maps the acquisition environment of multi-node signals through time and space correlation tags, providing high-quality input data with clear time and space alignment and topological structure for subsequent joint analysis, significantly improving the accuracy and efficiency of collaborative processing of multi-source weak signals.

[0188] In some embodiments, obtaining geographic location information of measurement nodes and a detection map and constructing a graph neural network node space includes:

[0189] Obtain geographic location information from weak signals and load a preset detection map containing terrain elevation, obstacle distribution, and electromagnetic environment characteristics.

[0190] Spatial registration is performed between the geographic location information and the detection map to obtain registration features, including:

[0191] Calculate the precise projection coordinates of geographic location information in the detection map;

[0192] Establishing a correlation mapping between geographic location information and map features;

[0193] Mark electromagnetic sensitive areas and signal blocking areas;

[0194] Generate registration features based on precise projection coordinates, correlation mapping, electromagnetic sensitive areas, and signal-blocked areas;

[0195] The GNN node space is constructed based on the registration features, and the GNN node space containing spatial topology and electromagnetic characteristics is obtained, including:

[0196] Each measurement node is a graph node, and the spatial relationship between measurement nodes is an edge connection relationship;

[0197] Calculate the edge weight based on the distance between the measured nodes and the line-of-sight condition;

[0198] The map features in the fused detection map are used as node attributes.

[0199] In this embodiment, the geographic location information includes the latitude and longitude coordinates for processing weak signal packages. The terrain elevation in the detection map can be characterized by a digital elevation model (DEM) to characterize the impact of surface undulations on signal propagation. The obstacle distribution marks the location information of buildings or natural barriers. The electromagnetic environment characteristic information includes the interference source intensity and spectrum occupancy status.

[0200] Preferably, spatial registration converts longitude and latitude into precise projection coordinates of the detection map plane coordinate system through a geographic coordinate conversion algorithm (such as UTM projection), and an association mapping establishes a correspondence between the measurement nodes and the obstacle contours and elevation grids in the map. The electromagnetic sensitive area refers to an area with high interference or susceptible to multipath effects, and the signal blocking area is determined by the obstacle three-dimensional model and ray tracing algorithm.

[0201] Preferably, the edge weight calculation is combined with the geodetic distance between the measurement nodes and the line-of-sight condition. The line-of-sight condition can be understood as whether the line-of-sight propagation is blocked by obstacles. The line-of-sight condition is determined by terrain elevation interpolation and the spatial position of the obstacle.

[0202] Node attributes incorporate terrain elevation values, obstacle occlusion coefficients, and electromagnetic interference intensity levels. The graph neural network node space constrains the strength of information transmission between nodes through edge weights and leverages node attributes to enhance the ability to jointly model spatial topology and electromagnetic environments. For example, in mountainous scenarios, where visibility between nodes is poor due to terrain occlusion, edge weights are reduced. The graph neural network uses this to suppress feature propagation from unreliable nodes. It also integrates elevation attributes to optimize signal attenuation compensation strategies, improving the accuracy of spatial correlation modeling in complex environments.

[0203] This embodiment achieves high-precision spatial registration of geographic location information and detection maps through UTM projection and association mapping, combines geodetic distance calculation and terrain elevation interpolation to determine visibility conditions, and dynamically adjusts the weights of graph neural network edges to constrain the intensity of information transmission between nodes; integrates terrain elevation values, obstacle occlusion coefficients, and electromagnetic interference intensity levels as node attributes to enhance the joint modeling capabilities of spatial topology and electromagnetic environment, solves the problem of insufficient adaptability of traditional methods to complex terrain and multi-source interference, suppresses the propagation of unreliable node features in terrain occlusion scenarios such as mountainous areas, optimizes signal attenuation compensation strategies, and improves the accuracy of spatial association modeling and anti-interference performance.

[0204] In some embodiments, inputting a signal feature set into a graph neural network node space to form a joint signal feature includes:

[0205] Mapping the second processed signal to a graph node corresponding to a measurement node in the graph neural network;

[0206] Based on the edge connection relationship in the node space of the graph neural network, the graph attention network is used to calculate the signal spatial dependency between the measurement nodes. The signal spatial dependency is configured to be represented by spatial attention weights, including:

[0207] Taking the spatial distance and visibility conditions of the measurement nodes as prior knowledge, the correlation score between the graph node features corresponding to each pair of connected measurement nodes is calculated;

[0208] Normalize the relevance scores to generate spatial attention weights;

[0209] Based on the spatial attention weight, the signal features between measurement nodes are fused to obtain spatial enhancement features, including:

[0210] performing weighted aggregation on the second processed signal features of adjacent measurement nodes according to the spatial attention weights;

[0211] retaining the second processed signal characteristics of the current measurement node itself;

[0212] The aggregated adjacent node features are concatenated with the current node features to generate spatial enhancement features;

[0213] The spatial enhancement features are processed at multiple levels to obtain joint signal features, including:

[0214] In the first graph attention layer, the signal features of directly adjacent measurement nodes are fused using spatial attention weights;

[0215] In the second graph attention layer, the signal features of indirectly connected measurement nodes are fused based on the output features of the first layer;

[0216] Maintaining the feature information of the original second processed signal through cross-layer residual connections;

[0217] The fused joint signal feature is output, and the joint signal feature includes the collaborative representation of the second processed signal of each measurement node after the spatial topological relationship is enhanced.

[0218] In this embodiment, the second processed signal is mapped to a graph node in the graph neural network node space via a spatial identifier. The graph attention network dynamically characterizes the dependency strength of signal propagation between nodes using spatial attention weights. The correlation score calculation combines the spatial distance of the measurement nodes with the line-of-sight condition. The spatial distance of the measurement nodes is weighted inversely, and the line-of-sight condition can be quantified using a Boolean variable to determine the line-of-sight obstruction state.

[0219] Preferably, the normalization uses the Softmax function to generate probabilistic attention weights.

[0220] Weighted aggregation refers to the linear superposition of adjacent node features according to weight coefficients, and the splicing operation retains the current node's own features to alleviate over-smoothing.

[0221] In multi-level processing, indirectly connected nodes refer to remote nodes that can be reached by jumping through intermediate nodes. Cross-layer residual connections add the original second-processed signal features to the deep features to prevent gradient disappearance.

[0222] The above steps can be understood with the following example: Nodes A and B in a mountainous area have no direct line of sight due to terrain obstruction, but are indirectly connected through node C. The second graph attention layer can capture the implicit association of the ACB path, and the residual connection ensures that the original spectral features of node A are not lost by high-level features.

[0223] This embodiment constructs dynamic spatial attention weights based on spatial distance and line-of-sight conditions to achieve adaptive quantification of the spatial dependency of signals between nodes; a multi-level graph attention mechanism is used to fuse the features of directly adjacent and indirectly connected nodes to capture implicit propagation paths under complex terrain; cross-layer residual connections retain the original signal features while enhancing deep semantic expression, solving the feature degradation problem caused by traditional single-layer aggregation, improving the robustness of multi-node collaborative representation to geographic occlusion and multipath interference, and providing spatial topology-enhanced input features for high-precision joint signal reconstruction.

[0224] In some embodiments, the joint signal features are input into a trained signal reconstruction model to obtain a reconstructed original signal. The signal reconstruction model is configured as a multi-scale memory network, including:

[0225] Construct a short-term memory module to obtain short-term memory features, including:

[0226] A temporal convolutional network is used to process the joint signal features. The temporal convolutional network captures the local waveform features by dilating the convolution kernel, which is expressed by formula (1). Formula (1) is as follows:

[0227] ST = TCN(J);

[0228] In formula (1), ST is the short-term memory feature, J is the joint signal feature, and TCN(J) is the temporal convolution operation;

[0229] Construct a long-term memory module to obtain long-term memory features, including:

[0230] The Transformer encoder is used to model the dependencies across measurement nodes, which is expressed by formula (2). Formula (2) is as follows:

[0231]

[0232] In formula (2), LT is the long-term memory feature, F tr (J+F at (Q,K,V)) is the encoding function of Transformer, F at (Q, K, V) is the expression of dependency relationship, is the attention mechanism function, Q is the query matrix, K is the key matrix, V is the value matrix, d is the feature dimension, K T is the transposed matrix of the key matrix;

[0233] Build a prototype memory library and obtain retrieval prototype features, including:

[0234] Construct an updateable signal category prototype set, and retrieve the signal category prototype in the signal category prototype set through the attention mechanism, which is expressed by formula (3). Formula (3) is as follows:

[0235]

[0236] In formula (3), PT is the retrieval prototype feature, α l is the contribution weight of the lth signal category prototype in signal reconstruction, m l is the prototype feature of the prototype of the l-th signal category, sim(ST,m l ) is the cosine similarity function, exp(sim(ST,m l )) is ST and m l The matching degree, ∑ p exp(sim(ST,m p )) is normalized, p is the counting unit, m P ∈M,M={m1,m2,…,m l}, M is the prototype set of signal categories;

[0237] The short-term memory feature, long-term memory feature and retrieval prototype feature are subjected to feature fusion and signal reconstruction to obtain the reconstructed original signal, which is expressed by formula (4). Formula (4) is as follows:

[0238]

[0239] In formula (4), G is the fusion feature of short-term memory feature, long-term memory feature and retrieval prototype feature. To reconstruct the original signal, F de (G) is the convolution function.

[0240] In this embodiment, Q, K, and V in formula (2) are generated by linear transformation and expressed as Q = JF·WQ , K=JF·W K , V=JF·W V , where W Q 、W K 、W V is a trainable weight matrix; the expression F for calculating the dependency at (Q, K, V), and then output the long-term memory feature LT.

[0241] The short-term memory module expands the receptive field in the time dimension through the dilated convolution kernel of the temporal convolution operation TCN(J), capturing local waveform patterns (such as rising edges of pulses or periodic jitter) in the joint signal features. Its dilation step size is adaptively adjusted according to the signal sampling rate to balance detail preservation and computational efficiency. The Transformer encoder of the long-term memory module uses a self-attention mechanism to model global dependencies across measurement nodes. The query matrix Q, key matrix K, and value matrix V are generated by linearly transforming the joint signal features. The attention weight calculation focuses on the synergy of signal propagation between nodes (such as spectral complementarity or phase synchronization).

[0242] The signal category prototype set in the prototype memory database stores typical signal patterns (such as frequency modulation and frequency hopping signal primitives). The prototype features are retrieved by calculating the matching degree between the short-term memory features and each prototype through cosine similarity. The contribution weight α l Reflects the correlation strength between the current signal and the historical prototype.

[0243] Feature fusion combines multi-dimensional features into fusion features G and convolution function F de (G) Signal waveform reconstruction is achieved through deconvolution or transposed convolution operations. For example, in a burst interference scenario, the prototype memory library guides the reconstruction model to suppress noise and restore the true signal shape by retrieving similar interference pattern prototypes.

[0244] This embodiment uses a temporal convolutional network to capture local waveform details and a Transformer encoder to model global node dependencies, thereby achieving multi-scale feature extraction of signals. The prototype memory dynamically retrieves and matches prototypes based on historical signal patterns, enhancing the model's generalization ability for known interference types. The feature fusion mechanism combines short-term fluctuations, long-range correlations, and prior knowledge to solve the problems of detail loss or over-smoothing caused by single-scale modeling, thereby improving the waveform fidelity and anti-interference robustness of weak signal reconstruction in complex electromagnetic environments.

[0245] In some embodiments, training the signal reconstruction model using the few-shot adversarial model includes:

[0246] Build a small sample training module, including:

[0247] A support set is constructed by randomly selecting several signal category prototypes from the signal category prototype set. A preset number of samples are provided for each category to form a small sample training task. The support set is used to simulate the small sample conditions in actual scenarios.

[0248] Calculate the prototype center of each signal category prototype in the support set. The prototype center is obtained by aggregating the features of the support set samples through the feature extraction function of the multi-scale memory network, and is used to establish the baseline feature representation of the signal category.

[0249] Prototype loss is calculated based on the prototype center, which minimizes the distance between the same type of signal features and the prototype center, maximizes the distance between different types of signal features, and enhances the ability to distinguish signal categories.

[0250] Build an adversarial training module, including:

[0251] Generate adversarial samples based on the original training samples. The adversarial samples are obtained by calculating the gradient of the loss function with respect to the input sample and adding a perturbation of a limited magnitude along the gradient direction.

[0252] The adversarial sample is input into the multi-scale memory network, and the adversarial loss between the reconstructed signal and the true signal is calculated. The adversarial loss is used to evaluate the robustness of the model to interference signals. By maximizing the adversarial loss, the model's anti-interference ability is improved.

[0253] Execute the joint optimization process until training is complete, including:

[0254] Alternate between small-sample training in the small-sample training module and adversarial training in the adversarial training module. Small-sample training adapts the model to small-sample conditions by optimizing the prototype loss, while adversarial training improves model robustness by optimizing the adversarial loss.

[0255] The objective function is constructed by combining prototype loss, adversarial loss and reconstruction loss. The reconstruction loss is used to ensure the fidelity of signal reconstruction.

[0256] Dynamically adjust the weight parameters corresponding to the combined prototype loss, adversarial loss, and reconstruction loss, and adaptively balance the optimization intensity of different training objectives according to the training stage.

[0257] In this embodiment, the prototype center is used as the feature benchmark of the signal category to establish a stable feature representation space under small sample conditions; the adversarial sample is used to simulate the interference signal in the actual environment, and the stability of the model in a complex electromagnetic environment is improved through adversarial training; the joint optimization process achieves the coordinated improvement of signal reconstruction accuracy and model robustness.

[0258] The support set is a set of several category prototypes and their associated samples randomly selected from the signal category prototype set. The preset number of each category is set according to the scarcity of labeled samples available in the actual scenario, and is used to simulate the model training environment under small sample conditions.

[0259] The prototype center performs feature aggregation (such as mean pooling or attention-weighted averaging) on ​​the support set samples through the feature extraction function of the multi-scale memory network (such as the combination of temporal convolution and Transformer encoder) to form a baseline feature that characterizes the commonality of the signal category.

[0260] Prototype loss enhances the model's ability to distinguish categories under small sample conditions by calculating the cosine similarity (maximization) between the features of similar samples and the prototype center and the Euclidean distance (maximization) between the features of inter-class samples.

[0261] Adversarial examples are generated by calculating the gradient direction of the model reconstruction loss for the original sample input and superimposing a signed perturbation on the gradient. The perturbation amplitude is controlled by a preset adversarial strength coefficient. The adversarial loss measures the mean squared error between the reconstructed signal of the adversarial example and the true signal. Maximizing this error forces the model to learn an interference suppression strategy.

[0262] During the joint optimization process, the reconstruction loss is a combination of a time-domain waveform similarity metric and a frequency-domain energy distribution consistency constraint. The objective function weight parameters are dynamically adjusted according to the training phase (e.g., initial emphasis on reconstruction fidelity, mid- to late-stage enhancements to adversarial robustness), adaptively balancing the synergistic relationship between small-sample learning and anti-interference optimization. For example, when a certain type of signal has only a single-digit number of labeled samples, prototype center aggregation can extract the common spectral features of that type, while adversarial training simulates frequency offset interference to force the model to automatically correct for frequency shift distortion during reconstruction.

[0263] This embodiment establishes signal category benchmark features under small sample conditions through support set construction and prototype center aggregation, and combines prototype loss optimization to enhance the model's feature identification ability for sparsely labeled data; the adversarial sample generation mechanism simulates complex interference environments and improves the model's robust reconstruction performance for disturbed signals by maximizing adversarial loss; the joint optimization process dynamically adjusts the weight parameters of reconstruction loss, prototype loss and adversarial loss to achieve coordinated optimization of signal fidelity, category discrimination and anti-interference ability, solves the problems of overfitting or insufficient generalization caused by traditional single training objectives, and significantly improves the environmental adaptability and stability of the signal reconstruction model in small sample scenarios.

[0264] In some embodiments, obtaining an abnormality ratio in the joint signal feature includes:

[0265] Perform time-frequency domain feature extraction on the joint signal features to obtain a time-frequency distribution feature matrix containing instantaneous frequency, spectral entropy and modulation depth features;

[0266] Calculate the Mahalanobis distance between the signal feature vector in the time-frequency distribution feature matrix and the standard signal vector to obtain the anomaly score corresponding to the signal feature vector;

[0267] Count the number of signal feature vectors whose anomaly scores exceed the dynamic anomaly threshold in the current detection cycle, and record it as the number of anomaly features;

[0268] The ratio of the number of abnormal features to the total number of signal feature vectors in the detection period is calculated as the abnormality ratio.

[0269] In this embodiment, an anomaly detection module is first constructed, and the time-frequency distribution characteristics of the joint signal characteristics are extracted through time-frequency domain features. The instantaneous frequency characterizes the instantaneous change rate of the signal by analyzing the signal phase difference calculation. The spectral entropy measures the complexity of the signal spectral energy distribution, and the modulation depth reflects the depth ratio of the signal amplitude modulation.

[0270] The row vectors of the time-frequency distribution feature matrix correspond to time points, while the column vectors contain the aforementioned multidimensional features. The Mahalanobis distance quantifies the degree to which a signal deviates from a normal pattern by calculating the statistical distance between the signal feature vector and the standard signal vector in covariance space. The covariance matrix is ​​estimated based on historical normal data, and the standard signal vector is derived from a normal signal template. The update process of the normal signal template is synchronized with the update of the signal category prototype set in the aforementioned embodiment.

[0271] Secondly, a dynamic threshold mechanism is established. The dynamic anomaly threshold is adaptively adjusted based on the current electromagnetic environment noise level. The detection period is aligned with the synchronization period of the signal feature set in the aforementioned embodiment to ensure timing consistency. For example, if a node experiences a sudden increase in spectrum entropy due to sudden electromagnetic interference, its Mahalanobis distance anomaly score will exceed the dynamic threshold, triggering an update of the anomaly ratio and retraining of the signal reconstruction model.

[0272] This embodiment extracts multidimensional signal modal characteristics through the time-frequency distribution feature matrix and combines it with the Mahalanobis distance to quantify the degree of signal deviation from the normal mode, thereby solving the problem of insufficient adaptability of traditional single threshold detection to complex signals. The dynamic anomaly threshold is adaptively adjusted according to the noise level of the electromagnetic environment and synchronized with the update cycle of the signal feature set to ensure the timeliness of anomaly detection and environmental matching. The anomaly ratio calculation is linked to the model retraining mechanism to achieve rapid response to abnormal signals and system self-calibration, breaking through the limitations of high false alarm rate or insufficient sensitivity caused by static threshold detection, and improving the accuracy of abnormal signal detection in complex electromagnetic environments and system robustness.

[0273] In some embodiments, updating sample data in the small sample-adversarial mode and triggering the small sample-adversarial mode to retrain the signal reconstruction model after the update includes:

[0274] Based on the detection results whose anomaly ratio exceeds the preset anomaly threshold, the signal feature vectors whose anomaly scores meet the requirements within the current detection cycle are selected as new sample data;

[0275] Add the newly added sample data to the training dataset of the small sample-adversarial model, and remove some of the original sample data in the training dataset to complete the update of the sample data;

[0276] The signal reconstruction model is retrained using the updated training dataset, and the updated signal reconstruction model is used to continue processing the new joint signal features.

[0277] In this embodiment, the screening criteria for newly added sample data include anomaly score and feature distribution consistency; the update ratio of the training data set is dynamically adjusted according to the actual detection results of the anomaly ratio; and the number of iterations of the retraining process is associated with the model performance improvement effect.

[0278] The signal feature vectors whose anomaly scores are higher than the dynamic anomaly threshold and whose consistency with the current feature distribution is consistent are screened. The feature distribution consistency is quantified by calculating the KL divergence or cluster center distance between the new sample and the training dataset features to ensure that the new sample data is compatible with the existing data model.

[0279] The newly added sample data are feature vectors that have significant abnormal characteristics and represent environmental changes during the current detection cycle. The removal of the original sample data follows the first-in-first-out or low-contribution priority principle, and the update ratio is adjusted linearly according to the abnormality ratio.

[0280] Preferably, the number of retraining iterations is dynamically set by monitoring the convergence rate of the validation set reconstruction error. Performance improvement is evaluated by the improvement in the reconstructed signal's signal-to-noise ratio or the reduction in the anomaly detection false alarm rate. For example, when a sudden new interference pattern causes a sharp increase in the anomaly rate, samples with high anomaly scores are prioritized for inclusion in the training set, replacing outdated data to strengthen the model's adaptability to the new interference pattern.

[0281] This embodiment screens new sample data based on the dual standards of anomaly score and feature distribution consistency to ensure that the training set update takes into account both anomaly significance and pattern compatibility; dynamically adjusts the data replacement ratio and links it with the anomaly ratio to achieve environmental adaptive iterative optimization of the training sample set; the retraining process dynamically controls the number of iterations based on the performance improvement effect to solve the overfitting or underconvergence problems caused by fixed training strategies, and enhances the generalization ability and real-time adaptability of the signal reconstruction model to unknown interference by continuously incorporating the latest anomaly features.

[0282] In a second aspect, this embodiment further provides a radio detection sensitivity enhancement system based on AI computing, applicable to the method of the first aspect, comprising:

[0283] Multiple measurement nodes are arranged in a preset area according to a preset method. The preset area is the area to be detected. The measurement nodes are used to obtain original weak signals of the measurement nodes.

[0284] Multiple edge processors, each edge processor is set at a measurement node, and the edge processor is used to pre-process the original weak signal;

[0285] Multiple communication units, each communication unit is set at a measurement node, and the communication unit is used to synchronously upload multiple processed weak signals;

[0286] The server is used to organize the processed weak signals of the same time period into a signal feature set; obtain the geographic location information and detection map of the measurement node, and construct the graph neural network node space; input the signal feature set into the graph neural network node space to form a joint signal feature; input the joint signal feature into the trained signal reconstruction model to obtain the reconstructed original signal. The signal reconstruction model is configured as a multi-scale memory network construction, and the signal reconstruction model is trained using a small sample-adversarial mode; and, obtain the anomaly ratio in the joint signal feature. If the anomaly ratio exceeds the preset anomaly threshold, the sample data in the small sample-adversarial mode is updated, and after the update, the small sample-adversarial mode is triggered to retrain the signal reconstruction model.

[0287] In this embodiment, the preset area is an area to be detected divided according to the complexity of the electromagnetic environment (such as a densely populated urban area or an open suburban area), and the preset layout method of the measurement nodes includes a grid-based uniform deployment or an adaptive layout strategy based on the historical interference source distribution.

[0288] Preferably, the edge processor has a built-in preprocessing algorithm module, including a noise reduction unit, a feature extraction unit and a data encapsulation unit, wherein the noise reduction unit is used to implement the adaptive filtering of the aforementioned method, the feature extraction unit is used to integrate the lightweight convolutional network of the aforementioned method, and the data encapsulation unit is used to perform the spatiotemporal information standardization of the aforementioned method.

[0289] The communication unit uses a timestamp synchronization protocol to ensure the time accuracy of multi-node signal acquisition, and its transmission protocol matches the cycle of synchronously uploading multiple processed weak signals in the aforementioned method.

[0290] Preferably, the server includes a feature set management module, a graph neural network computing engine, a signal reconstruction model library, and an anomaly detection and retraining scheduler, wherein the feature set management module is used to implement the spatiotemporal feature matrix construction of the aforementioned method, the graph neural network computing engine is used to execute the graph attention aggregation of the aforementioned method, the signal reconstruction model library is used to store the parameters of the signal reconstruction model trained in the aforementioned method, and the anomaly detection and retraining scheduler is linked to the update mechanism of sample data in the small sample-adversarial mode.

[0291] For example, when an edge processor detects sudden strong interference, its pre-processed signal is synchronously uploaded to the server through the communication unit. The server integrates multi-node features based on the graph neural network node space and triggers an abnormal ratio exceeding the limit judgment. The scheduler automatically calls the newly added samples to start model retraining. The updated model is distributed to each edge processor through the server to achieve closed-loop optimization.

[0292] This embodiment reduces the redundancy of original signal transmission through adaptive layout of measurement nodes and edge preprocessing, and the precise timing of the communication unit ensures the spatiotemporal alignment of multi-source data; the server integrates the graph neural network computing engine and the dynamic retraining scheduler to realize the full-link systematic deployment of the method described in the first aspect, forming a "end-edge-cloud" collaborative closed-loop optimization architecture, while improving the sensitivity of weak signal detection, ensuring the system's adaptive response capability to complex electromagnetic environment disturbances and new interference modes and cross-regional collaborative processing efficiency.

[0293] By adopting the above technical solution, the present invention is different from the existing technology and has the following beneficial effects:

[0294] The above technical solution constructs a graph neural network node space to fuse the geographic location information and electromagnetic environment characteristics of multiple measurement nodes, and uses the spatial correlation modeling between nodes and the spatiotemporal joint analysis of the signal feature set to break through the limitations of the traditional single-node isolated processing mode that cannot capture the spatiotemporal correlation characteristics of weak signals; combines the multi-level feature extraction capabilities of the multi-scale memory network with the dynamic optimization mechanism of small sample-adversarial training to achieve high-fidelity reconstruction of weak signals in complex electromagnetic environments; triggers sample data updates and model retraining based on anomaly ratio detection to form a closed-loop self-optimization system to solve the static model's lack of adaptive ability to abnormal signal fluctuations and new interference patterns.

[0295] At the system level, through adaptive deployment of measurement nodes, edge preprocessing noise reduction, and precise timing of communication units, combined with the server-side graph neural network computing engine and dynamic retraining scheduler, a "end-edge-cloud" collaborative closed-loop processing architecture is constructed. While maintaining high-sensitivity detection performance, it significantly improves the global optimization capabilities of multi-source interference suppression, signal attenuation compensation, and cross-node collaborative analysis, achieving a coordinated enhancement of the anti-interference robustness and environmental adaptability of the radio detection system in complex scenarios.

[0296] In addition, the functional units in various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0297] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the various embodiments of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0298] The above descriptions are only some embodiments of the present invention and do not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for improving the sensitivity of radio detection based on AI computing, characterized in that: include: Acquire an original weak signal from a measurement node, and preprocess the original weak signal to obtain a processed weak signal, wherein the preprocessing includes performing noise reduction, feature extraction, and data encapsulation on the original weak signal; Synchronously uploading the plurality of processed weak signals, and arranging the processed weak signals in the same period into a signal feature set; Obtain the geographic location information and detection map of the measurement nodes and construct the graph neural network node space; Inputting the signal feature set into the graph neural network node space to form a joint signal feature; Inputting the joint signal features into a trained signal reconstruction model to obtain a reconstructed original signal, wherein the signal reconstruction model is configured as a multi-scale memory network and trained using a small sample-adversarial model; Also, obtain the abnormality ratio in the joint signal feature. If the abnormality ratio exceeds the preset abnormality threshold, update the sample data in the small sample-adversarial mode, and trigger the small sample-adversarial mode to retrain the signal reconstruction model after the update.

2. The method for improving the sensitivity of radio detection based on AI computing according to claim 1, characterized in that: Acquiring an original weak signal from a measurement node and preprocessing the original weak signal to obtain a processed weak signal includes: Adopting an adaptive constant false alarm rate detection algorithm to eliminate background noise in the original weak signal to obtain a primary filtered signal, wherein the adaptive constant false alarm rate detection algorithm dynamically adjusts a detection threshold value to keep a false alarm probability constant and calculates a local noise power estimate within a sliding window; performing a joint time-frequency analysis on the primary filtered signal, separating aliased signal components using a Wigner-Ville distribution algorithm to obtain a separated signal, wherein the Wigner-Ville distribution algorithm calculates the energy distribution of the signal on a time-frequency plane and achieves effective separation of signal components through cross-term suppression processing; Performing deep noise suppression processing on the separated signal using a long short-term memory autoencoder to obtain a noise-reduced signal, wherein the long short-term memory autoencoder extracts signal timing features through an encoder and reconstructs a pure signal waveform in a decoder; Performing multi-dimensional feature extraction on the noise reduction signal through a lightweight convolutional neural network to generate a feature vector including time domain features, frequency domain features, and modulation features. The lightweight convolutional neural network uses a channel pruning method to remove redundant convolution kernels and retain key feature extraction channels; The feature vector is encapsulated with the corresponding timestamp and geographic location information to form the processed weak signal in a standard format. The data encapsulation process includes feature dimension normalization and metadata verification steps.

3. The method for improving the sensitivity of radio detection based on AI calculation according to claim 1, characterized in that: Synchronously uploading a plurality of the processed weak signals and arranging the processed weak signals in the same period into a signal feature set includes: Performing time alignment processing based on the timestamp information contained in the processed weak signal to obtain a first processed signal; Establishing a spatial index based on the geographical location information in the processed weak signal to determine the relative position relationship of each measurement node; Sorting the first processed signals according to the spatial index to construct a spatiotemporal correlation feature matrix; Performing integrity check on the feature matrix to remove missing or abnormal data entries; The feature matrix is ​​converted into a signal feature set output, where the signal feature set includes a plurality of second processed signals.

4. The method for improving the sensitivity of radio detection based on AI calculation according to claim 3, characterized in that: Obtain the geographic location information and detection map of the measurement node and construct the graph neural network node space, including: Obtaining geographic location information from the processed weak signal and loading a preset detection map, wherein the detection map includes terrain elevation, obstacle distribution, and electromagnetic environment characteristic information; The geographic location information is spatially aligned with the detection map to obtain alignment features, including: Calculate the precise projection coordinates of geographic location information in the detection map; Establishing a correlation mapping between geographic location information and map features; Mark electromagnetic sensitive areas and signal blocking areas; Generating the registration features according to the precise projection coordinates, the correlation map, the electromagnetic sensitive area, and the signal blocking area; A graph neural network node space is constructed based on the registration features to obtain a graph neural network node space containing spatial topology and electromagnetic characteristics, including: Each measurement node is a graph node, and the spatial relationship between measurement nodes is an edge connection relationship; Calculate the edge weight based on the distance between the measured nodes and the line-of-sight condition; The map features in the fused detection map are used as node attributes.

5. The method for improving the sensitivity of radio detection based on AI calculation according to claim 4, characterized in that: The signal feature set is input into the graph neural network node space to form a joint signal feature including: Mapping the second processed signal to a graph node corresponding to a measurement node in a graph neural network; Based on the edge connection relationship in the node space of the graph neural network, a graph attention network is used to calculate the signal space dependency between the measurement nodes, and the signal space dependency is configured to be represented by a spatial attention weight, including: Taking the spatial distance and visibility conditions of the measurement nodes as prior knowledge, the correlation score between the graph node features corresponding to each pair of connected measurement nodes is calculated; Normalize the relevance scores to generate spatial attention weights; Based on the spatial attention weight, signal features between measurement nodes are fused to obtain spatial enhancement features, including: Performing weighted aggregation on the second processed signal features of adjacent measurement nodes according to the spatial attention weight; retaining the second processed signal characteristics of the current measurement node itself; The aggregated adjacent node features are concatenated with the current node features to generate spatial enhancement features. The spatial enhancement features are subjected to multi-level processing to obtain joint signal features, including: In the first graph attention layer, the signal features of directly adjacent measurement nodes are fused using the spatial attention weights; In the second graph attention layer, the signal features of indirectly connected measurement nodes are fused based on the output features of the first layer; Maintaining the feature information of the original second processed signal through cross-layer residual connections; The fused joint signal feature is output, where the joint signal feature includes a collaborative representation of the second processed signals of each measurement node after the spatial topological relationship is enhanced.

6. The method for improving the sensitivity of radio detection based on AI calculation according to claim 1, characterized in that: The joint signal features are input into a trained signal reconstruction model to obtain a reconstructed original signal, wherein the signal reconstruction model is configured as a multi-scale memory network construction, including: Construct a short-term memory module to obtain short-term memory features, including: A temporal convolutional network is used to process the joint signal features. The temporal convolutional network captures local waveform features by dilating the convolution kernel, which is expressed by formula (1). The formula (1) is as follows: ST = TCN(J); In formula (1), ST is the short-term memory feature, J is the joint signal feature, and TCN(J) is the temporal convolution operation; Construct a long-term memory module to obtain long-term memory features, including: The Transformer encoder is used to model the dependencies across measurement nodes, which is expressed by formula (2). The formula (2) is as follows: In formula (2), LT is the long-term memory feature, F tr (J+F at (Q,K,V)) is the encoding function of Transformer, F at (Q, K, V) is the expression of dependency relationship, is the attention mechanism function, Q is the query matrix, K is the key matrix, V is the value matrix, d is the feature dimension, K T is the transposed matrix of the key matrix; Build a prototype memory library and obtain retrieval prototype features, including: An updateable signal category prototype set is constructed, and the signal category prototype in the signal category prototype set is retrieved through the attention mechanism, which is expressed by formula (3). The formula (3) is as follows: In formula (3), PT is the retrieval prototype feature, α l is the contribution weight of the lth signal category prototype in signal reconstruction, m l is the prototype feature of the prototype of the l-th signal category, sim(ST,m l ) is the cosine similarity function, exp(sim(ST,m l )) is ST and m l The matching degree, ∑ p exp(sim(ST,m p )) is normalized, p is the counting unit, m P ∈M,M={m1,m2,…,m l }, M is the prototype set of signal categories; The short-term memory feature, the long-term memory feature and the retrieval prototype feature are subjected to feature fusion and signal reconstruction to obtain the reconstructed original signal, which is expressed by formula (4). The formula (4) is as follows: In formula (4), G is the fusion feature of short-term memory feature, long-term memory feature and retrieval prototype feature. To reconstruct the original signal, F de (G) is the convolution function.

7. The method for improving the sensitivity of radio detection based on AI calculation according to claim 6, characterized in that: Training the signal reconstruction model using the small sample-adversarial model includes: Build a small sample training module, including: Randomly selecting a number of signal category prototypes from the signal category prototype set to construct a support set, providing a preset number of samples for each category to form a small sample training task, and the support set is used to simulate the small sample condition in the actual scenario; Calculate the prototype center of each signal category prototype in the support set. The prototype center is obtained by aggregating the support set samples through the feature extraction function of the multi-scale memory network, and is used to establish a baseline feature representation of the signal category. Calculating the prototype loss based on the prototype center, by minimizing the distance between the same type of signal features and the prototype center and maximizing the distance between different types of signal features, thereby enhancing the signal category differentiation capability; Build an adversarial training module, including: Generate adversarial samples based on the original training samples. The adversarial samples are obtained by calculating the gradient of the loss function with respect to the input sample and adding a perturbation of a limited magnitude along the gradient direction. The adversarial sample is input into the multi-scale memory network, and the adversarial loss between the reconstructed signal and the true signal is calculated. The adversarial loss is used to evaluate the robustness of the model to interference signals. The anti-interference ability of the model is improved by maximizing the adversarial loss. Execute the joint optimization process until training is complete, including: Alternately perform small sample training in the small sample training module and adversarial training in the adversarial training module. The small sample training adapts the model to small sample conditions by optimizing the prototype loss, and the adversarial training improves the robustness of the model by optimizing the adversarial loss. The objective function is constructed by combining prototype loss, adversarial loss and reconstruction loss, where the reconstruction loss is used to ensure the fidelity of signal reconstruction; Dynamically adjust the weight parameters corresponding to the combined prototype loss, adversarial loss, and reconstruction loss, and adaptively balance the optimization intensity of different training objectives according to the training stage.

8. The method for improving the sensitivity of radio detection based on AI calculation according to claim 1, characterized in that: Obtaining the abnormality ratio in the joint signal feature includes: Extracting the time-frequency domain features of the joint signal features to obtain a time-frequency distribution feature matrix including instantaneous frequency, spectral entropy and modulation depth features; Calculating the Mahalanobis distance between the signal feature vector in the time-frequency distribution feature matrix and the standard signal vector to obtain an anomaly score corresponding to the signal feature vector; Count the number of signal feature vectors whose anomaly scores exceed the dynamic anomaly threshold in the current detection cycle, and record it as the number of anomaly features; The ratio of the number of abnormal features to the total number of signal feature vectors in the detection period is calculated as the abnormality ratio.

9. The method for improving the sensitivity of radio detection based on AI calculation according to claim 8, characterized in that: Updating the sample data in the small sample-adversarial mode and triggering the small sample-adversarial mode to retrain the signal reconstruction model after the update includes: Based on the detection result that the anomaly ratio exceeds the preset anomaly threshold, the signal feature vectors whose anomaly scores meet the requirements in the current detection cycle are selected as new sample data; Adding the newly added sample data to the training data set of the small sample-adversarial mode, and removing part of the original sample data in the training data set to complete the update of the sample data; The signal reconstruction model is retrained using the updated training data set, and the updated signal reconstruction model is used to continue processing new joint signal features.

10. A radio detection sensitivity enhancement system based on AI computing, characterized in that: The method according to any one of claims 1 to 9, wherein the system comprises: A plurality of measurement nodes are arranged in a preset area according to a preset manner, wherein the preset area is an area to be detected, and the measurement nodes are used to obtain original weak signals of the measurement nodes; A plurality of edge processors, each of the edge processors is arranged at one of the measurement nodes, and the edge processor is used to pre-process the original weak signal; a plurality of communication units, each of the communication units being arranged at one of the measurement nodes, and the communication unit being configured to synchronously upload the plurality of processed weak signals; The server is used to organize the processed weak signals of the same time period into a signal feature set; obtain the geographic location information and detection map of the measurement node, and construct a graph neural network node space; input the signal feature set into the graph neural network node space to form a joint signal feature; input the joint signal feature into the trained signal reconstruction model to obtain the reconstructed original signal, the signal reconstruction model is configured to be constructed as a multi-scale memory network, and the signal reconstruction model is trained using a small sample-adversarial mode; and, obtain the anomaly ratio in the joint signal feature. If the anomaly ratio exceeds the preset anomaly threshold, the sample data in the small sample-adversarial mode is updated, and the small sample-adversarial mode is triggered after the update to retrain the signal reconstruction model.

Images