Radar target analytic calculation method based on multi-dimensional data fusion and radar device

Abstract

The invention relates to the technical field of radar signal processing, in particular to a radar target analytical calculation method based on multi-dimensional data fusion and a radar device. Comprising the following steps: deploying a multi-band radar sensor array comprising an X band, a C band and a Ku band in a radar monitoring area; performing pulse compression and Doppler processing on the time domain echo signal, and extracting a time domain feature; spectral analysis is carried out on the frequency domain signals, and frequency domain features are extracted; performing angle estimation on the spatial signals, and extracting spatial features; a dynamic weight distribution model is constructed, a fusion weight is calculated through an adaptive algorithm based on three-dimensional quality indexes of a real-time signal-to-noise ratio (SNR), feature stability (SI) and data integrity (CI), and a joint representation vector containing time domain, frequency domain and space multi-dimensional information is generated. According to the invention, by deploying the multi-band radar sensor array, the recognition capability of the subtle feature difference of the target is improved.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing technology, and in particular to a radar target analysis calculation method based on multi-dimensional data fusion and a radar device. Background Art

[0002] Modern radar systems undertake the core tasks of target detection, identification, and tracking in military reconnaissance, air traffic control, intelligent transportation, and other fields. With the increasing diversity of target characteristics and dynamic scenarios in complex electromagnetic environments, traditional single-band, single-mode radars face the challenge of insufficient multi-dimensional information fusion.

[0003] For example, Chinese patent CN202211634430.X discloses a method for obtaining geometric parameters of hidden targets based on multiple resonance information of ground penetrating radar echoes; the invention fully explores the multiple resonance echo mechanism generated by ground penetrating radar echo detection targets, establishes a target geometric parameter acquisition model based on internal harmonics and external harmonics, and can directly obtain the geometric parameter information of hidden targets from the time domain diagram of multiple resonance echoes, providing a new method for ground penetrating radar target detection and acquisition, and effectively solving technical problems such as the difficulty of target analysis and geometric parameter acquisition in existing geological radar echoes; for example, Chinese patent CN201910589162.6 discloses a method, device and radar system for fusing AIS targets and radar targets; the invention obtains the latitude and longitude of radar targets and AIS targets Position information and obtain the heading and speed information of the radar target; judge the longitude and latitude information of the radar target and the longitude and latitude information of the AIS target; if the longitude and latitude information of the target is greater than the preset position threshold, then determine that the radar target and the AIS target are not the same target; if the longitude and latitude information of the target is not greater than the preset position threshold, then further judge the movement trend of the heading and speed of the radar target and the AIS target; if the movement trends are consistent, then determine that the radar target and the AIS target are the same target; if the movement trends are inconsistent, then determine that the radar target and the AIS target are not the same target; if the radar target and the AIS target are the same target, then fuse and display the radar target and the AIS target as one target; if the radar target and the AIS target are not the same target, then display the radar target and the AIS target separately;

[0004] Although the aforementioned comparative patents all have their respective advantages, they all have the following technical problems: 1. The depth of multi-dimensional feature fusion is insufficient. Existing methods mostly use traditional signal processing or simple machine learning models (such as support vector machines) to extract features, making it difficult to capture the temporal dependence of time domain waveforms, the nonlinear distribution of frequency domain spectra, and the geometric correlation of spatial signals. For example, patent CN202211634430.X only obtains geometric parameters through time domain echo analysis, and does not explore the value of frequency domain features (such as center frequency and bandwidth) for target material identification; 2. Existing methods mostly use traditional signal processing or simple machine learning models (such as support vector machines) to extract features, making it difficult to capture the temporal dependence of time domain waveforms, the nonlinear distribution of frequency domain spectra, and the geometric correlation of spatial signals. For example, patent CN202211634430.X only obtains geometric parameters through time-domain echo analysis, and does not explore the value of frequency-domain features (such as center frequency and bandwidth) for target material identification. 3. Lack of spatiotemporal correlation modeling: Traditional methods often treat spatiotemporal information as independent dimensions and do not establish a dynamic graph structure to capture the interactive relationship between target motion trajectory and multi-dimensional features. For example, patent CN201910589162.6 only determines target association by position and motion trend, and does not consider the impact of the temporal evolution of time-domain features (such as pulse repetition interval) and frequency-domain features (such as Doppler shift) on target classification. In view of the above, how to achieve accurate analysis of target type, motion parameters and threat level in complex environments through deep fusion of multi-dimensional data, dynamic weight allocation and spatiotemporal correlation modeling in multi-band radar collaborative detection scenarios is a core issue that needs to be solved in the current field of radar technology. Summary of the Invention

[0005] The purpose of the present invention is to provide a radar target analysis calculation method and a radar device based on multi-dimensional data fusion to solve the problems raised in the above background technology.

[0006] To solve the above technical problems, one of the objectives of the present invention is to provide a radar target parsing calculation method based on multi-dimensional data fusion, comprising the following steps:

[0007] S100, deploying a multi-band radar sensor array including X-band, C-band, and Ku-band in the radar monitoring area, wherein the sensor array includes a phased array antenna unit, a signal receiving module, and a synchronization trigger device;

[0008] S200, realizing collaborative acquisition of multi-band radar data through a clock synchronization mechanism, obtaining multi-dimensional echo signals including time domain waveform, frequency domain spectrum, and spatial orientation, and establishing a three-dimensional mapping relationship between time, frequency, and space;

[0009] S300, performing pulse compression and Doppler processing on the time domain echo signal and extracting time domain features (such as pulse width, pulse repetition interval, etc.); performing spectrum analysis on the frequency domain signal and extracting frequency domain features (such as center frequency, bandwidth, etc.); performing angle estimation on the spatial signal and extracting spatial features (such as azimuth angle, elevation angle, etc.);

[0010] S400, through the deep learning model, extracts three types of single-dimensional deep features in the time domain, frequency domain, and space domain:

[0011] S500, randomly combining any two of the three types of single-dimensional deep features in the time domain, frequency domain, and space domain to form a two-dimensional feature pair, and performing normalization and feature alignment on the two-dimensional feature pair, eliminating dimensional differences through Z-score normalization, and generating a two-tuple feature vector containing the correlation between the dimensions;

[0012] S600: Build a dynamic weight allocation model. Based on the three-dimensional quality indicators of real-time signal-to-noise ratio (SNR), feature stability (SI), and data integrity (CI), calculate the fusion weight through an adaptive algorithm to generate a joint representation vector containing multi-dimensional information in the time domain, frequency domain, and space domain.

[0013] S700, mapping the single-dimensional deep features, the binary feature vectors, and the joint representation vectors into a spatiotemporal tensor, defining a dynamic graph structure including single-modal nodes, bimodal fusion nodes, and full-modal global nodes, where the edge connections between nodes are determined by feature cosine similarity and temporal correlation;

[0014] S800, aggregates neighborhood features through a spatiotemporal multi-head attention mechanism. Each attention head independently calculates the correlation weights between dimensions, combines residual connections with layer normalization to optimize feature expression, and generates a target feature vector containing spatiotemporal correlations.

[0015] S900: Based on the spatiotemporal joint feature expression, the target type probability distribution is calculated through a multi-level classifier, similarity matching is performed in combination with the target feature knowledge base, and the target type recognition result is output;

[0016] S1000. Generate a structured decision report including target type, motion parameters and threat level, and output a visual target situation map with a data traceability chain.

[0017] As a further improvement to the present technical solution, the multi-band radar sensor array in S100 synchronously collects echo signals using a hybrid technology of time division multiplexing (TDM) and frequency division multiplexing (FDM), including the following steps:

[0018] S110, based on time division multiplexing technology, uses FPGA high-precision timers to achieve microsecond time slot division (such as X-band 0-5ms, C-band 5-10ms, Ku-band 10-15ms), or generates multi-band trigger signals through the TICDCE913 synchronization chip to achieve orthogonal sampling in the time domain of multi-band radar;

[0019] S120, spectrum resource allocation based on frequency division multiplexing technology introduces orthogonal frequency division multiplexing (OFDM) technology to improve spectrum utilization through subcarrier division, and adopts software-defined radio (SDR) architecture to achieve dynamic frequency band configuration and build a two-dimensional orthogonal signal acquisition framework in time and frequency.

[0020] S130, perform adaptive sampling frequency control, first ensure that the sampling frequency meets the Nyquist sampling criterion to avoid signal aliasing, and then predict the target motion parameters (such as position, velocity vector, acceleration) through the Kalman filter algorithm according to the target motion state (such as speed, acceleration) to optimize the sampling frequency in real time.

[0021] As a further improvement of the present technical solution, in S400, three types of single-dimensional deep features in the time domain, frequency domain and space are extracted respectively through a deep learning model, which specifically includes:

[0022] Use the long short-term memory network (LSTM) to model the time domain signal, capture the temporal dependencies in the sequence data, and generate LSTM feature vectors;

[0023] Process frequency domain signals (such as spectrograms, time-frequency graphs, etc.) through convolutional neural networks (CNNs), extract their spectral distribution characteristics and local patterns, and output CNN feature vectors;

[0024] For spatial signals (such as the coordinate sequence and geometric topological structure of the motion trajectory), a graph neural network (GNN) or a geometric deep learning model is used to learn the spatial geometric features of the motion trajectory (such as positional relationship, shape description, trajectory change pattern, etc.) and obtain the geometric feature vector.

[0025] In the S400:

[0026] The time domain feature extraction network adopts a bidirectional gated recurrent unit (Bi-GRU) structure, which extracts time-dependent features by stacking two layers of hidden units (128 neurons per layer) and setting a dropout layer of 0.2.

[0027] The frequency domain feature extraction network adopts the ResNet-18 residual architecture, which realizes hierarchical mapping of spectral features through four residual convolution blocks containing 3×3 convolution kernels and batch normalization (BN) layers.

[0028] As a further improvement of the present technical solution, the dynamic weight allocation model in S600 calculates the fusion weight through a three-layer quality assessment system, including the following steps:

[0029] S610, using the periodogram method to estimate the signal power spectral density and calculate the real-time signal-to-noise ratio (SNR); calculating the feature stability (SI) by the rate of change of the cosine similarity of the feature vectors within a sliding window of 3-10 (preferably 5) scanning cycles; using the Kalman filter to predict missing data points and combining the IQR outlier detection algorithm to calculate the data integrity (CI);

[0030] S620: Calculate the initial fusion weight using the weighted product method, adjust the weight sensitivity using the exponential normalization algorithm, and generate the final weight vector.

[0031] As a further improvement of this technical solution, the construction of the space-time tensor in S700 includes the following steps:

[0032] S710, performing dimension mapping on the time domain features (such as pulse width and repetition interval sequence) to generate a time dimension tensor with the scanning period (10 periods) as the time axis and the time domain features (8 dimensions) as the feature axis;

[0033] S720. Perform dimension mapping on the frequency domain features (such as Doppler shift and spectrum energy) to generate a frequency dimension tensor with frequency points (256 points) as the frequency axis and frequency domain features (12 dimensions) as the feature axis.

[0034] S730: Dimensionally map the spatial features (such as azimuth, elevation, and distance) to generate a spatial dimension tensor with the spatial grid (divided into 64×64 pixels) as the spatial axis and the spatial features (6 dimensions) as the feature axis.

[0035] S740. Synthesize the time, frequency, and space dimension tensors into a three-dimensional space-time tensor according to the correspondence between period, frequency bin, and spatial grid, where each tensor element aggregates multi-dimensional features corresponding to the time period, frequency interval, and spatial position.

[0036] As a further improvement of this technical solution, the spatiotemporal attention mechanism in S800 adopts a graph attention network architecture, including the following steps:

[0037] S811, for each node feature , respectively by querying the weight matrix , key weight matrix , value weight matrix Generate the corresponding 64-dimensional query vector , key vector , value vector ;

[0038] S812: Node With neighboring nodes , through the formula , calculate the attention coefficient, where, Representation node For neighbor nodes The attention weight of For nodes The query vector (QueryVector) is used to capture the query representation of node features; For nodes The key vector is used to calculate the attention score with the query vector; For nodes The first-order neighbor set (i.e. the set of directly connected nodes), For nodes The first-order neighbor number index of ; For nodes The key vector of Represents the transpose of a vector matrix; It is a leaky rectified linear unit activation function used to enhance the nonlinear expression of attention scores;

[0039] S813: Passed , aggregate neighbor features, and introduce residual connection and layer normalization (LayerNorm) to optimize feature expression; among them, For nodes Output feature vector after aggregating neighbor features; An activation function (such as Sigmoid, ReLU, etc., which depends on the model configuration) is used to perform nonlinear transformation on the aggregation result; For nodes The feature vector represents the original features of the neighboring nodes or the features processed by the previous layer.

[0040] As a further improvement of this technical solution, the spatiotemporal multi-head attention mechanism in S800 is collaboratively calculated by 8 parallel attention heads, including the following steps:

[0041] S821. Perform multi-head feature mapping on the 36-dimensional input features, and generate 8 sets of 64-dimensional query, key, and value matrices through 8 sets of independent weight matrices;

[0042] S822. Perform single-head attention calculation on each set of feature matrices, use the Softmax function to normalize the dot product of the query and the key, and obtain the single-head attention output;

[0043] S823, concatenate 8 single-head outputs and map them back to a 36-dimensional feature vector through the output matrix to complete the multi-head attention feature fusion.

[0044] As a further improvement of the present technical solution, the target feature knowledge base in S900 is constructed by an improved K-means++ clustering algorithm, including the following steps:

[0045] S911. Perform Z-score normalization on the 12-dimensional time-frequency features. The calculation formula is: , and reduced to 6 dimensions through PCA; is the characteristic mean, is the standard deviation, normalizing the features to the standard normal distribution; is the eigenvalue after Z-score standardization;

[0046] S912, through the Euclidean distance formula ), measure the similarity between the sample and the center, and iteratively generate clustering templates; where, For the sample, is the cluster center; Representation sample and cluster centers the distance between them; Represents the sum index variable;

[0047] S913: Use the cosine similarity algorithm to determine the matching degree between the new sample and the template, and trigger the template update mechanism.

[0048] As a further improvement of the present technical solution, the multi-level classifier in S900 adopts a two-stage decision architecture, including the following steps:

[0049] S911: Input spatial features (azimuth angle change rate, velocity vector, acceleration) into the SVM classifier, use the RBF kernel function to distinguish low-altitude targets (change rate > 10° / s) from high-altitude targets (change rate < 5° / s), and output the coarse classification results.

[0050] S912: The fused 12-dimensional time-frequency features (such as pulse width and Doppler shift peak) are input into the random forest classifier, which outputs the target type probability distribution through a voting mechanism of 500 decision trees.

[0051] S913: Calculate the cosine similarity between the current feature and the knowledge base template (threshold ≥ 0.92), correct the classification probability and output the final target type (such as "multi-rotor drone") and confidence level (such as 0.98).

[0052] A second object of the present invention is to provide a radar device system based on multi-dimensional data fusion, which is loaded with a corresponding computer program and can execute the steps of the radar target analysis calculation method based on multi-dimensional data fusion as described above.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. The present invention deploys a multi-band radar sensor array to synchronously collect multi-dimensional echo signals in the time domain, frequency domain, and spatial orientation, and construct a three-dimensional mapping relationship between time, frequency, and space. It can extract features including time-series dependencies, spectral characteristics, and geometric features of motion trajectories to form a multi-dimensional joint representation vector, thereby improving the ability to identify subtle differences in target features, and is particularly suitable for small, slow, and low-lying targets such as birds and drones. At the same time, through a dynamic graph structure and a spatiotemporal multi-head attention mechanism, it achieves deep interaction between time, frequency, and spatial features, effectively distinguishing the motion characteristics and electromagnetic scattering differences of different targets, and significantly reducing the target misjudgment rate in complex electromagnetic environments.

[0055] 2. This invention is based on the three-dimensional indicators of real-time signal-to-noise ratio, feature stability, and data integrity, and dynamically adjusts the fusion weights of each dimension through a sliding window mechanism. When a certain frequency band radar is interfered with, the model can automatically reduce the weight of that dimension to avoid noise data interference, thereby improving the system's anti-interference performance and data fusion stability. For different application scenarios, the dynamic weight allocation model can adaptively adjust the contribution of each dimensional feature, accurately distinguish different target types, effectively suppress clutter interference such as ground reflections, and reduce the false alarm rate of target detection.

[0056] 3. This invention uses spatiotemporal tensors and dynamic graph structures to align the target's multi-dimensional features with its motion trajectory in spatiotemporal order, combining temporal dependency with motion trajectory modeling to accurately predict changes in the target's motion state and improve the accuracy of motion parameter estimation. The spatiotemporal attention network effectively solves the problem of spatiotemporal information fragmentation in traditional methods. In scenarios where multiple targets intersect or are blocked, it can distinguish the feature interaction patterns of different targets in real time, thereby improving target resolution and tracking continuity. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is an overall step diagram of the present invention. DETAILED DESCRIPTION

[0058] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0059] Example 1

[0060] like Figure 1 As shown, this embodiment provides a radar target parsing calculation method based on multi-dimensional data fusion, including the following steps:

[0061] S100. Deploy a multi-band radar sensor array encompassing X-band, C-band, and Ku-band in the radar monitoring area. The sensor array comprises a phased array antenna unit, a signal receiving module, and a synchronization trigger device. The X-band, C-band, and Ku-band cover different frequency band characteristics (e.g., X-band high resolution, Ku-band strong resistance to rain attenuation). Synchronous acquisition through TDM or FDM technology can complement each other to obtain multi-dimensional target information, enhancing target perception capabilities in complex environments.

[0062] In this embodiment, the multi-band radar sensor array in S100 synchronously collects echo signals using a hybrid technology of time division multiplexing (TDM) and frequency division multiplexing (FDM), including the following steps:

[0063] S110, based on time division multiplexing technology, uses FPGA high-precision timers to achieve microsecond time slot division (such as X-band 0-5ms, C-band 5-10ms, Ku-band 10-15ms), or generates multi-band trigger signals through the TICDCE913 synchronization chip to achieve orthogonal sampling in the time domain of multi-band radar;

[0064] S120, spectrum resource allocation based on frequency division multiplexing technology introduces orthogonal frequency division multiplexing (OFDM) technology to improve spectrum utilization through subcarrier division, and adopts software-defined radio (SDR) architecture to achieve dynamic frequency band configuration and build a two-dimensional orthogonal signal acquisition framework in time and frequency.

[0065] S130, perform adaptive sampling frequency control, first ensure that the sampling frequency meets the Nyquist sampling criterion to avoid signal aliasing, and then predict the target motion parameters (such as position, velocity vector, acceleration) through the Kalman filter algorithm according to the target motion state (such as speed, acceleration) to optimize the sampling frequency in real time.

[0066] As a further illustration of this embodiment, a layered cascade phased array antenna array is used in this embodiment, and the radar antenna units of each band are independently designed according to the function and performance requirements:

[0067] X-band radar (high-precision close-range detection): uses an 8×8 planar phased array antenna with a unit spacing of half a wavelength, forming a narrow beam with an azimuth angle of ±30° and an elevation angle of ±15°, covering a distance range of 0-500 meters, and supporting microsecond-level beam agility (such as dynamically adjusting the beam pointing when tracking a target).

[0068] C-band radar (medium-range multi-target detection): uses a 4×4 planar phased array antenna with a unit spacing of λ (λ≈6cm, spacing 6cm), a beam width of ±45° in azimuth and ±20° in elevation, covering a distance of 500-2000 meters, and has the ability to simultaneously scan multiple beams (such as time-division multiplexing to generate three independent beams, covering the front, left, and right areas respectively).

[0069] Ku-band radar (long-range, wide-area monitoring): uses a 2×2 linear phased array antenna, arranged in the vertical direction, with a unit spacing of 1.5λ (λ≈2cm, spacing 3cm), a beam azimuth angle of ±60°, and an elevation angle of ±30°, covering a distance of 2000-5000 meters. It achieves wide-area rapid search through a combination of frequency scanning and phase scanning (for example, a full airspace scan is completed every 100ms).

[0070] As a further explanation of this embodiment, the time division multiplexing technology in this embodiment can use a high-precision timer based on FPGA to achieve microsecond time slot division, or generate a multi-band synchronization trigger signal through a dedicated synchronization chip (such as TICDCE913); the frequency division multiplexing technology in this embodiment can introduce orthogonal frequency division multiplexing (OFDM) technology to further improve spectrum utilization through subcarrier division; or adopt a software-defined radio (SDR) architecture to achieve dynamic configuration of frequency bands.

[0071] Furthermore, if TDM technology is adopted, the acquisition cycle can be divided into multiple time slots (e.g., 0-5ms for X-band, 5-10ms for C-band, and 10-15ms for Ku-band) using an FPGA timer. Synchronous triggering can be used to achieve time-sharing acquisition of multi-band signals. In electromagnetic interference environments, TDM technology can dynamically switch acquisition time slots to avoid interference, while FDM technology can ensure data integrity through frequency band hopping. For multi-target scenarios, the differences in target reflection characteristics of different frequency bands (e.g., detecting small drones in the Ku-band and large passenger aircraft in the C-band) can be exploited to achieve precise multi-target perception. The Nyquist sampling theorem is used to calculate the Doppler frequency based on target velocity (e.g., a target velocity of 200m / s corresponds to an X-band Doppler frequency of 13.3kHz, so the sampling frequency is set to 30kHz). Alternatively, an adaptive sampling strategy can be used to dynamically adjust the sampling frequency based on the target's motion state.

[0072] S200: A clock synchronization mechanism enables coordinated acquisition of multi-band radar data, acquiring multi-dimensional echo signals encompassing time-domain waveforms, frequency-domain spectra, and spatial orientation, and establishing a three-dimensional time-frequency-space mapping relationship. This clock synchronization mechanism ensures the temporal alignment of X-band, C-band, and Ku-band radar data, preventing target feature misalignment due to acquisition timing deviations and improving the accuracy of multi-band data fusion. By simultaneously acquiring time-domain waveforms (reflecting signal temporal characteristics), frequency-domain spectra (revealing frequency components and Doppler information), and spatial orientation (locating the target's spatial position), a comprehensive target characterization is achieved, providing a rich data foundation for subsequent analysis. Furthermore, by establishing a three-dimensional time-frequency-space mapping relationship, feature data from different dimensions is integrated into structured information in a unified coordinate system, facilitating subsequent cross-dimensional feature fusion and analysis, and enhancing the understanding and prediction of target dynamic behavior.

[0073] In this embodiment, the clock synchronization mechanism uses GPS / Beidou satellite-synchronized clocks to obtain an absolute time reference; alternatively, the IEEE1588 precision clock protocol is used to achieve high-precision time synchronization across multiple radar devices within the network. The mapping association method constructs a three-dimensional mapping model based on a spatiotemporal tensor network, efficiently processing the relationships between multi-dimensional data through tensor operations; alternatively, a graph neural network is used to construct a dynamic graph structure from time-frequency-space feature nodes, enabling nonlinear correlation analysis between features.

[0074] S300 performs pulse compression and Doppler processing on the time-domain echo signal, extracting time-domain features (such as pulse width and pulse repetition interval); performs spectral analysis on the frequency-domain signal, extracting frequency-domain features (such as center frequency and bandwidth); and performs angle estimation on the spatial signal, extracting spatial features (such as azimuth and elevation angle). This design extracts features from radar echo signals in three dimensions: time, frequency, and space, enabling a comprehensive description of target characteristics. Time-domain features reflect the signal's temporal distribution, frequency-domain features reveal the signal's frequency components, and spatial features determine the target's position. The combination of these three provides a rich information foundation for target resolution. The simultaneous use of pulse compression and Doppler processing to extract time-domain features improves signal resolution and the ability to detect target velocity. Spectral analysis is used to extract frequency-domain features, effectively identifying signal frequency components and Doppler shift. Angle estimation enables spatial feature extraction, accurately locating the target's spatial position. These techniques are highly targeted and improve feature extraction accuracy. In addition, the extracted time domain, frequency domain, and spatial features are important inputs for subsequent processing steps such as data fusion and target classification. Accurate feature extraction helps improve the accuracy and reliability of radar target resolution.

[0075] For example, in a port ship monitoring scenario, for time-domain echo signals, matched filtering pulse compression technology is used to compress the originally wide transmission pulse into a narrow pulse at the receiving end, thereby improving the distance resolution and accurately measuring the distance between the ship and the radar. Through Doppler processing, the Doppler frequency shift of the echo signal is analyzed to obtain the ship's speed and direction. For frequency-domain signals, fast Fourier transform (FFT) is used for spectral analysis to extract the center frequency and bandwidth of the signal and determine the characteristics of electromagnetic signals generated by equipment such as ship engines. In spatial signal processing, single-pulse angle measurement technology is used for angle estimation to accurately obtain the ship's azimuth and pitch angles, thereby achieving all-round monitoring and identification of ships in the port.

[0076] In the above-mentioned pulse compression process, the design of the matched filter is carried out according to the waveform parameters of the transmitted pulse (such as pulse width and carrier frequency) to achieve the best compression effect; the number of FFT transformation points is determined according to the signal sampling frequency and frequency resolution requirements. For example, if the sampling frequency is 10MHz and the frequency resolution is required to be 10Hz, the number of FFT points is ; When estimating angles, the accuracy of single-pulse angle measurement is related to parameters such as the beam width and gain of the radar antenna. By reasonably designing antenna parameters (such as a beam width of 2°), the angle measurement accuracy can be improved.

[0077] S400: This design uses deep learning models to extract single-dimensional deep features in the time, frequency, and spatial domains. This design utilizes deep learning models (LSTM, CNN) to automatically extract deep features from time, frequency, and spatial signals. Compared to traditional methods that manually design features, it can discover more complex and abstract patterns in signals, improve feature representation capabilities, and thus enhance the accuracy of target resolution.

[0078] In this embodiment, the S400 extracts three types of single-dimensional deep features in the time domain, frequency domain, and space respectively through a deep learning model, which specifically includes:

[0079] Use the long short-term memory network (LSTM) to model the time domain signal, capture the temporal dependencies in the sequence data, and generate LSTM feature vectors;

[0080] Process frequency domain signals (such as spectrograms, time-frequency graphs, etc.) through convolutional neural networks (CNNs), extract their spectral distribution characteristics and local patterns, and output CNN feature vectors;

[0081] For spatial signals (such as the coordinate sequence and geometric topological structure of the motion trajectory), a graph neural network (GNN) or a geometric deep learning model is used to learn the spatial geometric features of the motion trajectory (such as positional relationship, shape description, trajectory change pattern, etc.) and obtain the geometric feature vector.

[0082] As a further illustration of this embodiment, an LSTM network is used to extract the temporal dependence features of time domain signals. This network is suitable for processing data with time series characteristics and can effectively capture the dynamic changes of radar echo signals in the time dimension. A CNN network is used to extract the spectral characteristics of frequency domain signals. Its convolution operation can automatically learn the local features and hierarchical structure of frequency domain data, enhancing the extraction effect of frequency domain features. The geometric features of the motion trajectory are extracted from spatial signals, providing effective information for the spatial positioning and behavior analysis of the target.

[0083] As a further illustration of this embodiment, in addition to the standard LSTM network, the LSTM variant in this embodiment may adopt a gated recurrent unit (GRU), which has a relatively simple structure and higher computational efficiency; or use a bidirectional LSTM (Bi-LSTM) to learn timing features from both forward and reverse directions to more comprehensively capture the temporal dependencies of signals; at the same time, the CNN architecture in this embodiment may adopt a classic CNN architecture such as LeNet, AlexNet, VGGNet, etc., or may use a more advanced residual network (ResNet) or a densely connected network (DenseNet). By introducing residual connections or dense connections, the gradient vanishing problem in deep network training is solved, and the feature extraction capability of the network is improved; in addition, this embodiment may use a graph neural network (GNN) to model the positional relationship of spatial targets as a graph structure, and extract spatial correlation features between targets through graph convolution operations; or adopt a Transformer architecture to use a self-attention mechanism to learn the global dependencies of target spatial features.

[0084] Furthermore, in urban traffic monitoring, for example, there are multiple targets (cars, buses, motorcycles, etc.) with complex and variable motion states. A time-domain LSTM network can learn the temporal characteristics of echo signals from different vehicles as they travel, such as the periodic variations in signal strength caused by the frequent starts and stops of buses. A frequency-domain CNN network can identify the unique spectral signatures produced by different vehicle engines and distinguish vehicle types. A spatial feature extraction network combines target location information to extract geometric features such as vehicle trajectories and lane-changing behavior, enabling effective identification and tracking of different targets in urban traffic flows.

[0085] For example, in severe weather conditions (such as heavy rain or dust storms), radar echo signals are severely interfered with, reducing the signal-to-noise ratio (SNR). In these situations, LSTM networks can leverage their ability to memorize sequential data to extract useful temporal features from noisy time-domain signals. CNN networks, through multi-layer convolution and pooling operations, enhance robustness to frequency-domain features and suppress noise interference. Furthermore, during network training, data augmentation techniques (such as adding Gaussian noise and random shifting) are employed to improve the model's feature extraction capabilities in low SNR environments.

[0086] S500. Select any two of the three types of single-dimensional deep features in the time domain, frequency domain, and space domain to form a two-dimensional feature pair, and perform normalization and feature alignment on the two-dimensional feature pairs. Z-score standardization is used to eliminate dimensional differences and generate a binary feature vector containing the correlation between dimensions. Normalizing the two-dimensional features through Z-score standardization can eliminate the dimensional influence between different features, so that each feature is at the same order of magnitude, which is convenient for subsequent fusion and comparison, and improves the stability and convergence speed of the model. By aligning the two-dimensional features, it is ensured that the features of different dimensions are consistent in time, space or other reference systems. The generated binary feature vector contains the correlation information between dimensions, which helps to explore the potential relationship between features and improve the feature expression ability. At the same time, the normalized and aligned feature vectors provide a good data foundation for subsequent cross-dimensional feature fusion, so that features of different dimensions can be processed under a unified framework, improving the effect and accuracy of multi-dimensional data fusion.

[0087] In this embodiment, in addition to Z-score normalization, the normalization method in this embodiment can also adopt Min-Max normalization to scale the eigenvalues ​​to the [0,1] interval; or use L2 normalization to normalize the feature vector to a unit vector; at the same time, the feature alignment method can be based on timestamp alignment, which is suitable for time series data; it can also be based on spatial position alignment, for example, combining different dimensional features of the same spatial area; it can also be based on event triggering alignment, when a specific event occurs, the relevant different dimensional features are aligned; in addition, the generation of binary feature vectors can adopt a splicing method to directly splice the feature vectors of two dimensions into a longer vector; it can also adopt a weighted combination method to assign different weights according to the importance of the features to generate a weighted combination of binary feature vectors.

[0088] As a further explanation of this embodiment, the method for generating a binary feature vector in this embodiment includes: generating a binary feature vector by splicing, weighted combination, etc., for example, the time domain feature vectors at the same time are and frequency domain eigenvector Combination ; In multi-dimensional scenarios, it can be extended to generate triples or higher-dimensional feature vectors.

[0089] Furthermore, in an airport airspace monitoring scenario, for example, there are multiple different types of targets (civilian airliners, fighter jets, drones, etc.). For each target, its time, frequency, and spatial features are extracted separately. When performing two-dimensional feature processing, features are first grouped by target ID. Then, features of the same target in different dimensions are normalized and aligned using the Z-score. For example, for a particular civil airliner at time t, its time domain features (e.g., pulse width) and frequency domain features (e.g., Doppler shift) are normalized and aligned to generate a binary feature vector for subsequent target classification and recognition.

[0090] S600, build a dynamic weight allocation model. Based on the three-dimensional quality indicators of real-time signal-to-noise ratio (SNR), feature stability (SI), and data integrity (CI), the fusion weight is calculated through an adaptive algorithm to generate a joint representation vector containing multi-dimensional information in the time domain, frequency domain, and space. This design builds a dynamic weight allocation model based on the three-dimensional quality indicators of real-time signal-to-noise ratio (SNR), feature stability (SI), and data integrity (CI). It can adaptively adjust the fusion weight according to the real-time quality of data in each dimension, increase the reliance on high-quality data in complex environments, and enhance the robustness of the model.

[0091] In this embodiment, the dynamic weight allocation model in S600 calculates the fusion weight through a three-layer quality assessment system, including the following steps:

[0092] S610: Use the periodogram method to estimate the signal power spectral density and calculate the real-time signal-to-noise ratio (SNR). Within a sliding window of 3-10 scanning cycles, calculate the feature stability (SI) by the rate of change of the cosine similarity of the feature vectors. A rate of change < 0.1 is considered stable. Use the Kalman filter to predict missing data points and calculate the data integrity (CI) in combination with the IQR outlier detection algorithm. A missing rate < 5% is considered complete.

[0093] It is understandable that the scanning cycle in this embodiment adopts a sliding window of 5 scanning cycles. The core principle is to balance the real-time performance and stability of feature evaluation:

[0094] Real-time performance: 5 cycles (e.g., 500ms for a 100ms cycle) can quickly respond to changes in target characteristics, meeting the radar system's real-time processing requirements for dynamic targets.

[0095] Stability: Through multi-cycle data statistics, the impact of single-cycle noise interference on feature evaluation is reduced, ensuring the reliability of indicators such as signal-to-noise ratio (SNR) and feature stability (SI).

[0096] Adjustment logic: Users can dynamically adjust the window length based on target movement speed (e.g., shorten to 3 cycles for high-speed targets and extend to 7 cycles for low-speed targets) or the real-time requirements of the scene.

[0097] In this embodiment, the threshold value of the cosine similarity change rate of the feature vector is set to 0.1. The technical logic is based on the physical characteristics of the radar target:

[0098] Stationary target characteristics: When the target is in a stable motion state such as uniform speed or straight line, the cosine similarity change rate of its time domain, frequency domain or spatial characteristics is usually low (for example, the characteristics of a repetitive periodic signal change smoothly).

[0099] Dynamic target differentiation: A threshold of 0.1 can effectively distinguish stable features from dynamically changing features (for example, when the target maneuvers or changes its posture, the feature change rate is usually greater than 0.1), providing a reliable stability criterion for multi-dimensional feature fusion.

[0100] Adaptive Adjustment: For high-dynamic scenarios (such as rapid drone maneuvers), the threshold can be appropriately increased to 0.15; for low-speed or stationary target scenarios, it can be reduced to 0.08 to enhance sensitivity.

[0101] In this embodiment, data integrity is determined by a loss rate of less than 5%. The theoretical basis is derived from the performance boundary of the data recovery algorithm:

[0102] Kalman filter effectiveness: When the data missing rate is ≤5%, the Kalman filter can effectively recover the missing points through time series prediction. Combined with the IQR outlier detection algorithm, it can ensure that the accuracy of the recovered data meets the target analysis requirements (for example, the motion parameter estimation error is within an acceptable range).

[0103] Error control: A missing rate exceeding 5% may lead to cumulative prediction errors and affect the accuracy of feature fusion. Therefore, this threshold is set to balance the data recovery cost and parsing accuracy.

[0104] Scenario adaptation: In strong interference or high noise environments, the loss rate can be reduced by optimizing the anti-interference algorithm (such as increasing redundant sampling), and then combined with threshold adjustment to achieve data integrity assessment.

[0105] S620: Calculate the initial fusion weight using the weighted product method, adjust the weight sensitivity using the exponential normalization algorithm, and generate the final weight vector.

[0106] As a further illustration of this embodiment, in the dynamic weight allocation model S600, the fusion weight calculation formula is:

[0107] ;

[0108] in, , Represents the three dimensions of time domain, frequency domain and space, ; Indicates the Real-time signal-to-noise ratio of dimensions; represents the characteristic stability index of the i-th dimension; Indicates the Data completeness index of the dimension;

[0109] In order to avoid the denominator being zero (for example, all the indicators of a dimension are zero), a very small constant is introduced in the actual calculation = , the corrected formula is: ;

[0110] Furthermore, in the S600 dynamic weight allocation model, the fusion weight is determined by the three-dimensional quality indicators. Taking the frequency domain as an example, its weight The calculation of depends on:

[0111] : Perform an FFT transform on the C-band spectrum signal using the periodogram method described in S610, estimate the power spectrum density using an N=1024-point Hanning window, and calculate the ratio of signal power to noise power (unit: dB);

[0112] :In a sliding window of 3-10 scanning cycles (10ms per cycle), calculate the cosine similarity change rate of the spectral feature vectors of adjacent cycles. If the mean change rate is 0.05 (<0.1), then =1-0.05=0.95;

[0113] :Predict missing points in frequency domain data through Kalman filtering and detect outliers with IQR algorithm. If 3-10 of the 256 frequency points are missing, then =1-0.0195=0.9805.

[0114] In this embodiment, the signal receiving module transmits the frequency domain raw data to the data cache module (cache period 10ms) through the PCIe4.0 bus. The quality assessment module predicts missing data using the built-in Kalman filter unit of the FPGA and marks abnormal values ​​through the IQR detection unit of the ASIC architecture. Finally, the signal receiving module transmits the frequency domain raw data to the data cache module (cache period 10ms) through the LVDS parallel interface. 、 、 The data is then transmitted to the weight calculation module (transmission rate 1Gbps). The weight calculation module feeds the fusion weights back to the feature fusion module via DDR4 shared memory, achieving dynamic weighted fusion with temporal and spatial features (update frequency 100Hz, synchronized with the scanning cycle).

[0115] As a further illustration of this embodiment, the real-time signal-to-noise ratio (SNR) in this embodiment can use the periodogram method or the Welch method to estimate the signal power spectral density, and calculate the SNR by distinguishing the signal and noise spectral peaks; in addition to the variance, the feature stability (SI) can also use the cosine similarity change rate of the feature vector to evaluate the stability, and the smaller the similarity change, the more stable it is; the data integrity (CI) is evaluated by detecting the data missing rate, the proportion of outliers, etc., and the missing data can be interpolated in combination with the Kalman filter prediction value; in addition, in addition to exponential function normalization, the weight calculation method in this embodiment can also use the Sigmoid function or the Softmax function to calculate the weight, or design a hierarchical weight allocation strategy (such as first performing a rough allocation based on the SNR, and then fine-tuning in combination with the SI and CI).

[0116] Furthermore, the dynamic weight in this embodiment may adopt an exponential normalization formula, which is calculated as follows: ;in For the The quality score of each dimension (obtained by weighting SNR, SI, and CI) Adjust sensitivity; or use the Softmax function or hierarchical weight allocation strategy (first coarse SNR allocation, then SI or CI fine tuning).

[0117] S700. Map single-dimensional deep features, binary feature vectors, and joint representation vectors into spatiotemporal tensors, and define a dynamic graph structure containing single-modal nodes, dual-modal fusion nodes, and full-modal global nodes. The edge connections between nodes are determined by feature cosine similarity and temporal correlation. Through the spatiotemporal tensor and dynamic graph structure, single-dimensional features (time domain, frequency domain, space) and fusion features are integrated to form a hierarchical feature expression, which can not only retain the unique information of each dimension, but also capture the correlation between dimensions, thereby improving the comprehensive representation ability of the target.

[0118] In this embodiment, the construction of the spatiotemporal tensor in S700 includes the following steps:

[0119] S710, performing dimension mapping on the time domain features (such as pulse width and repetition interval sequence) to generate a time dimension tensor with the scanning period (10 periods) as the time axis and the time domain features (8 dimensions) as the feature axis;

[0120] S720. Perform dimension mapping on the frequency domain features (such as Doppler shift and spectrum energy) to generate a frequency dimension tensor with frequency points (256 points) as the frequency axis and frequency domain features (12 dimensions) as the feature axis.

[0121] S730: Dimensionally map the spatial features (such as azimuth, elevation, and distance) to generate a spatial dimension tensor with the spatial grid (divided into 64×64 pixels) as the spatial axis and the spatial features (6 dimensions) as the feature axis.

[0122] S740. Synthesize the time, frequency, and space dimension tensors into a three-dimensional space-time tensor according to the correspondence between period, frequency bin, and spatial grid, where each tensor element aggregates multi-dimensional features corresponding to the time period, frequency interval, and spatial position.

[0123] As a further illustration of this embodiment, the feature dimension division includes: mapping the time domain features (such as pulse width, pulse repetition interval sequence) into the time dimension of the tensor, the dimension size is ,in is the number of continuous acquisition cycles (for example, set to 10 scan cycles), is the time domain feature dimension (such as 8 dimensions); the frequency domain features (such as Doppler shift, spectrum energy distribution) are mapped to the frequency dimension, and the dimension size is ,in is the number of frequency domain bins (such as 256 frequency points), is the frequency domain feature dimension (such as 12 dimensions); the spatial features (such as azimuth, pitch angle, distance) are mapped to the spatial dimension, and the dimension size is ,in is the number of spatial grids (e.g. the radar monitoring area is divided into 64×64 grids), is the spatial feature dimension (such as 6 dimensions).

[0124] As a further explanation of this embodiment, the mathematical definition of the space-time tensor is: construct a three-dimensional space-time tensor , where each element Indicates the Time period, frequency bin, The multi-dimensional feature aggregation value of a spatial grid.

[0125] Furthermore, the dual determination mechanism for edge connections in this embodiment is:

[0126] Feature similarity measurement, using cosine similarity to calculate node feature matching, the formula is: ; For example, when the domain node Frequency Domain Node When the cosine similarity is greater than 0.8, it is determined that there is a strong similarity connection between the two nodes;

[0127] Timing correlation measurement: Timing dependency is evaluated by sliding window cross-correlation coefficient. The window length is set to 3-10 scan cycles. The formula is: ,in For the The characteristic value of the period, is the characteristic mean and standard deviation, when When it is >0.7, it is determined that there is a temporal correlation.

[0128] It should be added that the edge weights in this embodiment It is a weighted synthesis of similarity and relevance, and the formula is: For example, in the UAV maneuvering flight scenario, the temporal correlation between the time domain characteristics (pulse repetition interval) and the spatial characteristics (acceleration) Rising to 0.8, driving the edge weight Increased from 0.5 to 0.7 to strengthen the interaction between two-dimensional features.

[0129] In addition, the update mechanism of the dynamic graph structure in this embodiment is as follows: the graph attention network (GAT) is used to update the node representation, and the specific steps are as follows;

[0130] For each node ,Generate query vector through linear transformation , key vector , value vector ,in is the weight matrix (set is the total feature dimension);

[0131] Compute nodes With neighboring nodes Attention coefficient: ;

[0132] Aggregate neighbor features to generate new node representations:

[0133] in For nodes The first-order neighbor set of a node is dynamically strengthened through this mechanism (such as edges with strong similarity or strong temporal association).

[0134] S800, aggregates neighborhood features through a spatiotemporal multi-head attention mechanism. Each attention head independently calculates the correlation weights between dimensions, combines residual connections with layer normalization to optimize feature expression, and generates a target feature vector containing spatiotemporal correlations.

[0135] In this embodiment, the spatiotemporal attention mechanism in S800 adopts a graph attention network architecture, including the following steps:

[0136] S811, for each node feature , respectively by querying the weight matrix , key weight matrix , value weight matrix Generate the corresponding 64-dimensional query vector , key vector , value vector ;

[0137] S812, for the node With neighboring nodes , through the formula , calculate the attention coefficient, where, Representation node For neighbor nodes The attention weight of For nodes The query vector (QueryVector) is used to capture the query representation of node features; For nodes The key vector is used to calculate the attention score with the query vector; For nodes The first-order neighbor set (i.e. the set of directly connected nodes), For nodes The first-order neighbor number index of ; For nodes The key vector of Represents the transpose of a vector matrix; It is a leaky rectified linear unit activation function used to enhance the nonlinear expression of attention scores;

[0138] S813, through , aggregate neighbor features, and introduce residual connection and layer normalization (LayerNorm) to optimize feature expression; among them, For nodes Output feature vector after aggregating neighbor features; An activation function (such as Sigmoid, ReLU, etc., which depends on the model configuration) is used to perform nonlinear transformation on the aggregation result; For nodes The feature vector represents the original features of the neighboring nodes or the features processed by the previous layer.

[0139] In this embodiment, the spatiotemporal multi-head attention mechanism in S800 is collaboratively calculated by 8 parallel attention heads, including the following steps:

[0140] S821. Perform multi-head feature mapping on the 36-dimensional input features, and generate 8 sets of 64-dimensional query, key, and value matrices through 8 sets of independent weight matrices;

[0141] S822. Perform single-head attention calculation on each set of feature matrices, use the Softmax function to normalize the dot product of the query and the key, and obtain the single-head attention output;

[0142] S823, concatenate 8 single-head outputs and map them back to a 36-dimensional feature vector through the output matrix to complete the multi-head attention feature fusion.

[0143] As a further illustration of this embodiment, in this embodiment, the multi-head attention mechanism uses 8 parallel attention heads. Each head maps the input features (such as 36-dimensional spatiotemporal tensor node features) into 64-dimensional query (Q), key (K), and value (V) matrices through linear transformation. The formula is: ;in, is the input feature matrix; 、 、 Indicates the The learnable weight matrix of the head; 、 、 Indicates the The query, key, and value matrix of the header;

[0144] Among them, the calculation formula for each head to calculate the attention weight independently is: ;in, Represents the similarity matrix between nodes; Represents the scaling factor; softmax represents the normalization operation and outputs the attention weight; Attentioni represents the Weighted features of the head;

[0145] Among them, the output of 8 heads is fused and mapped back to the original dimension ;in, Indicates concatenating 8 head outputs along the feature dimension; represents the learnable mapping matrix.

[0146] S900: Based on the spatiotemporal joint feature expression, the target type probability distribution is calculated through a multi-level classifier, similarity matching is performed in combination with the target feature knowledge base, and the target type recognition result is output;

[0147] In this embodiment, the target feature knowledge base in S900 is constructed using an improved K-means++ clustering algorithm, including the following steps:

[0148] S911. Perform Z-score normalization on the 12-dimensional time-frequency features. The calculation formula is: , and reduced to 6 dimensions through PCA; is the characteristic mean, is the standard deviation, normalizing the features to the standard normal distribution; is the eigenvalue after Z-score standardization;

[0149] S912, through the Euclidean distance formula ), measure the similarity between the sample and the center, and iteratively generate clustering templates; where, For the sample, is the cluster center; Representation sample and cluster centers the distance between them; Represents the sum index variable;

[0150] S913: Use the cosine similarity algorithm to determine the matching degree between the new sample and the template, and trigger the template update mechanism.

[0151] In this embodiment, the multi-level classifier in S900 adopts a two-stage decision architecture, including the following steps:

[0152] S921: Input the spatial features (azimuth angle change rate, velocity vector, acceleration) into the SVM classifier, use the RBF kernel function to distinguish low-altitude targets (change rate > 10° / s) from high-altitude targets (change rate < 5° / s), and output the coarse classification results.

[0153] S922: Input the fused 12-dimensional time-frequency features (such as pulse width and Doppler shift peak) into the random forest classifier, and output the target type probability distribution through a voting mechanism of 500 decision trees;

[0154] S923: Calculate the cosine similarity between the current feature and the knowledge base template (threshold ≥ 0.92), correct the classification probability and output the final target type (such as "multi-rotor drone") and confidence level (such as 0.98).

[0155] As a further explanation of this embodiment, the classification decision process in this embodiment includes the following contents:

[0156] Primary filtering: SVM quickly eliminates non-targets based on spatial features and outputs coarse classification results (such as "low-altitude maneuvering target" and "high-altitude stable target");

[0157] Fine recognition: Random forest performs in-depth time-frequency feature analysis on the roughly classified targets and outputs the probability distribution of the target type;

[0158] Knowledge base assistance: Combine the feature template similarity matching results, correct the classification probability, and finally output the target type (such as "multi-rotor drone" and "small piston aircraft").

[0159] It should be added that in the construction of the target feature knowledge base, the K-means clustering algorithm generates a multi-dimensional feature template of a typical target through unsupervised learning. The specific implementation is as follows: First, the number of clusters is preset based on historical data. (For example, if the target type is set to 5 categories), the K-means++ algorithm is used to initialize the cluster center to avoid local optimality, and the similarity between the sample and the center is measured by the Euclidean distance (Formula ), iterate through “sample allocation-center update” until convergence (center change <10 or up to 50 iterations). In the preprocessing stage, the 12-dimensional time-frequency features are Z-score normalized ( ) to eliminate the dimensionality effect, combined with principal component analysis (PCA) to reduce it to 6 dimensions (retaining 295% variance) and filter key dimensions (such as pulse width and Doppler shift peak) according to feature importance.

[0160] At the same time, the knowledge base is calculated by cosine similarity technology: , dynamically update the template, where: represents the new sample feature vector, represents the known feature vectors in the template library, represents the vector dot product, represents the L2 norm; new samples with a similarity ≥ 0.92 match the known template; a similarity below the threshold triggers an update—if the number of samples in the cluster is > 20, the outlier template is replaced; otherwise, a new template is added to avoid noise contamination. In practice, an offline training-online optimization model is employed, using Mini-Batch K-means to reduce real-time computational load, ensuring the knowledge base adapts to new target feature variations and providing a reliable feature matching benchmark for multi-dimensional target parsing.

[0161] S1000. Generate a structured decision report including target type, motion parameters and threat level, and output a visual target situation map with a data traceability chain.

[0162] As a further illustration of this embodiment, in the structured decision report generation mechanism, the system generates standardized output based on target type, motion parameters and feature matching results: First, the threat value is calculated by the threat level quantification model. Each parameter is normalized and mapped to the [0, 1] range (e.g., TypeScore = 0.8, ProximityScore = 0.9). The report integrates multi-dimensional information in XML / JSON format. A typical structure includes target type (with confidence level), motion parameters (3D coordinates, velocity, acceleration), threat level (Levels I-IV and description), and data provenance (feature source and classification basis).

[0163] Example 2

[0164] This embodiment also provides a radar device based on multi-dimensional data fusion, which is loaded with a corresponding computer program and can execute the steps of the radar target analysis calculation method based on multi-dimensional data fusion as described above.

[0165] Those skilled in the art will appreciate that the process of implementing all or part of the steps of the above embodiments may be accomplished by hardware, or by instructing related hardware through a program, which may be stored in a computer-readable storage medium.

[0166] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A radar target parsing and calculation method based on multi-dimensional data fusion, characterized in that: The steps include: S100, deploying a multi-band radar sensor array including X-band, C-band, and Ku-band in the radar monitoring area; S200, realizing collaborative acquisition of multi-band radar data through a clock synchronization mechanism, obtaining multi-dimensional echo signals including time domain waveform, frequency domain spectrum, and spatial orientation, and establishing a three-dimensional mapping relationship between time, frequency, and space; S300, performing pulse compression and Doppler processing on the time domain echo signal, and extracting time domain features; Perform spectrum analysis on frequency domain signals and extract frequency domain features; Estimating the angle of spatial signals and extracting spatial features; S400, through the deep learning model, extracts three types of single-dimensional deep features in the time domain, frequency domain, and space domain: S500, randomly combining any two of the three types of single-dimensional deep features in the time domain, frequency domain, and space domain to form a two-dimensional feature pair, and performing normalization and feature alignment on the two-dimensional feature pair, eliminating dimensional differences through Z-score normalization, and generating a two-tuple feature vector containing the correlation between the dimensions; S600: Build a dynamic weight allocation model. Based on the three-dimensional quality indicators of real-time signal-to-noise ratio (SNR), feature stability (SI), and data integrity (CI), calculate the fusion weight through an adaptive algorithm to generate a joint representation vector containing multi-dimensional information in the time domain, frequency domain, and space domain. S700, mapping the single-dimensional deep features, the binary feature vectors, and the joint representation vectors into a spatiotemporal tensor, and defining a dynamic graph structure including single-modal nodes, dual-modal fusion nodes, and full-modal global nodes; S800, aggregates neighborhood features through a spatiotemporal multi-head attention mechanism. Each attention head independently calculates the correlation weights between dimensions, combines residual connections with layer normalization to optimize feature expression, and generates a target feature vector containing spatiotemporal correlations. S900: Based on the spatiotemporal joint feature expression, the target type probability distribution is calculated through a multi-level classifier, similarity matching is performed in combination with the target feature knowledge base, and the target type recognition result is output; S1000. Generate a structured decision report including target type, motion parameters and threat level, and output a visual target situation map with a data traceability chain.

2. The radar target analysis and calculation method based on multi-dimensional data fusion according to claim 1 is characterized in that: The multi-band radar sensor array in S100 synchronously collects echo signals through a hybrid technology of time division multiplexing and frequency division multiplexing, including the following steps: S110, based on time division multiplexing technology, uses FPGA high-precision timer to achieve microsecond time slot division, or generates multi-band trigger signals through TICDCE913 synchronization chip to achieve orthogonal sampling of multi-band radar in the time domain; S120, spectrum resource allocation based on frequency division multiplexing technology introduces orthogonal frequency division multiplexing technology, improves spectrum utilization through subcarrier division, and adopts software-defined radio architecture to achieve dynamic frequency band configuration and build a two-dimensional orthogonal signal acquisition framework in time and frequency. S130 , performing adaptive sampling frequency control, first ensuring that the sampling frequency meets the Nyquist sampling criterion to avoid signal aliasing, and then predicting the target motion parameters through the Kalman filter algorithm according to the target motion state to optimize the sampling frequency in real time.

3. The radar target analysis and calculation method based on multi-dimensional data fusion according to claim 1 is characterized in that: In S400, the deep learning model is used to extract three types of single-dimensional deep features in the time domain, frequency domain and space, which specifically include: Use long short-term memory networks to model time domain signals, capture temporal dependencies in sequence data, and generate LSTM feature vectors; Process the frequency domain signal through convolutional neural network, extract its spectrum distribution characteristics and local patterns, and output CNN feature vector; For spatial signals, graph neural networks or geometric deep learning models are used to learn the spatial geometric features of motion trajectories and obtain geometric feature vectors.

4. The radar target analysis and calculation method based on multi-dimensional data fusion according to claim 1 is characterized in that: The dynamic weight allocation model in S600 calculates the fusion weight through a three-layer quality assessment system, including the following steps: S610, using the periodogram method to estimate signal power spectral density and calculate real-time signal-to-noise ratio (SNR); calculating feature stability (SI) by the rate of change of cosine similarity of feature vectors within a sliding window of 3-10 scanning cycles; using Kalman filtering to predict missing data points and combining the IQR outlier detection algorithm to calculate data integrity (CI); S620: Calculate the initial fusion weight using the weighted product method, adjust the weight sensitivity using the exponential normalization algorithm, and generate the final weight vector.

5. The radar target analysis and calculation method based on multi-dimensional data fusion according to claim 1 is characterized in that: The construction of the space-time tensor in S700 includes the following steps: S710, performing dimension mapping on the time domain features to generate a time dimension tensor with the scanning period as the time axis and the time domain features as the feature axis; S720: Perform dimension mapping on the frequency domain features to generate a frequency dimension tensor with the frequency points as the frequency axis and the frequency domain features as the feature axis; S730, performing dimension mapping on the spatial features to generate a spatial dimension tensor with the spatial grid as the spatial axis and the spatial features as the feature axis; S740. Synthesize the time, frequency, and space dimension tensors into a three-dimensional space-time tensor according to the correspondence between period, frequency bin, and spatial grid, where each tensor element aggregates multi-dimensional features corresponding to the time period, frequency interval, and spatial position.

6. The radar target analysis and calculation method based on multi-dimensional data fusion according to claim 1 is characterized in that: The spatiotemporal attention mechanism in S800 adopts a graph attention network architecture, including the following steps: S811, for each node feature , respectively by querying the weight matrix , key weight matrix , value weight matrix Generate the corresponding 64-dimensional query vector , key vector , value vector ; S812, for the node With neighboring nodes , through the formula , calculate the attention coefficient, where Representation node For neighbor nodes The attention weight of For nodes The query vector is used to capture the query representation of node features; For nodes The key vector is used to calculate the attention score with the query vector; For nodes The first-order neighbor set of For nodes The first-order neighbor number index of ; For nodes The key vector of Represents the transpose of a vector matrix; It is a leaky rectified linear unit activation function used to enhance the nonlinear expression of attention scores; S813, through , aggregate neighbor features, and introduce residual connections and layer normalization to optimize feature expression; among them, For nodes Output feature vector after aggregating neighbor features; is the activation function; For nodes The eigenvector of .

7. The radar target analysis and calculation method based on multi-dimensional data fusion according to claim 6 is characterized in that: The spatiotemporal multi-head attention mechanism in S800 is a collaborative computation using eight parallel attention heads, including the following steps: S821. Perform multi-head feature mapping on the 36-dimensional input features, and generate 8 sets of 64-dimensional query, key, and value matrices through 8 sets of independent weight matrices; S822. Perform single-head attention calculation on each set of feature matrices, use the Softmax function to normalize the dot product of the query and the key, and obtain the single-head attention output; S823, concatenate 8 single-head outputs and map them back to a 36-dimensional feature vector through the output matrix to complete the multi-head attention feature fusion.

8. The radar target analysis and calculation method based on multi-dimensional data fusion according to claim 1 is characterized in that: The target feature knowledge base in S900 is constructed by using an improved K-means++ clustering algorithm, including the following steps: S911. Perform Z-score normalization on the 12-dimensional time-frequency features. The calculation formula is: , and reduced to 6 dimensions through PCA; is the characteristic mean, is the standard deviation; is the eigenvalue after Z-score standardization; S912, through the Euclidean distance formula ), measure the similarity between the sample and the center, and iteratively generate clustering templates; where, For the sample, is the cluster center; Represents a sample and cluster centers the distance between them; Represents the sum index variable; S913: Use the cosine similarity algorithm to determine the matching degree between the new sample and the template, and trigger the template update mechanism.

9. The radar target analysis and calculation method based on multi-dimensional data fusion according to claim 8, characterized in that: The multi-level classifier in S900 adopts a two-stage decision architecture, including the following steps: S921: Input the spatial features into the SVM classifier, use the RBF kernel function to distinguish low-altitude targets from high-altitude targets, and output the rough classification results; S922: Input the fused 12-dimensional time-frequency features into the random forest classifier, and output the target type probability distribution through a voting mechanism of 500 decision trees; S923: Calculate the cosine similarity between the current feature and the knowledge base template, correct the classification probability and output the final target type and confidence.

10. A radar device based on multi-dimensional data fusion, characterized in that: The method is loaded with a corresponding computer program and can execute the steps of the radar target analysis and calculation method based on multi-dimensional data fusion as described in any one of claims 1 to 9.

Images