Electric detection-radar heterogeneous signal positioning method based on encoder-decoder network

Abstract

The invention relates to the technical field of signal processing and data fusion, and discloses an electric detection-radar heterogeneous signal positioning method based on an encoder-decoder network, and the method comprises the steps: obtaining a radar and an electric detection data sequence which are aligned in a space-time manner; performing deep interactive fusion by using a hybrid feature encoder, and outputting a fused feature sequence; decoding the fusion feature sequence into a probabilistic result containing a position mean value and a covariance matrix by using a probability positioning decoder; carrying out supervised training on the coding and decoding network model by adopting a negative log-likelihood loss function; and inputting real-time data into the trained model, and directly outputting a positioning result through forward propagation. Through deep adaptive fusion and probabilistic decoding, the problems of shallow fusion level and weak interference resistance are solved, and the precision, robustness and reliability of heterogeneous signal positioning are improved.

Description

Technical Field

[0001] This invention relates to the field of signal processing and data fusion technology, specifically to a method for locating heterogeneous signals from electronic reconnaissance and radar based on encoder-decoder networks. Background Technology

[0002] In the field of modern electronic warfare, rapid and accurate location of radiation source targets is a core element for gaining battlefield information superiority and implementing effective decision-making. Currently, radar and electronic reconnaissance (referred to as "electronic reconnaissance") are two mainstream and highly complementary detection methods.

[0003] To leverage synergistic advantages, the industry typically employs heterogeneous fusion positioning technology. Radar, as an active sensor, can provide high-precision radial distance and velocity information for targets, but its own position is easily revealed during operation. Electronic reconnaissance, as a passive sensor, can covertly intercept signals and accurately measure their direction of arrival, but it usually cannot directly obtain distance information, resulting in limited positioning accuracy and spatial resolution when used alone. Traditional fusion positioning methods often employ classical filtering algorithms and other techniques to post-process and fuse track data from two independent processing links: radar and electronic reconnaissance.

[0004] However, existing fusion localization methods have inherent limitations. First, these methods often involve loosely coupled, shallow fusion at the data or decision level, lacking mechanisms for deep interaction and mining of heterogeneous data at the feature level, making it difficult to fully utilize the complex temporal correlations between data. Second, traditional methods cannot dynamically evaluate and adjust the information contribution weights of different sensors based on real-time signal quality; when any data source is interfered with or its quality degrades, the robustness of the fusion system significantly decreases. More importantly, they typically only output a deterministic location estimate, failing to simultaneously provide a quantitative assessment of the uncertainty of that estimate. This greatly limits the reliability and practical value of the localization results in real-world decision-making applications.

[0005] Therefore, this invention proposes a method for locating heterogeneous signals from electronic reconnaissance and radar based on encoder-decoder networks to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network. This method aims to solve the problems in existing technologies, such as insufficient positioning accuracy and robustness due to shallow feature fusion of heterogeneous signals, inability to adaptively weight signals based on real-time quality, and positioning results that are only deterministic point estimates, as well as the difficulty in quantifying the reliability of the results.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network, comprising the following steps: Acquire radar data sequences and electronic reconnaissance data sequences that have been timestamped and spatially aligned, and associate the radar data sequences and the electronic reconnaissance data sequences as heterogeneous signal sample pairs; A hybrid feature encoder is constructed, and the hybrid feature encoder is used to perform feature extraction and deep interactive fusion on the heterogeneous signal sample pairs to output the fused radar feature sequence and electronic reconnaissance feature sequence. Construct a probabilistic localization decoder, and use the probabilistic localization decoder to receive the fused feature sequence, and decode the fused feature sequence into the probabilistic localization result of the target, wherein the probabilistic localization result of the target contains the mean and covariance matrix of the position coordinates; Based on a labeled dataset containing real target locations, a negative log-likelihood loss function is used to supervise the training of the encoder-decoder network model composed of the hybrid feature encoder and the probabilistic localization decoder until the encoder-decoder network model converges. The real-time acquired radar data sequences and electronic reconnaissance data sequences are input into the encoder-decoder network model that has been trained and converged. Through the forward propagation calculation of the encoder-decoder network model, the probabilistic localization result of the target is directly output.

[0008] Preferably, the step of acquiring the radar data sequence and electronic reconnaissance data sequence after timestamp and spatial alignment specifically includes: The radar data sequence is formed by extracting kinematic and property features from the raw radar echo signal; Key parameters are extracted from the original pulse descriptor or signal stream of the electronic reconnaissance system to form the electronic reconnaissance data sequence; Timestamp alignment is achieved through a unified time base and sampling interval, and spatial alignment is achieved by setting a spatial association gate.

[0009] Preferably, the radar data sequence includes at least one of the target's radial range, radial velocity, and signal-to-noise ratio; the electronic reconnaissance data sequence includes at least one of the target's direction of arrival, signal strength, carrier frequency, and pulse width.

[0010] Preferably, the step of using the hybrid feature encoder for feature extraction and deep interactive fusion specifically includes: Local feature extraction based on a one-dimensional convolutional neural network is performed on the radar data sequence and the electronic reconnaissance data sequence, respectively. Temporal dependency modeling based on long short-term memory network is performed on the sequence after local feature extraction; Deep interactive fusion based on a gated guided sparse cross-attention transformer is performed on the sequence after the temporal dependency modeling.

[0011] Preferably, the deep interaction fusion step based on the gated guided sparse cross-attention transformer specifically includes: Explicitly estimate the characteristic uncertainty of two time-series characteristic sequences; Based on the aforementioned feature uncertainty, calculate the corresponding confidence mask; Perform gated reciprocal sparse cross-attention computation.

[0012] Preferably, the confidence mask is calculated as follows: The feature uncertainty is multiplied by a learnable scalar parameter and inverted, then mapped using the sigmoid function.

[0013] Preferably, the estimation of the feature uncertainty is achieved through an uncertainty prediction network; the uncertainty prediction network is a multilayer perceptron and uses the Softplus activation function to ensure that the output feature uncertainty is always positive.

[0014] Preferably, the step of decoding using the probabilistic localization decoder specifically includes: The fused radar feature sequence and electronic reconnaissance feature sequence are aggregated to form a fixed-length fused feature vector. The fused feature vector is processed using a multilayer perceptron network with two independent output branches, which outputs the mean of the location coordinates and the covariance matrix, respectively.

[0015] Preferably, the covariance matrix is ​​output in the following manner: The output vector of one output branch of the multilayer perceptron network; Construct a lower triangular matrix based on the vector.

[0016] Preferably, the encoder-decoder network model training and optimization uses a negative log-likelihood loss function, which includes a Mahalanobis distance term and a regularization term.

[0017] This invention provides a method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network. It offers the following advantages: 1. This invention constructs an innovative hybrid feature encoder within an encoder-decoder network. This encoder utilizes a gated, guided sparse cross-attention mechanism to achieve deep interactive fusion of two heterogeneous signals: radar and electronic reconnaissance. This method can explicitly evaluate the feature uncertainties of each signal at different time steps and actively suppress the interference of low-confidence information on the fusion result. Thus, in complex heterogeneous signal localization tasks, it achieves higher-quality feature synergy and complementarity, providing a solid foundation for subsequent accurate localization.

[0018] 2. This invention significantly improves the accuracy and completeness of heterogeneous signal localization by designing a probabilistic localization decoder at the end of the encoder-decoder network. This method effectively integrates precise range information from radar and precise azimuth information from electronic reconnaissance, and directly calculates the probabilistic localization result of the target—namely, the mean and covariance matrix of the position coordinates—through a specific network structure. Compared to traditional methods, this approach not only fundamentally improves the accuracy of the localization results but also provides a quantitative measure of localization uncertainty, making the results more complete and reliable.

[0019] 3. This invention significantly enhances the robustness and anti-interference capability of heterogeneous signal localization systems by introducing a gated, reciprocal, sparse, cross-attention mechanism into the encoder. This encoder-decoder network can estimate and utilize feature uncertainties online. When the quality of any sensor signal deteriorates or is interfered with, this attention mechanism automatically adjusts the weights and actively suppresses the contribution of low-confidence information. This adaptive fusion strategy enables the method to maintain stable and reliable heterogeneous signal localization capabilities even in complex electromagnetic environments or when some data is damaged. Attached Figure Description

[0020] Figure 1 This is a flowchart of the electronic reconnaissance-radar heterogeneous signal localization method based on encoder-decoder network of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the encoder-decoder network model of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 This invention provides a method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network, comprising the following steps: S1. Data Preprocessing and Association. This involves acquiring timestamped and spatially aligned radar and electronic reconnaissance data sequences, and then association them into heterogeneous signal sample pairs. S2. Construct a hybrid feature encoder. The hybrid feature encoder is used to extract features and perform deep interactive fusion on the heterogeneous signal sample pairs output from step S1, to output fused radar feature sequences and electronic reconnaissance feature sequences. S3. Construct a probabilistic localization decoder. Receive the fused feature sequence output from step S2 using the probabilistic localization decoder and decode it into a probabilistic localization result for the target, which includes the mean of the position coordinates and the covariance matrix. S4. Model Training and Optimization. Based on a labeled dataset containing real target locations, the encoder-decoder network model, composed of the encoder from step S2 and the decoder from step S3, is trained in a supervised manner using the negative log-likelihood loss function until the model converges. S5. Target Localization. The real-time acquired radar and electronic reconnaissance data sequences are input into the trained and converged encoder-decoder network model. Through the forward propagation calculation of the model, the probabilistic localization result of the target is directly output.

[0023] The following will provide a detailed description of each of the steps described in the embodiments of the present invention.

[0024] In detail, the data preprocessing and association in step S1 includes the following sub-steps: S1.1: Acquisition of radar data sequences.

[0025] The received raw radar echo signal undergoes a series of signal processing operations, including but not limited to pulse compression, Doppler filtering, constant false alarm rate (CFAR) detection, and target tracking. From the stably tracked target trajectory, signals are processed at preset time intervals. Sampling is performed to extract key kinematic and property features, forming radar data sequences.

[0026] Radar data sequences are represented as Its mathematical form is: ; in, This represents the time step of the sequence, i.e., the length of the sequence. This represents the characteristic dimension of radar data at each time step; It is the first The radar feature vector at each time step, here It is an index used to represent the vector in the time series. The position in the middle, This refers to the sequence number of this time step, which ranges from 1 to... The radar feature vector contains features such as the target's radial range, radial velocity, and signal-to-noise ratio (SNR).

[0027] S1.2: Acquisition of electronic reconnaissance data sequence.

[0028] The passively intercepted raw pulse descriptor words (PDWs) or signal streams from electronic surveillance are processed, including parameter measurement, pulse sorting, and radiation source identification. From the identified specific radiation source signals, the signals are also processed at time intervals. Perform sampling or time alignment to extract key parameters.

[0029] Electronic surveillance data sequences are represented as Its mathematical form is: ; in, It has the same sequence length definition as radar data sequences; This represents the feature dimension of the electronic surveillance data at each time step; It is the first The electronic reconnaissance feature vector at each time step, The value range is from 1 to The electronic reconnaissance feature vector contains features such as the target's direction of arrival (DOA, such as azimuth and elevation angles), signal strength, carrier frequency, pulse width, and other parameters.

[0030] S1.3: Data association.

[0031] This step aligns the timestamp with the spatial location.

[0032] Timestamp alignment is achieved through a unified time base and sampling interval. To achieve, to ensure and The first in element and It describes the target state within the same time window.

[0033] Spatial alignment is achieved by setting spatial association gates. For example, using azimuth information provided by electronic reconnaissance data. This serves as a spatial threshold, filtering radar tracking tracks within that azimuth area. This ensures that both data sequences point to the same physical target.

[0034] After the above steps, the final result is This forms a spatiotemporally aligned heterogeneous signal sample pair, which serves as the input to the subsequent encoder-decoder network model.

[0035] See attached document Figure 1 and Figure 2 Step S2 involves constructing a hybrid feature encoder, which serves as the encoder part of the encoder-decoder network model. The hybrid feature encoder employs a dual-branch structure, corresponding to the radar data sequences obtained in step S1. and electronic reconnaissance data sequence Each branch contains three sequential processing stages to achieve a progressive, layer-by-layer processing from local feature extraction and temporal dependency modeling to deep interaction of heterogeneous features.

[0036] S2.1: Local feature extraction based on one-dimensional convolutional neural network (CNN).

[0037] This stage utilizes a one-dimensional convolutional neural network (1D-CNN) to process the two input sequences separately. The 1D-CNN slides its convolutional kernels along the time dimension to capture local patterns or short-term correlation features in the signal sequence, such as the pulse edge features of radar signals or the parameter transient features of electronic reconnaissance signals.

[0038] For radar data sequences Radar local feature sequence The calculation process is as follows: ; in It is the input radar data sequence; This represents a one-dimensional convolution operation; and These are the learnable convolutional kernel weights and biases of the CNN layer in the radar branch; It is a modified linear unit activation function, used to introduce nonlinearity; It is the radar local feature sequence output by the CNN layer.

[0039] Similarly, for electronic surveillance data sequences Apply independent 1D-CNN layers (with independent weights) and bias ), to obtain the local feature sequence of electronic detection .

[0040] S2.2: Temporal dependency modeling based on Long Short-Term Memory (LSTM) network.

[0041] This stage receives the local feature sequence output from stage one. and They are then processed separately using Long Short-Term Memory (LSTM) networks. LSTM, through its unique internal gating mechanism (input gate, forget gate, output gate), can effectively learn and capture long-term contextual dependencies and dynamic evolution patterns in feature sequences.

[0042] At each time step of the sequence The LSTM unit receives the local features at the current time step. (from) The (vectors of each time step) and the vector of the previous time step Hidden state Its internal calculation process is defined as follows:

[0043] in, This indicates that two vectors are concatenated; , , They are The activation values ​​of the input gate, forget gate, and output gate at any given time; yes The state of candidate cells at any given time; and They are Cellular state and hidden state at any given moment; yes Cellular state at any given moment; , , , It is the learnable weight matrix of the LSTM unit; , , , It is the learnable bias vector of the LSTM unit; It is the Sigmoid activation function; It is the hyperbolic tangent activation function; This indicates element-wise multiplication.

[0044] This embodiment retains all time steps. Hidden state The complete sequence is used as the output of the LSTM layer, and is denoted as the radar time-series feature sequence. and electronic detection time series feature sequences .

[0045] S2.3: Deep interactive integration based on GSCA-Former.

[0046] This stage is the core of the hybrid feature encoder of this invention; it receives the two temporal feature sequences output from stage two. and And through a gated sparse cross-attention transformer (GSCA-Former), they are subjected to deep, reciprocal interaction and fusion. GSCA-Former is composed of... layer( For example, positive integers. It is composed of stacked GSCA blocks with the same structure.

[0047] Let the first layer( The input to the GSCA block is and (where the input of the first layer) ). In the Inside the GSCA layer block, the following calculation steps are performed sequentially: First, multi-head self-attention (MHSA) is independently computed for each of the two feature sequences, combined with residual connections and layer normalization. This step aims to enable information exchange between elements within each sequence (i.e., features at different time steps) to capture temporal correlations within their respective modalities.

[0048] ; ; in, and It is the first The input feature sequence of the GSCA block; This represents a standard multi-head self-attention operation; Presentation layer normalization operation; and It is the feature sequence after in-modal self-attention update.

[0049] Secondly, for and For each time step of the feature vector, its uncertainty is explicitly estimated as a measure of its reliability. This estimation is achieved through an uncertainty prediction network (UncertaintyNet), which can be a small multilayer perceptron (MLP) with shared or independent parameters.

[0050] ; ; in, and It is an uncertainty prediction network; and It is the estimated characteristic uncertainty sequence, whose dimension is related to the sequence length. Consistent, each element represents the prediction variance of the feature at the corresponding time step; It is a smooth activation function, defined as follows: Used to ensure the characteristic uncertainty of the output It is always a positive value.

[0051] Next, gate-guided reciprocal sparse cross-attention computation is performed. This step is crucial for achieving deep collaboration of heterogeneous features. This computation is guided by the feature uncertainty estimated in the previous step.

[0052] Specifically, firstly, based on the estimated characteristic uncertainty... Calculate the corresponding credibility mask : ; ; in, It is a learnable scalar parameter used to adjust the scale of the impact of uncertainty on credibility; The function maps the input to the interval (0, 1); and It is a confidence mask sequence. This calculation introduces feature uncertainty. The higher the feature's confidence level, the higher its corresponding confidence mask value. The closer a feature is to 0 (indicating low confidence), the lower its uncertainty. The closer the value is to 1 (indicating high confidence).

[0053] Then, in the Gated Multi-Head Cross-Attention (GMHCA) module, reciprocal cross-attention and gating modulation of attention scores are performed. This reciprocity computation includes: 1) Radar sequence update: Receive As a query (Queny, Q), As keys (K) and values ​​(V), and using the credibility mask corresponding to the electronic detection sequence. As a gating signal (Mask, M).

[0054] 2) Electronic reconnaissance sequence update: Receive As Q, As K and V, and using the confidence mask corresponding to the radar sequence. As M.

[0055] Within the gated multi-head cross-attention module, taking radar sequence update as an example, the gated multi-head cross-attention module calculates Q(from... ) and K (from After the original attention scores between K and V, this score, before applying Softmax normalization, is compared with the confidence mask corresponding to the feature sequence (i.e., the electro-detection sequence) that serves as K and V. Element-wise multiplication is performed. This gated modulation operation actively suppresses the contribution of information from the high-uncertainty (low-confidence) portion of the electronic reconnaissance sequence to the attention weight calculation.

[0056] Combining residual connections and layer normalization, the output of the gated multi-head cross-attention module is obtained: ; ; in, and It is the feature sequence after gating and cross-attention update.

[0057] Finally, the output of the gated multi-head cross-attention module... A Position-wise Feed-Forward Network (FFN) is applied for further nonlinear transformations to enhance the model's representational power. This FFN typically consists of two fully connected layers and a ReLU activation function between them.

[0058] ; ; in, Indicates the position feedforward network operation; and That is, the first The final output of the GSCA layer block.

[0059] S2.4: Encoder output.

[0060] No. Layer output Will be the first Layer input Repeat the calculation in S2.3. After... After stacking the GSCA blocks, the hybrid feature encoder finally outputs the first GSCA block. Layer and As two feature sequences that have undergone deep fusion, let them be denoted as and These two sequences, containing fully interacted and refined heterogeneous information, will be sent to the probabilistic localization decoder in step S3 for processing.

[0061] See attached document Figure 1 and Figure 2 Step S3 constructs a probabilistic localization decoder, which serves as the decoder part of the encoder-decoder network model. This decoder receives the two deep fusion feature sequences ultimately output from step S2. and The solution is then converted into a probabilistic localization result for the target. This process includes two stages: feature aggregation and probability output.

[0062] S3.1: Feature aggregation.

[0063] This stage will use two variable-length time-series feature sequences. and It is converted into a fixed-length fused feature vector.

[0064] Concatenate the two feature sequences along the feature dimension: ; in, and These are the final fused feature sequences from radar and electronic reconnaissance, respectively. This indicates that the concatenation is performed on the second dimension (feature dimension); It is the spliced ​​temporal feature sequence.

[0065] spliced ​​time-series feature sequences Apply one-dimensional average pooling to it in the time dimension. The aggregation is performed to obtain a fixed-length fused feature vector. : ; in, This represents the average value of the feature vectors calculated along the time dimension for all time steps; It is the fused feature vector after aggregation.

[0066] S3.2: Output of probabilistic positioning results.

[0067] This stage utilizes a multilayer perceptron (MLP) network (as the main decoder) for processing. The MLP network outputs the probabilistic localization result of the target. It has two independent output branches (output heads), used for predicting the mean and covariance matrix of the position coordinates, respectively.

[0068] Fusion feature vectors First, high-level abstract features are obtained through the shared hidden layer of the MLP network. : : in, and These are the learnable weights and biases of the hidden layers in an MLP; These are the hidden features of the output.

[0069] Mean branch: This mean branch consists of a linear fully connected layer that receives... and output Mean of position coordinates of dimension (For example, when) When, corresponding plane coordinates

[0070] ; in, and These are the learnable weights and biases of the mean branch output layer; It is the average of the predicted location coordinates.

[0071] Covariance branch: This covariance branch is used for output. dimensional covariance matrix To ensure the accuracy of the predictions The matrix is ​​always valid (i.e., a positive definite matrix), and this embodiment uses the Cholesky decomposition method. The output layer of this branch (e.g., a linear fully connected layer) outputs a matrix containing... A vector of elements .

[0072] Here, the Softplus function is used to ensure the lower triangular matrix. The diagonal elements are positive. The definition of the Softplus function is: ; in, This is the input parameter of the function.

[0073] by For example, , containing elements Applying the Softplus function, the lower triangular matrix... The structure is as follows: ; in, , , It is the direct output of the covariance branch network layer; and They are and The result calculated by the Softplus function ensures that... The diagonal elements of a matrix are always positive.

[0074] The final covariance matrix pass and its transpose The product of these two products is: ; This constructor method guarantees that It is always a symmetric positive definite matrix.

[0075] In summary, the final output of step S3 is: The resulting probabilistic localization. Among them, As the best estimate of the target location, and This quantifies the uncertainty of the location estimate (e.g., The diagonal elements represent the prediction variance for each coordinate dimension, and the off-diagonal elements represent the correlation between prediction errors in different dimensions.

[0076] See attached document Figure 1 and Figure 2 S4.1: Training Preparation: Step S4 involves supervised, end-to-end training and optimization of the complete encoder-decoder network model, which consists of the hybrid feature encoder constructed in step S2 and the probabilistic localization decoder constructed in step S3. The training process is based on a... Annotated dataset of 100 samples: ; in, It is the first A heterogeneous signal sample pair after processing and association following step S1; It is the first The true target location coordinates (i.e., labeled values ​​or ground truth values) corresponding to each sample. For the dimensions of the positioning coordinates (e.g.) ).

[0077] S4.2: Definition of loss function.

[0078] This embodiment employs the Negative Log-Likelihood (NLL) loss function to optimize the encoder-decoder network model. This method aims to maximize the ground truth location. The likelihood of the probability distribution predicted by the encoder-decoder network model. Specifically, the probability distribution predicted by the model is a distribution based on the mean of the position coordinates predicted in step S3. Covariance Matrix Let the parameter be a multivariate normal distribution, denoted as . .

[0079] For a single training sample, its negative log-likelihood loss The calculation is as follows (ignoring the constant term). ): ; in, It is the training sample 3D true position vector It is the average of the predicted location coordinates; It is the predicted covariance matrix; It is the prediction error vector transpose; It is the predicted covariance matrix The inverse matrix; It is the prediction covariance matrix The determinant of; It is a natural logarithm operation.

[0080] The negative log-likelihood loss function The first item The mean of the predicted location coordinates Coordinates of the actual target location The square of the Mahalanobis distance between them, which uses the covariance matrix predicted by the model itself. To normalize the prediction error. Second term. It is a regularization term that penalizes excessive uncertainty in the model output, thereby preventing the model from blindly increasing the predicted covariance in order to reduce Mahalanobis distance.

[0081] S4.3: Training iteration process.

[0082] During training, from the labeled dataset A small batch of samples is randomly selected from the dataset. For each sample in this batch, the following operations are performed: Forward propagation: transferring sample data The input is fed into the encoder-decoder network model, and through the calculations in steps S2 and S3, the probabilistic localization result of the sample is obtained. .

[0083] Loss calculation: using the method defined in S4.2 The formula is used to calculate the loss value for this sample. Then, the average loss for this mini-batch is calculated.

[0084] Backpropagation: Based on the calculated batch average loss, the gradient of the loss with respect to all learnable parameters in the encoder-decoder network model is calculated using the backpropagation algorithm. These learnable parameters include the weights and biases of the CNN layers, LSTM layers, and modules within the GSCA-Former in S2, as well as the weights and biases of the MLP decoder in S3.

[0085] Parameter update: An optimization algorithm, such as the Adam (Adaptive Moment Estimation) optimizer or the RMSprop optimizer, is used to update all learnable parameters in the encoder-decoder network model based on the calculated gradients.

[0086] S4.4: Model convergence.

[0087] Repeat the training iteration process of S4.3. After each or several iterations, evaluate the model on a reserved independent validation dataset. Loss. When the loss value on the validation set stabilizes and no longer decreases significantly, the model can be considered to have converged. At this point, training is stopped, the current model parameters are saved, and the finally converged encoder-decoder network model is obtained.

[0088] See attached document Figure 1 Step S5: Target localization.

[0089] S5.1: This step utilizes the encoder-decoder network model that has been trained and converged in step S4 to perform real-time localization calculations on newly received, unknown radiation source target observation data.

[0090] S5.2: Online data processing and model input.

[0091] When a new target needs to be located, the radar echo signal and electronic reconnaissance intercept signal of the target are first acquired in real time. This new observation data will then follow the "data preprocessing and correlation" process in step S1 to generate a spatiotemporally aligned heterogeneous signal sample pair. .

[0092] Subsequently, this real-time acquired and preprocessed radar data sequence was correlated. and electronic reconnaissance data sequence As a complete input sample, it is fed into the trained encoder-decoder network model.

[0093] S5.3: Model forward propagation and localization solution.

[0094] After receiving an input sample, the encoder-decoder network model performs a complete forward propagation (inference) computation. The computation process is as follows: 1. Input data First, the hybrid feature encoder constructed in step S2, after local feature extraction, temporal dependency modeling, and deep interaction fusion, outputs a fused feature sequence containing sufficient heterogeneous information. and .

[0095] 2. Next, the fused feature sequence is fed into the probabilistic localization decoder constructed in step S3. The probabilistic localization decoder aggregates the features and ultimately outputs the probabilistic localization result of the radiation source target. This result is a pair of numbers containing two parts. ,in: It is the average of the predicted location coordinates.

[0096] It is the predicted covariance matrix.

[0097] S5.4: Interpretation and application of positioning results.

[0098] The probabilistic localization result output in step S5.3 It provides a complete description of the target location, including both the best estimate of the location and a quantitative measure of the uncertainty of that estimate.

[0099] Mean of location coordinates This is used as the final, definitive estimate of the geographic location of the radiation source target. For example, in a two-dimensional planar positioning scenario ( ), This is the best estimate of the target plane coordinates calculated by the model.

[0100] covariance matrix This is used to quantify the final deterministic estimate of the aforementioned location. The prediction uncertainty is represented by the diagonal elements of the covariance matrix, which represent the variance of the predicted values ​​in each coordinate dimension, while the off-diagonal elements represent the correlation between prediction errors in different coordinate dimensions.

[0101] Based on this covariance matrix A positioning error confidence region can be constructed, which is typically an error ellipse. ) or error ellipsoid ( The positioning error confidence region represents the range within which the true target location falls with a specific probability. The boundary of this error ellipse is defined by the following equation: ; in, It refers to the position vector of any point on the boundary of the error ellipse; It is the mean of the predicted position coordinates, and also the center point of the error ellipse; It is the predicted covariance matrix, which determines the size, shape, and rotation direction of the error ellipse; It is a scalar constant whose value determines the confidence level of the error ellipse. This value is typically derived from the set of degrees of freedom. Chi-squared The distribution is selected from the given information. For example, to construct a two-dimensional error ellipse with a confidence level of 95% (…), the following method is used: ), The value should be The quantile corresponding to a cumulative probability of 0.95 with 2 degrees of freedom (this value is approximately 5.991).

[0102] In this way, the present invention not only provides a target location for point estimation It also provides an explicit, data-driven covariance matrix for quantifying uncertainty. This greatly improves the reliability of the positioning results and their value in practical applications.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for locating heterogeneous signals from electronic reconnaissance and radar based on encoder-decoder networks, characterized in that, Includes the following steps: Acquire radar data sequences and electronic reconnaissance data sequences that have been timestamped and spatially aligned, and associate the radar data sequences and the electronic reconnaissance data sequences as heterogeneous signal sample pairs; A hybrid feature encoder is constructed, and the hybrid feature encoder is used to perform feature extraction and deep interactive fusion on the heterogeneous signal sample pairs to output the fused radar feature sequence and electronic reconnaissance feature sequence. Construct a probabilistic localization decoder, and use the probabilistic localization decoder to receive the fused feature sequence, and decode the fused feature sequence into the probabilistic localization result of the target, wherein the probabilistic localization result of the target contains the mean and covariance matrix of the position coordinates; Based on a labeled dataset containing real target locations, a negative log-likelihood loss function is used to supervise the training of the encoder-decoder network model composed of the hybrid feature encoder and the probabilistic localization decoder until the encoder-decoder network model converges. The real-time acquired radar data sequences and electronic reconnaissance data sequences are input into the encoder-decoder network model that has been trained and converged. Through the forward propagation calculation of the encoder-decoder network model, the probabilistic localization result of the target is directly output.

2. The method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network according to claim 1, characterized in that, The steps of acquiring the timestamped and spatially aligned radar data sequences and electronic reconnaissance data sequences specifically include: The radar data sequence is formed by extracting kinematic and property features from the raw radar echo signal; Key parameters are extracted from the original pulse descriptor or signal stream of the electronic reconnaissance system to form the electronic reconnaissance data sequence; Timestamp alignment is achieved through a unified time base and sampling interval, and spatial alignment is achieved by setting a spatial association gate.

3. The method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network according to claim 2, characterized in that, The radar data sequence includes at least one of the target's radial range, radial velocity, and signal-to-noise ratio; the electronic reconnaissance data sequence includes at least one of the target's direction of arrival, signal strength, carrier frequency, and pulse width.

4. The method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network according to claim 1, characterized in that, The steps of feature extraction and deep interactive fusion using the hybrid feature encoder specifically include: Local feature extraction based on a one-dimensional convolutional neural network is performed on the radar data sequence and the electronic reconnaissance data sequence, respectively. Temporal dependency modeling based on long short-term memory network is performed on the sequence after local feature extraction; Deep interactive fusion based on a gated guided sparse cross-attention transformer is performed on the sequence after the temporal dependency modeling.

5. The method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network according to claim 4, characterized in that, The steps of deep interactive fusion based on the gated guided sparse cross-attention transformer specifically include: Explicitly estimate the characteristic uncertainty of two time-series characteristic sequences; Based on the aforementioned feature uncertainty, calculate the corresponding confidence mask; Perform gated reciprocal sparse cross-attention computation.

6. The method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network according to claim 1, characterized in that, The confidence mask is calculated as follows: The feature uncertainty is multiplied by a learnable scalar parameter and inverted, then mapped using the sigmoid function.

7. The method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network according to claim 5, characterized in that, The estimation of the feature uncertainty is achieved through an uncertainty prediction network; the uncertainty prediction network is a multilayer perceptron and uses the Softplus activation function to ensure that the output feature uncertainty is always positive.

8. The method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network according to claim 1, characterized in that, The steps for decoding using the probabilistic localization decoder specifically include: The fused radar feature sequence and electronic reconnaissance feature sequence are aggregated to form a fixed-length fused feature vector. The fused feature vector is processed using a multilayer perceptron network with two independent output branches, which outputs the mean of the location coordinates and the covariance matrix, respectively.

9. The method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network according to claim 8, characterized in that, The covariance matrix is ​​output in the following manner: The output vector of one output branch of the multilayer perceptron network; Construct a lower triangular matrix based on the vector.

10. The method for locating heterogeneous signals from electronic reconnaissance and radar based on an encoder-decoder network according to claim 1, characterized in that, The encoder-decoder network model is trained and optimized using a negative log-likelihood loss function, which includes a Mahalanobis distance term and a regularization term.

Images