A Multi-Source Target Trajectory Fusion Method and System Based on Deep Learning

Abstract

The present invention relates to the technical field of data processing, and in particular, to a multi-source target trajectory fusion method and system based on deep learning. It includes constructing a spatio-temporal alignment training set containing radar echo data, satellite telemetry data, and ground truth trajectory data; designing a multi-channel deep learning network architecture, including a radar echo data feature extraction channel and a satellite telemetry data feature extraction channel, and establishing dual-channel feature correlation through an attention mechanism; optimizing network parameters through backpropagation to obtain a weight prediction model; receiving radar echo data and satellite telemetry data in real time, and realizing spatio-temporal alignment of sampled data through a sliding window mechanism and coordinate system transformation; inputting the preprocessed dual-source features into the trained weight prediction model to generate radar-telemetry fusion weight coefficients; performing weighted processing and optimization on the dual-source trajectory coordinates, and outputting the fused target motion trajectory. It solves the problem of insufficient accuracy of traditional single-source trajectory measurement systems in complex environments.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a multi-source target trajectory fusion method and system based on deep learning. Background Art

[0002] In the prior art, the target trajectory measurement system mainly relies on a single sensor data source, and there are significant limitations in measurement accuracy in complex environments. When satellite telemetry data is used, it is limited by the transmission conditions of satellite signals. When encountering occlusion, electromagnetic interference, etc., the signal is easily lost or deviated, resulting in incomplete or inaccurate trajectory data. Moreover, the time resolution of satellites is usually low, and it is easily affected by the atmosphere to produce centimeter-level spatial offsets, making it difficult to accurately capture the instantaneous pose changes of high-speed maneuvering targets. When radar echo data is used, although the radar has a millisecond-level time resolution, affected by environmental factors such as multipath effects and ground object occlusion, the target scattering characteristics will be distorted in a complex electromagnetic environment, making it difficult to extract the target scattering characteristics, thus affecting the judgment of the target position and trajectory. In addition, there are also certain limitations in the detection range and accuracy of the radar. For the trajectory measurement of long-distance or small targets, its accuracy will decrease significantly.

[0003] Advanced target trajectory measurement requires integrating the complementary advantages of radar and satellites, but the prior art faces severe challenges in the deep fusion of heterogeneous data. Since the radar echo data is essentially a time-frequency domain scattering feature matrix, while the satellite telemetry data is discrete positioning parameters in the geographical coordinate system, there are essential differences between the two in terms of data structure dimension, spatio-temporal reference, error characteristics, etc. The existing fusion methods cannot establish an association model for cross-modal features, and the traditional fixed-weight fusion mechanism is also difficult to adaptively adjust the confidence weights of radar and telemetry data according to dynamic factors such as environmental interference intensity and target maneuvering state. For a target trajectory measurement system with high accuracy requirements, the prior art cannot effectively provide accurate, timely, and effective trajectory fusion support, thus making it difficult to meet the high-precision tracking and positioning requirements of targets in complex environments in practical applications, and limiting the performance and reliability of the trajectory measurement system in actual operation. Summary of the Invention

[0004] In order to solve the problem of insufficient accuracy of the traditional single-source trajectory measurement system in complex environments, the present invention provides a multi-source target trajectory fusion method and system based on deep learning.

[0005] In the first aspect, a multi-source target trajectory fusion method based on deep learning provided by the present invention adopts the following technical solution:

[0006] A multi-source target trajectory fusion method based on deep learning includes:

[0007] Construct a spatio-temporal alignment training set containing radar echo data, satellite telemetry data, and true value trajectory data;

[0008] Design a multi-channel deep learning network architecture, including an echo data feature extraction channel and a telemetry data feature extraction channel, and establish dual-channel feature correlation through an attention mechanism;

[0009] Optimize the network parameters through backpropagation to obtain a weight prediction model;

[0010] Receive radar echo data and satellite telemetry data in real time, and achieve spatio-temporal alignment of sampled data through a sliding window mechanism and coordinate system transformation;

[0011] Input the preprocessed dual-source features into the trained weight prediction model to generate radar-telemetry fusion weight coefficients;

[0012] Perform weighted processing and optimization on the dual-source trajectory coordinates, and output the fused target motion trajectory.

[0013] Furthermore, the construction of the spatio-temporal alignment training set containing radar echo data, satellite telemetry data, and true value trajectory data includes:

[0014] Perform pulse compression processing on the original radar echo data and extract error parameters to construct a feature matrix;

[0015] Perform coordinate system transformation processing on the satellite telemetry data and extract error parameters to construct a feature matrix;

[0016] Adopt a sliding window mechanism and cubic spline interpolation method to align and fill the time axis of the dual-source data, and establish training samples synchronized at the millisecond level.

[0017] Furthermore, the design of the multi-channel deep learning network architecture, including a radar echo data feature extraction channel and a satellite telemetry data feature extraction channel, and establishing dual-channel feature correlation through an attention mechanism includes:

[0018] The radar echo data feature extraction channel uses a three-dimensional convolutional neural network to extract spatio-temporal-error features;

[0019] The satellite telemetry data feature extraction channel uses a bidirectional LSTM network to decode the satellite positioning parameter sequence;

[0020] The cross-attention module establishes a correlation matrix of radar features and telemetry features with environmental quality perception;

[0021] The fully connected output layer generates a weight distribution tensor with dimensions [timestamp × spatial coordinates].

[0022] Furthermore, the cross-attention module calculates the correlation matrix of radar features and telemetry features, including:

[0023] Map the radar feature vector and the telemetry feature vector to the query space and the key-value space respectively;

[0024] Calculate the feature similarity matrix and generate the attention weight distribution through the softmax function;

[0025] Perform weighted fusion on the dual-channel features to generate a joint feature representation and input it into the fully connected layer.

[0026] Furthermore, the optimizing the network parameters through backpropagation to obtain the weight prediction model includes:

[0027] Initialize the network parameters and define the mean square error loss function for the fusion trajectory and the ground truth trajectory;

[0028] Update the network weights using backpropagation and the stochastic gradient descent algorithm;

[0029] Control model overfitting through early stopping and save the optimal weight prediction model.

[0030] Furthermore, the real-time receiving of radar echo data and satellite telemetry data and the spatio-temporal alignment of the sampled data through the sliding window mechanism and coordinate system transformation includes:

[0031] Perform pulse compression processing on the original radar echo data and extract the error parameters to construct a feature matrix;

[0032] Perform coordinate system transformation processing on the satellite telemetry data and extract the error parameters to construct a feature matrix;

[0033] Use the sliding window mechanism and the cubic spline interpolation method to align and fill in the time axis of the dual-source data.

[0034] Furthermore, the weighted processing and optimization of the dual-source trajectory coordinates to output the fused target motion trajectory includes:

[0035] Perform weighted calculation on the radar trajectory point Pr and the telemetry trajectory point Ps: Pfused = α·Pr + β·Ps, where Pfused is the fused point coordinate, and α and β are the fusion weight coefficients of the radar data and the telemetry data respectively;

[0036] Perform moving average filtering on the fused trajectory to remove noise and output the optimized target motion trajectory.

[0037] In a second aspect, a multi-source target trajectory fusion system based on deep learning includes:

[0038] A spatio-temporal alignment training set construction module, which is used to resample the radar echo data, satellite telemetry data and true value trajectory on the time axis and unify the coordinate system, so as to generate a spatio-temporally synchronized supervised learning data set;

[0039] A model network architecture construction module, which is connected to the spatio-temporal alignment training set construction module, and is used to build a deep learning model network architecture suitable for multi-source target trajectory fusion;

[0040] A model training module, which is connected to the model network architecture construction module, uses the constructed training set to train the built network architecture, generates a dynamic weight prediction model, and saves the optimal parameter model;

[0041] A sampled data spatio-temporal alignment module, which performs alignment processing on the multi-source trajectory data obtained in real time, so that data from different sources are comparable in terms of time and space dimensions;

[0042] A fusion weight coefficient generation module, which is connected to the model training module and the sampled data spatio-temporal alignment module, and is used to input the preprocessed dual-source features into the training model to generate weight coefficients for trajectory fusion;

[0043] A multi-source trajectory weighted processing module, which is connected to the fusion weight coefficient generation module, and is used to perform linear weighted calculation on radar trajectory points and telemetry trajectory points according to the weight coefficients to generate an initial fusion trajectory;

[0044] A fusion trajectory filtering and smoothing module, which is connected to the multi-source trajectory weighted processing module, performs filtering and smoothing operations on the weighted fusion trajectory, removes noise and fluctuations, and makes the trajectory smoother and more accurate;

[0045] A target fusion trajectory output module, which is connected to the fusion trajectory filtering and smoothing module, and is used to output the finally optimized target fusion trajectory.

[0046] In summary, the present invention has the following beneficial technical effects:

[0047] 1. The multi-source target trajectory fusion method based on deep learning proposed by the present invention, by constructing a spatio-temporal alignment training set containing radar echo, satellite telemetry and true value data, and designing a dual-channel deep learning network with a cross-attention mechanism, has achieved the cross-modal dynamic association modeling of radar time-frequency features and satellite positioning parameters for the first time. Compared with the traditional fixed-weight fusion method, this solution can adaptively generate optimal fusion weight coefficients according to dynamic factors such as target maneuvering state and environmental interference intensity, greatly improving the trajectory positioning accuracy in complex electromagnetic environments and effectively solving the problem of feature mismatch caused by the dimensional differences of heterogeneous data.

[0048] 2. The sliding window mechanism and coordinate system conversion method adopted by the present invention, combined with the cubic spline interpolation method, resamples the dual-source data in time axis, realizes high-precision spatiotemporal alignment of the sampled data, solves the problem of differences in spatiotemporal benchmarks between different data sources, provides a reliable basis for subsequent data fusion, and improves the accuracy and stability of the entire trajectory fusion process. Whether it is when the target's high-speed maneuvers cause drastic changes in data, or when the data fluctuates abnormally in a complex environment, the accuracy and integrity of the data can be improved, thereby improving the accuracy and stability of the entire trajectory fusion process, so that the fused trajectory can more truly reflect the actual movement of the target.

[0049] 3. The present invention optimizes network parameters through back propagation, defines the mean square error loss function of the fusion trajectory and the true value trajectory, and uses the early stopping method to control the overfitting of the model. The obtained weight prediction model can dynamically adjust the fusion weight according to different data characteristics. Compared with the traditional fixed weight fusion mechanism, the fusion weight can be adjusted more flexibly and dynamically. Whether facing sudden changes in the intensity of environmental interference or complex changes in the maneuvering state of the target, it can respond quickly and adjust the weight distribution to make the fused target motion trajectory more accurate, greatly improving the adaptability and accuracy of the model.

[0050] 4. The cross-attention module of the present invention constructs a joint mapping mechanism of radar features and telemetry features, and establishes a nonlinear association model of dual-source features through query-key space conversion and softmax weight allocation. The module significantly improves the trajectory continuity retention rate in scenarios where a single signal is blocked, significantly enhances the anti-interference ability in complex environments, and provides a strong guarantee for the application of multi-source target trajectory fusion in complex environments, ensuring more accurate acquisition of target motion trajectory information in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of the method of embodiment 1 of the present invention;

[0052] Figure 2 is a schematic diagram of the system structure of Embodiment 2 of the present invention; DETAILED DESCRIPTION

[0053] The present invention is further described in detail below in conjunction with the accompanying drawings. Example 1

[0054] Reference Figure 1 , a multi-source target trajectory fusion method based on deep learning in this embodiment includes:

[0055] S1. Construct a spatiotemporal alignment training set containing radar echo data, satellite telemetry data, and true trajectory data, including:

[0056] S11. Perform pulse compression processing on the original radar echo data, extract error parameters, and construct a feature matrix;

[0057] Perform matched filtering pulse compression processing on the original radar echo signal to generate a time-frequency two-dimensional feature matrix, extract the scattering feature vector to obtain the target position. During the feature extraction process, synchronously extract the signal-to-noise ratio (SNR) of the signal as the error feature parameter. By introducing the error feature parameter, the dynamic change of the radar measurement quality can be effectively characterized, providing a quantitative basis for subsequent weight allocation. Concatenate the scattering feature vector and the error parameter into a composite feature vector, with the dimension of [timestamp × target position × SNR].

[0058] S12. Perform coordinate system conversion processing on the satellite telemetry data, extract error parameters, and construct a feature matrix;

[0059] Use the conversion formula from the WGS84 coordinate system to the ENU coordinate system to convert the original satellite positioning data to the radar station-centered coordinate system. During the conversion process, synchronously extract the dilution of precision (DOP) parameter of the satellite system as the error feature parameter, and construct a three-dimensional feature matrix with the dimension of [timestamp × target position × DOP]. The DOP parameter includes components such as HDOP and VDOP, which are used to quantify the characteristics of the satellite positioning accuracy changing with the constellation geometry configuration, providing an error characterization basis for subsequent feature fusion.

[0060] S13. Use the sliding window mechanism and cubic spline interpolation method to align and fill the time axis of the dual-source data, and establish a training sample synchronized at the millisecond level.

[0061] Set the adaptive window length L = 2f_max / f_min, where f_max and f_min are the highest and lowest sampling frequencies of the dual-source respectively, and use the time axis sliding step Δt = 1 / (2f_max) to perform time domain alignment on the dual-source data. Then, within the sliding window, check whether there are missing values in the radar echo data and satellite telemetry data. If there are missing data, use the cubic spline interpolation method to fill them. This process, while ensuring the smoothness of the data, helps to reduce errors caused by environmental factors (such as high radar signal-to-noise ratio or weak satellite signal) by filling the missing data within the window, and also solves the problem of inconsistent sampling frequencies between the radar and the satellite, providing a high-quality data set for subsequent model training.

[0062] S2. Design a multi-channel deep learning network architecture, including an echo data feature extraction channel and a telemetry data feature extraction channel, and establish dual-channel feature correlation through the attention mechanism, including:

[0063] S21. Radar echo data feature extraction channel, using a three-dimensional convolutional neural network to extract spatio-temporal-error features;

[0064] The input layer receives a composite feature vector with dimensions [timestamp × target location × SNR]. The first-layer convolutional kernel is designed as a three-dimensional (3×3×16) spatio-temporal-error convolution, which can fully extract multi-dimensional features. The second layer uses a 3×3 dilated convolution, which can expand the receptive field and capture more extensive feature information without increasing the number of parameters. The last layer uses a 1×1 convolution for feature compression to remove redundant information and refine key features. An attention gating module is connected after each convolution layer, which dynamically adjusts the feature weights according to the SNR values of each time slice.

[0065] S22. For the satellite telemetry data feature extraction channel, a bidirectional LSTM network is used to decode the satellite positioning parameter sequence;

[0066] A bidirectional LSTM network with 64 hidden units is constructed, and the input sequence is a triple feature of [timestamp × target location × DOP]. During the time reverse propagation, a scaling factor with tanh activation is applied to the DOP parameter. When DOP > 3, the output feature decays by 30%, and when DOP < 2, the feature is enhanced by 20%. Through this DOP-sensitive feature modulation, the network can autonomously suppress the influence of positioning data in high-error periods.

[0067] S23. The cross-attention module establishes a correlation matrix between radar features and telemetry features with environmental quality perception, including:

[0068] S231. Map the radar feature vector and the telemetry feature vector to the query space and the key-value space respectively;

[0069] Design bilinear mapping matrices \(W_q\in R^{d\times d}\) and \(W_k\in R^{d\times d}\), where \(d\) is the feature dimension. The radar feature \(F_r\) generates a query vector through \(F_q = F_rW_q\), and the satellite feature \(F_s\) generates a key vector through \(F_k = F_sW_k\). In particular, the error parameters SNR of the radar echo data and DOP of the satellite telemetry data are retained as additional feature channels during the mapping process, enabling the attention mechanism to simultaneously perceive the environmental quality parameters.

[0070] S232. Calculate the feature similarity matrix and generate the attention weight distribution through the softmax function;

[0071] Calculate the similarity between the radar feature vector and the telemetry feature vector by dot product to obtain a feature similarity matrix. Each element in this matrix represents the similarity between the corresponding elements in the radar feature vector and the telemetry feature vector. Then, input the feature similarity matrix into the softmax function, and the softmax function normalizes each element in the similarity matrix so that the sum of all elements is 1, thereby generating an attention weight distribution.

[0072] S233. Weightedly fuse the dual-channel features to generate a joint feature representation and input it into the fully connected layer.

[0073] According to the generated attention weight distribution, weightedly fuse the radar feature vector and the telemetry feature vector to obtain a joint feature representation. This joint feature representation combines the information of the radar features and the telemetry features, effectively integrating the advantages of the dual-source data. Then input the joint feature representation into the fully connected layer for further processing and feature integration.

[0074] S24. The fully connected output layer generates a weight distribution tensor with dimensions [timestamp × spatial coordinates].

[0075] The fully connected layer performs further linear transformation and feature integration on the joint features through a series of operations of weight matrices and bias vectors. Finally, it outputs a weight distribution tensor with dimensions [timestamp × spatial coordinates], where each element in this tensor represents the weight allocation of the radar data and the telemetry data during the fusion process at different timestamps and spatial coordinates.

[0076] S3. Optimize the network parameters through backpropagation to obtain a weight prediction model, including:

[0077] S31. Initialize the network parameters and define the mean squared error loss function between the fusion trajectory and the ground truth trajectory;

[0078] During the training process, first initialize the parameters of the deep learning network and use the mean squared error (MSE) as the loss function. This loss function can measure the gap between the predicted trajectory and the ground truth trajectory, further optimizing the performance of the network. In addition, the loss function also takes into account the impact of environmental data, enabling the network to adaptively adjust during training and reduce the propagation of errors.

[0079] S32. Update the network weights using backpropagation and the stochastic gradient descent algorithm;

[0080] During the model training process, the dual-source feature matrix constructed from the training set is input into the network. Through feature extraction in each channel, correlation calculation in the cross-attention module, and processing in the fully connected layer, the prediction result of the fused trajectory is obtained. Then, according to the defined mean square error loss function, the error between the prediction result and the true trajectory is calculated. Using the backpropagation algorithm, in the opposite direction of the network's forward propagation, the error is propagated to each layer of the network in turn, the gradient of each layer's parameters is calculated, and then through the stochastic gradient descent algorithm, the weight parameters in the network are updated to gradually reduce the value of the loss function.

[0081] S33. Control model overfitting through early stopping and save the optimal weight prediction model.

[0082] The specific approach is to divide the training dataset into a training set and a validation set. After each training iteration, the validation set is used to evaluate the model and calculate the mean square error loss on the validation set. When it is found that the loss on the validation set no longer decreases or even increases continuously for several times (such as 3 times), it is considered that the model may have overfitting. At this time, stop the model training in time and save the weight parameters obtained from the current training. The model corresponding to these parameters is the optimal weight prediction model, ensuring the reliability and generalization ability of the model in practical applications.

[0083] S4. Real-time receive radar echo data and satellite telemetry data, and achieve spatio-temporal alignment of the sampled data through a sliding window mechanism and coordinate system transformation, including:

[0084] S41. Perform pulse compression processing on the original radar echo data and extract error parameters to construct a feature matrix;

[0085] Similar to step S11, perform matched filtering pulse compression processing on the original radar echo signal to generate a time-frequency two-dimensional feature matrix, extract the scattering feature vector to obtain the target position. During the feature extraction process, synchronously extract the signal-to-noise ratio (SNR) of the signal as an error feature parameter. Then, splice the scattering feature vector and the error parameter into a composite feature vector with a dimension of [timestamp × target position × SNR].

[0086] S42. Perform coordinate system transformation processing on the satellite telemetry data and extract error parameters to construct a feature matrix;

[0087] Similar to step S12, use the conversion formula from the WGS84 coordinate system to the ENU coordinate system to convert the original satellite positioning data to the radar station-centered coordinate system. During the conversion process, synchronously extract the dilution of precision (DOP) parameter of the satellite system as an error feature parameter and construct a three-dimensional feature matrix with a dimension of [timestamp × target position × DOP].

[0088] S43. Use the sliding window mechanism and cubic spline interpolation method to align and fill the time axis of the dual-source data.

[0089] Similar to step S13, set the adaptive window length L = 2f_max / f_min, where f_max and f_min are the highest and lowest sampling frequencies of the dual sources respectively, and perform time-domain alignment on the dual-source data with the time axis sliding step Δt = 1 / (2f_max). Then, within the sliding window, check whether there are missing values in the radar echo data and satellite telemetry data. If there are missing data, use the cubic spline interpolation method to fill them.

[0090] S5. Input the preprocessed dual-source features into the trained weight prediction model to generate the radar-telemetry fusion weight coefficients.

[0091] Input the radar echo data, a three-dimensional feature matrix in the format of [timestamp × target position × DOP], and the satellite telemetry data, a three-dimensional feature matrix in the format of [timestamp × target position × SNR], into the trained weight prediction model. After the calculation of the deep learning network, the model will generate the fusion weight coefficients α and β according to the characteristics of each data source. By adjusting these coefficients, the model can dynamically optimize the fusion ratio of radar and satellite data according to the environment and target state.

[0092] S6. Perform weighted processing and optimization on the dual-source trajectory coordinates, and output the fused target motion trajectory, including:

[0093] S61. Perform weighted calculation on the radar trajectory point P_r and the telemetry trajectory point P_s: P_fused = α·P_r + β·P_s, where P_fused is the fused point coordinate, and α and β are the fusion weight coefficients of radar data and telemetry data respectively;

[0094] In this step, the radar trajectory point P_r and the satellite telemetry trajectory point P_s will be weighted calculated according to the previously calculated weight coefficients α and β. Specifically, the final coordinate P_fused of each trajectory point is the result of the fusion of radar and satellite data, and the contributions of the two data sources are adjusted by the fusion weights α and β. This weighted process can be dynamically adjusted according to different environmental factors to ensure that in practical applications, the fusion of radar data and satellite data can accurately reflect the target's motion trajectory.

[0095] S62. Perform moving average filtering on the fused trajectory to remove noise and output the optimized target motion trajectory.

[0096] Use a Savitzky-Golay filter with an adaptive window length, and the window size , where SNR_avg is the average radar signal-to-noise ratio in the current window, and DOP_avg is the average satellite precision factor. The smoothing window is automatically increased to 15 points in low signal-to-noise ratio or high DOP periods, and reduced to 5 points in high signal-to-noise ratio and low DOP periods, to achieve environment-adaptive trajectory optimization. After filtering, the output target motion trajectory is smoother and more accurate, and can provide more reliable target position prediction, especially in real-time dynamic environment, the tracking of target motion has stronger robustness.

[0097] This embodiment provides a multi-source target trajectory fusion method based on deep learning. The method aims at the background that the traditional single-source trajectory measurement system has insufficient accuracy in complex environments and the existing multi-source fusion technology is difficult to effectively fuse radar echoes and satellite telemetry data. In order to accurately generate the target motion trajectory under complex and changeable electromagnetic environments and various interference conditions, a spatiotemporal alignment training set is constructed to perform spatiotemporal alignment processing on radar echo data and satellite telemetry data, and high-precision spatiotemporal synchronization is achieved by using a cubic spline interpolation method. A multi-channel deep learning network architecture is designed, and dual-channel feature association is established with the help of an attention mechanism, a cross-modal data association model is established, and feature information of the two data sources is deeply mined. The trained weight prediction model is applied to real-time data processing, and the radar echo data and satellite telemetry data received in real time are spatiotemporally aligned through a sliding window mechanism and coordinate system conversion. The model is then input to generate a fusion weight coefficient, and then the dual-source trajectory coordinates are weighted and optimized, and finally a high-precision fused target motion trajectory is output, which significantly improves the accuracy and stability of trajectory measurement. This method can also efficiently and accurately fuse radar and satellite data from different sources and with different characteristics in actual scenarios with high requirements for trajectory measurement accuracy, thereby providing accurate, timely and reliable target trajectory data for these fields with high precision requirements, and effectively improving the reliability of task execution and equipment effectiveness in related fields. Example 2

[0098] Reference Figure 2 , this embodiment provides a multi-source target trajectory fusion system based on deep learning, including:

[0099] The module for building a spatiotemporal alignment training set is used to resample the radar echo data, satellite telemetry data, and true value trajectory in time axis and coordinate system to generate a spatiotemporal synchronized supervised learning dataset;

[0100] The model network architecture building module is connected with the spatiotemporal alignment training set building module to build a deep learning model network architecture suitable for multi-source target trajectory fusion;

[0101] The model training module, connected to the model network architecture construction module, uses the constructed training set to train the built network architecture, generates a dynamic weight prediction model, and saves the optimal parameter model;

[0102] The sampling data spatio-temporal alignment module aligns the multi-source trajectory data obtained in real time, making the data from different sources comparable in the time and space dimensions;

[0103] The fusion weight coefficient generation module, connected to the model training module and the sampling data spatio-temporal alignment module, is used to input the preprocessed dual-source features into the training model to generate the weight coefficients for trajectory fusion;

[0104] The multi-source trajectory weighted processing module, connected to the fusion weight coefficient generation module, is used to perform linear weighted calculations on the radar trajectory points and telemetry trajectory points according to the weight coefficients to generate an initial fusion trajectory;

[0105] The fused trajectory filtering and smoothing module, connected to the multi-source trajectory weighted processing module, performs filtering and smoothing operations on the weighted fused trajectory to remove noise and fluctuations, making the trajectory smoother and more accurate;

[0106] The target fused trajectory output module, connected to the fused trajectory filtering and smoothing module, is used to output the finally optimized target fused trajectory.

[0107] The above are all preferred embodiments of the present invention. The protection scope of the present invention is not limited hereby. Therefore, all equivalent changes made according to the structure, shape, and principle of the present invention shall be covered within the protection scope of the present invention.

Claims

1. A multi-source target trajectory fusion method based on deep learning, characterized in that: include: Construct a spatiotemporal alignment training set containing radar echo data, satellite telemetry data, and true trajectory data; Design a multi-channel deep learning network architecture, including radar echo data feature extraction channel and satellite telemetry data feature extraction channel, and establish dual-channel feature association through attention mechanism; The network parameters are optimized through back propagation to obtain the weight prediction model; Receive radar echo data and satellite telemetry data in real time, and achieve spatial and temporal alignment of sampled data through sliding window mechanism and coordinate system conversion; The preprocessed dual-source features are input into the trained weight prediction model to generate radar-telemetry fusion weight coefficients; Perform weighted processing and optimization on the dual-source trajectory coordinates, and output the fused target motion trajectory; The construction of a spatiotemporal alignment training set containing radar echo data, satellite telemetry data, and true value trajectory data includes: Perform pulse compression processing on the original radar echo data and extract error parameters to construct a feature matrix; Perform coordinate system conversion on satellite telemetry data, extract error parameters, and construct feature matrix; The sliding window mechanism and cubic spline interpolation method are used to align and fill the time axis of the dual-source data, and establish millisecond-level synchronized training samples; The multi-channel deep learning network architecture is designed, including a radar echo data feature extraction channel and a satellite telemetry data feature extraction channel, and dual-channel feature association is established through an attention mechanism, including: Radar echo data feature extraction channel, using a three-dimensional convolutional neural network to extract time-space-error features; Satellite telemetry data feature extraction channel, using a bidirectional LSTM network to decode satellite positioning parameter sequences; Cross-attention module, which builds the correlation matrix between radar features and telemetry features containing environmental quality perception; The fully connected output layer generates a weight distribution tensor of dimension [timestamp × spatial coordinate]; The cross attention module calculates the correlation matrix between radar features and telemetry features, including: Mapping radar feature vectors and telemetry feature vectors to query space and key value space respectively; Calculate the feature similarity matrix and generate the attention weight distribution through the softmax function; The dual-channel features are weightedly fused to generate a joint feature representation and input into the fully connected layer.

2. According to a multi-source target trajectory fusion method based on deep learning according to claim 1, it is characterized in that: The method of optimizing the network parameters by back propagation to obtain the weight prediction model includes: Initialize network parameters and define the mean square error loss function between the fusion trajectory and the true value trajectory; Use back propagation and stochastic gradient descent algorithms to update network weights; The model overfitting is controlled by early stopping method, and the optimal weight prediction model is saved.

3. The multi-source target trajectory fusion method based on deep learning according to claim 1 is characterized in that: The real-time reception of radar echo data and satellite telemetry data and the realization of spatiotemporal alignment of the sampled data through a sliding window mechanism and coordinate system conversion include: Perform pulse compression processing on the original radar echo data and extract error parameters to construct a feature matrix; Perform coordinate system conversion on satellite telemetry data, extract error parameters, and construct feature matrix; The sliding window mechanism and cubic spline interpolation method are used to align and complete the time axis of the dual-source data.

4. The multi-source target trajectory fusion method based on deep learning according to claim 1 is characterized in that: The weighted processing and optimization of the dual-source trajectory coordinates and the output of the fused target motion trajectory include: Perform weighted calculation on the radar trajectory point P_r and the telemetry trajectory point P_s: P_fused = α·P_r + β·P_s, where P_fused is the coordinate of the fusion point, α and β are the fusion weight coefficients of radar data and telemetry data respectively; Perform sliding average filtering on the fused trajectory to remove noise and output the optimized target motion trajectory.

Images