A direct positioning method based on a multi-scale feature fusion neural network

Abstract

The application discloses a direct positioning method based on a multi-scale feature fusion neural network. Firstly, time domain signals received by a multi-station uniform linear array are acquired, the time domain signals are preprocessed, and a covariance matrix dimension of the preprocessed time domain signals is reshaped, and linear regression labels corresponding to source coordinates are generated. Then, the reshaped covariance matrix data is respectively extracted by a multi-scale sampling convolutional neural network and a multi-scale dilated convolutional neural network to extract local features and global features of the data, and the global features and the local features are selectively fused by a multi-scale selection to obtain deep features of the whole signal. Finally, based on the extracted deep features, the coordinates of the signal source are estimated by a linear regression method. The method realizes information fusion between multi-station without losing the resolution of the covariance matrix, obtains multi-scale signal features for convolution operation, and ensures the utilization rate of global feature information and local feature information.

Description

Technical Field

[0001] This invention relates to the field of direct localization technology in deep learning, specifically a direct localization method based on a multi-scale feature fusion neural network. Background Technology

[0002] Passive positioning technology, due to its characteristics of not emitting electromagnetic signals and strong concealment, plays a crucial role in fields such as search and rescue, autonomous driving, smart logistics, and IoT sensing. Traditional passive positioning methods are mainly divided into two-step positioning methods and one-step positioning methods. Among them, the one-step positioning method, also known as the direct position determination (DPD) method, constructs a cost function by directly processing the original sampled signals, avoiding errors in intermediate parameter estimation, and can achieve higher positioning accuracy under ideal conditions.

[0003] However, traditional DPD methods face severe challenges in practical applications. Their performance is highly dependent on an ideal signal model; when real-world environments are complex, such as those with low signal-to-noise ratios, few base stations, or limited snapshot data, their positioning performance deteriorates sharply, exhibiting poor robustness. Furthermore, the optimization process for their high-dimensional cost function is computationally complex, making it difficult to meet the demands of real-time processing.

[0004] In recent years, data-driven deep learning methods have provided a new paradigm for direct localization. By constructing the localization problem as a coordinate classification or regression problem, deep neural networks can directly learn the complex mapping relationship from data to location from array signals. However, existing deep learning-based direct localization methods still have significant limitations:

[0005] Limited Feature Extraction Capabilities: Mainstream Convolutional Neural Networks (CNNs) are limited by their fixed local receptive fields, making it difficult to effectively capture the global spatial dependencies between signal sources and base stations contained in the array covariance matrix. This makes it difficult for the model to extract discriminative deep features in complex electromagnetic environments, limiting its upper limit of positioning accuracy.

[0006] The trade-off between computational efficiency and global modeling: While the attention mechanism can capture global contextual information to overcome the limitations of the receptive field in CNNs, its computational complexity increases quadratically with the size of the input features. This leads to enormous computational overhead when dealing with high-dimensional array covariance matrices, hindering the deployment and real-time application of models on resource-constrained platforms.

[0007] Insufficient multi-site information fusion: Most existing deep learning methods process data from individual base stations independently or perform simple concatenation and fusion, failing to effectively utilize spatial diversity gain and signal correlation information among multiple base stations. This simplistic fusion strategy limits the model's localization potential in sparse observation scenarios with few base stations.

[0008] Therefore, designing a direct localization method that can efficiently capture global dependencies in the array covariance matrix, take into account local fine features, effectively integrate multi-site information, and maintain high accuracy, high robustness, and low computational complexity in complex environments has become a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0009] The purpose of this invention is to overcome the problems of local defects in feature extraction of the array covariance matrix due to the small receptive field of the traditional convolution kernel and the low direct positioning accuracy of traditional methods when there are few base stations. This invention proposes a direct positioning method based on a multi-scale feature fusion neural network. This method integrates the feature information of the covariance matrix through matrix operations without losing the resolution of the covariance matrix. By performing special processing on the covariance matrix, information fusion between multiple stations is achieved, and multi-scale signal features are obtained for convolution operation, ensuring the utilization rate of global and local feature information.

[0010] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows:

[0011] This invention provides a direct localization method based on a multi-scale feature fusion neural network, comprising the following steps:

[0012] S1. Acquire the time-domain signal received by a multi-site uniform linear array, preprocess the time-domain signal to generate linear regression labels corresponding to the source coordinates, and reshape the covariance matrix dimension of the time-domain signal; the preprocessing specifically involves obtaining the ideal covariance matrix from the signals received by each base station based on the number of base stations in the multi-site array, extracting the real part, imaginary part, and phase of the ideal covariance matrix respectively, generating a three-dimensional tensor of a preset dimension, and then concatenating the real part, imaginary part, and phase of the covariance matrix of each site in sequence to form a three-dimensional tensor adapted to the dimension of the network input data;

[0013] S2. For the reshaped covariance matrix data, local and global features of the data are extracted by multi-scale sampling convolutional neural network and multi-scale dilated convolutional neural network, respectively. Then, the global and local features are selectively fused by multi-scale to obtain the deep features of the entire signal.

[0014] S3. Based on the extracted deep features, the coordinates of the signal source are estimated by linear regression; wherein, steps S2 and S3 constitute a direct localization model.

[0015] Preferably, step S2, which involves extracting local features using a multi-scale sampling convolutional neural network, includes two steps: feature tensor preprocessing and deep feature extraction. Feature tensor preprocessing involves obtaining an input feature tensor containing batch size, number of channels, and the height and width dimensions of the feature map, and dividing it into three branches based on the number of channels. Deep feature extraction involves performing downsampling, depth convolution, and upsampling operations sequentially on each branch, which consists of two convolutional layers, an upsampling module, and a downsampling module, to restore the original data dimensions. Each branch is concatenated along the number of channels, processed by a convolutional kernel of a preset size, and then multiplied element-wise with the original input feature tensor to obtain local features. The downsampling uses adaptive max pooling with downsampling factors of 4, 8, and 16; the depth convolutional kernel size is 3×3 with a stride of 1 and padding of 1; each group has 22 channels; and the convolution process uses the ReLU activation function.

[0016] Preferably, the process of extracting global features using a multi-scale dilated convolutional neural network in step S2 is as follows: after processing the covariance matrix, the height and width of the feature tensor are transformed to a preset size. Based on the formula for calculating the receptive field of dilated convolution, dilated convolutional layers with dilation rates of 2, 4, and 8 are set, paired with convolutional kernels of a preset size. Simultaneously, two-dimensional batch normalization and the ReLU activation function are used during the dilation convolution process to obtain the global features. The size of the dilated convolutional kernel is 3×3, and the multi-scale dilated convolutional neural network has 3 dilated convolutional branches, with 3 input feature channels and 3 output feature channels.

[0017] Preferably, step S2 employs a multi-scale selective feature fusion method for the selective fusion of global and local features, including: 1. Global information aggregation and channel modeling: global average pooling is performed on all branch feature maps to be fused, and then the dependencies between channels are modeled using convolutional layers with shared weights to obtain unnormalized channel descriptors; 2. Multi-scale weight generation: all the channel descriptors are concatenated in the spatial dimension to form an integrated weight tensor, and the ReLU activation function and Softmax function are applied sequentially to generate normalized weights that satisfy the sum of the weights of each branch under each channel being 1; 3. Weighted feature reconstruction and projection: the normalized weights are used to perform element-wise multiplication and summation on each original feature branch to achieve adaptive feature fusion and obtain fused features; the fused features are residually connected with the original input features, and then feature mapping is performed through a 1×1 convolutional layer to obtain the deep features of the entire signal.

[0018] Preferably, the specific process of estimating the signal source coordinates through linear regression in step S3 is as follows: the final feature tensor is expanded into a one-dimensional vector, and then processed sequentially through multiple linear layers and activation functions. The last linear layer directly outputs a vector of coordinate estimates, thus obtaining the coordinate values ​​of each signal source. In the linear regression prediction module, the number of input feature channels is 3072, and the number of output categories is 6; the number of linear layers is set sequentially to 2048, 1024, 512, 256, and 6, and the linear layers use the ReLU activation function.

[0019] Preferably, the method of the present invention further includes a training step for the direct positioning model, employing a direct training strategy for end-to-end training, specifically including: 1. Construction of the training dataset: Based on a uniform linear array antenna structure, the signal-to-noise ratio parameter is discretized within a preset range with a fixed step size. The parameter ranges for the base station location and source coordinates are set, and the parameter ranges are divided into a preset number of grids. A fixed number of point coordinates are set within each grid, and each grid contains only one source. All possible combinations of multi-source grids without distinction of order or repetition are exhaustively enumerated to simulate different multi-source spatial distribution scenarios. For each scenario defined by a specific signal-to-noise ratio and multi-source coordinates, the corresponding ideal covariance matrix is ​​calculated, generating independent training samples, which are then integrated to obtain a multi-source ideal covariance matrix dataset; 2. Direct positioning model training: All parameters of the direct positioning model are trained iteratively on the multi-source ideal covariance matrix dataset. During training, the AdamW optimizer is used, an adaptive learning rate scheduler is set, and an early stopping strategy is configured to complete model training.

[0020] The direct localization model uses the mean squared error (MSE) as the loss function during training and the linear regression labels of the corresponding source coordinates as the true labels.

[0021] As a preferred option, the specific parameters for model training are configured as follows: 100 training epochs, batch size 256, validation set ratio 0.1; and an initial learning rate of 1x10 for the AdamW optimizer. -4 The weight decays to 1x10. -3 The number of early stops is 15 rounds.

[0022] This invention has the following characteristics and beneficial effects:

[0023] First, it combines high accuracy with strong robustness. This invention introduces a parallel structure of multi-scale dilated convolution and multi-scale sampling convolution, expanding the receptive field to capture global spatial dependencies while preserving the fine local structure of the covariance matrix. This enables the model to stably extract effective signal features even in complex and harsh environments such as low signal-to-noise ratio, limited snapshots, and few base stations, overcoming the poor robustness of traditional DPD methods and the limitations of feature extraction in ordinary CNNs, significantly improving the accuracy of direct multi-target localization.

[0024] Second, it balances global modeling capabilities with low computational complexity. This invention uses multi-scale dilated convolution instead of the traditional attention mechanism to extract global features, avoiding the problem of computational complexity increasing with the square of the size when dealing with high-dimensional covariance matrices. While ensuring the capture of long-range dependencies, it significantly reduces the model's computational overhead (FLOPs are significantly lower than similar models), making it easier to deploy on resource-constrained edge computing devices and meeting the application requirements for real-time processing.

[0025] Third, it achieves deep fusion of information from multiple stations. This invention organically integrates information from multiple stations into a unified three-dimensional tensor input network by performing structured preprocessing (separation and concatenation of real, imaginary, and phase components) on the covariance matrix of multiple base stations, rather than simply concatenating or processing them independently. This fusion strategy enables the neural network to fully explore and utilize the spatial diversity gain and signal correlation between different base stations, effectively improving the positioning capability and generalization performance in sparse observation scenarios such as those with few base stations.

[0026] Fourth, it boasts high feature utilization and fast convergence speed. The multi-scale selective feature fusion module designed in this invention assigns weights to feature maps of different scales through adaptive learning, dynamically enhancing useful information and suppressing redundant interference, and then performs residual connections between the fused features and the original input. This mechanism not only preserves the original information of the signal to the maximum extent but also promotes gradient flow during training, thereby accelerating model convergence and improving feature utilization efficiency.

[0027] In summary, this invention provides a novel direct positioning solution that is highly accurate, robust, low-complexity, and easy to deploy. It effectively solves the core problems in existing technologies, such as insufficient feature extraction, low computational efficiency, and difficulty in multi-site information fusion. It has broad application prospects in civilian fields such as autonomous driving, intelligent logistics, and search and rescue. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the uniform linear array structure involved in this invention.

[0029] Figure 2 This is a schematic diagram of the four base stations and three signal sources involved in this invention.

[0030] Figure 3 This is a schematic diagram of the deep learning model structure involved in this invention.

[0031] Figure 4 This is a performance comparison chart of the method of this invention with other algorithms under different snapshot numbers.

[0032] Figure 5 This is a performance comparison chart of the method of this invention with other algorithms under different signal-to-noise ratio conditions. Detailed Implementation

[0033] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0034] Example 1

[0035] This embodiment provides a direct localization method based on a multi-scale feature fusion neural network. The specific method of this embodiment is as follows:

[0036] S1. Obtain the time-domain signal received by the multi-site uniform linear array, preprocess the time-domain signal to generate linear regression labels corresponding to the source coordinates, and reshape the covariance matrix dimension of the time-domain signal.

[0037] Specifically, in this embodiment, a uniform linear array structure is used to complete the time-domain signal reception, such as... Figure 1 As shown, the uniform linear array antenna structure consists of It consists of several array elements, with the first sensor as the reference element, and the spacing between the array elements is... The wavelength is .

[0038] Assumption The sources are respectively Narrowband incoherent signals are incident on a signal with On a uniform linear array of array elements, then the th array element... The sensor at the first The complex envelope received in the next snapshot is:

[0039] (1)

[0040] in For the first The received signals from each sensor Represents the first One source, For the number of snapshots, , For the first The sensor receives the first signal relative to the reference element. Phase delay of the signal propagation.

[0041] The received time-domain signal is represented as:

[0042] (2)

[0043] For the received signal vector, Let be the direction matrix, where It is a direction vector. It is a signal vector. It is a noise vector. , This refers to the number of base stations.

[0044] Furthermore, the time-domain signal is preprocessed, specifically based on the number of base stations. For example, In this embodiment, by Each base station receives signals Obtaining the ideal covariance matrix The real, imaginary, and phase parts of the covariance matrix are extracted, resulting in a final generation of dimension [missing information]. The three-dimensional tensor, and then... The real part, imaginary part, and phase of the covariance matrix of each station are concatenated in sequence to form... The three-dimensional tensor adapts the data dimension of the network input, which is the covariance matrix after dimension reshaping.

[0045] At the same time, The specific coordinates of each information source , Concatenate them sequentially into one dimension. The one-dimensional vector is the linear regression label of the source coordinates.

[0046] S2. For the reshaped covariance matrix data, local and global features of the data are extracted by multi-scale sampling convolutional neural network and multi-scale dilated convolutional neural network, respectively. Then, the global and local features are selectively fused by multi-scale to obtain the deep features of the entire signal.

[0047] Specifically, the method for extracting local features using a multi-scale sampling convolutional neural network is as follows:

[0048] Feature tensor preprocessing:

[0049] Obtain the input feature tensor Its dimensions are ,in For batch size, For the number of channels, , The height and width of the feature map are defined; the input features are divided into three branches according to the number of channels.

[0050] Deep feature extraction:

[0051] Each branch consists of two convolutional layers, an upsampling module, and a downsampling module, with convolutional kernel sizes of [sizes to be filled in]. Depth convolution and The standard convolution operation retains the same number of channels before and after convolution. Each branch is downsampled, then subjected to a depthwise convolution followed by upsampling to restore the original data dimensions. Finally, each branch is concatenated along its channel count. The convolution kernel is then multiplied element-wise with the original input features. This is because multi-scale sampling in space leads to the loss of spatial information, while multiplying with the input features preserves some spatial details. The multi-scale sampling convolution module extracts local spatial features and global channel features from the processed covariance matrix data, progressively transforming local features in the real, imaginary, and phase parts into highly refined and most effective features, achieving an end-to-end mapping from the array covariance matrix to deep spatial features. Its mathematical expression is as follows:

[0052] (3)

[0053] In the formula, For input data, For max pooling layer, For upsampling, The kernel size is Deep convolutional layers, These are the high-level features after encoding.

[0054] Furthermore, global features are extracted using a multi-scale dilated convolutional neural network, as follows:

[0055] Set the dilated convolution parameters:

[0056] After processing the covariance matrix, the characteristic tensor Height and width from become According to the formula for calculating the receptive field of dilated convolution, dilated convolutional layers include those with dilation rates of 2, 4, and 8. The convolutional layers, while using two-dimensional batch normalization and Activation function.

[0057] (4)

[0058] In the formula, For input data, For two-dimensional batch normalization layer, The kernel size is dilated convolutional layers, These are the high-level features after encoding.

[0059] Finally, the process of selectively fusing global and local features at multiple scales, which is the basis for the multi-scale selective feature fusion method, is as follows:

[0060] Global information aggregation and channel modeling:

[0061] To capture the global contextual information of each feature branch, global average pooling is first performed on the feature maps of all branches to be fused. Subsequently, using a shared weight Convolutional layers are used to model the dependencies between channels. Let... Given the four feature branches of the input, this step can be represented as:

[0062] (5)

[0063] in, It is an unnormalized channel descriptor.

[0064] Multi-scale weight generation:

[0065] To obtain the relative importance across different scales, the model concatenates the five channel descriptors calculated above in spatial dimensions to form an integrated weight tensor. Next, apply them in sequence. Activation function and The function is used to generate the final normalized weights. Used for initially mapping numerical values ​​to Interval, activate channel features. Normalization is performed along the scale dimension (i.e., the dimension of the stitching) to ensure that for each channel... The sum of the weights of the four branches is 1. This allows the network to make competitive choices between different scales.

[0066] (6)

[0067] Weighted feature reconstruction and projection:

[0068] Using the calculated normalized weights, element-wise multiplication and summation are performed on the original four feature branches to achieve adaptive feature fusion. This step allows the network to dynamically enhance useful scale features and suppress irrelevant features based on the specific content of the input image. The fused features... The calculation is as follows:

[0069] (7)

[0070] Finally, to facilitate gradient flow and further integrate features, the fused features are... With the original input Perform residual join, and through a The convolutional layer performs the final feature mapping:

[0071] (8)

[0072] S3. Based on the extracted deep features, the coordinates of the signal source are estimated by linear regression.

[0073] Specifically, firstly, the final feature tensor is obtained... Expanded into a one-dimensional vector, then passed through multiple linear layers and activation functions The last linear layer directly outputs a vector of coordinate estimates. Let the target have... One, whose mathematical expression is:

[0074] (9)

[0075] In the formula, These are the coordinates of the information source. , The last linear layer function This is a pre-neural network.

[0076] Understandably, the direct positioning model constructed in this embodiment is used to implement steps S2 and S3.

[0077] In this embodiment, the direct localization model employs a direct training strategy. The model will be trained end-to-end on a sampled dataset containing real-world scene features. Through a single-stage learning process, the model directly adapts to the target scene to achieve the expected prediction accuracy. The specific steps are as follows:

[0078] Training phase: Based on the ideal covariance matrix dataset of three sources.

[0079] The training dataset was constructed based on a uniform linear array antenna structure, with the signal-to-noise ratio parameter starting from... arrive by Discretize the step size; simultaneously, the positions of the four base stations are respectively... Coordinates of each signal source and The parameter range is between 221 and 281. In this embodiment, 221 to 281 is divided into 12 grids, and the coordinates... and The entire two-dimensional plane consists of 144 grids, with 25 point coordinates per grid. Each signal source represents a point within a grid, and each grid can only contain one signal source. All possible, non-repeating, three-pair grid combinations are enumerated to simulate different spatial distribution scenarios of three signal sources. For each scenario defined by a specific signal-to-noise ratio and three signal coordinate pairs, an independent training sample is generated by calculating its corresponding ideal covariance matrix. By iterating through all parameter combinations, the three-source ideal covariance matrix dataset is integrated.

[0080] During training, all model parameters were iterated on the aforementioned three-source ideal covariance matrix dataset. Given the low signal-to-noise ratio of this data, the model can quickly identify effective features in the covariance matrix under low signal-to-noise ratio conditions and construct a robust input-to-output mapping logic, thereby achieving the expected training effect.

[0081] When training the direct localization model, the mean squared error of MSE is used as the loss function, and the linear regression label of the corresponding source coordinates is used as the true label.

[0082] Training environment setup and parameter initialization.

[0083] Specific parameters of the network architecture:

[0084]

[0085] Table 1 shows the detailed parameter configuration of the multi-scale sampling convolution module in the neural network structure of this method.

[0086]

[0087] Table 2 shows the specific parameter configurations of the multi-scale dilated convolution module in the neural network structure of this method.

[0088]

[0089] Table 3 shows the specific parameter configurations for the regression prediction module in the neural network structure of this method.

[0090]

[0091] Table 4 shows the specific parameter configurations for the formal training of this method.

[0092]

[0093] Table 5 shows the specific experimental environment configuration.

[0094] Comparative Example

[0095] The model methods compared with the method of this invention are respectively based on Direct localization estimation and deep convolutional neural networks ( Table 6 provides the specific model index parameters for the model of this invention and the comparative model.

[0096] Model Parameter Quantity and Complexity Analysis

[0097] The formula for calculating the total number of parameters in the global dynamic convolutional neural network is as follows:

[0098] (5)

[0099] In the formula, For multi-scale sampling convolution parameters, The parameters for multi-scale dilated convolution. For the number of parameters in multi-scale feature fusion, This refers to the number of parameters in the linear regression layer. Calculations show that the total number of parameters in the model of this invention is 9.05M, which is lower than... Model.

[0100] The computational complexity of the model is the number of floating-point operations during the forward propagation process. The overall model complexity is measured as follows:

[0101] (6)

[0102] In the formula, The number of floating-point operations for multi-scale sampling convolution. The number of floating-point operations for multi-scale dilated convolution. The number of floating-point operations for multi-scale feature fusion This represents the number of floating-point operations performed on the linear regression layer. Calculations show that the model of this invention performs... The total FLOPs under the input dimension is 280.06M.

[0103]

[0104] Table 6 shows the comparison data of the model indicators of the present invention and the comparative examples.

[0105] To verify the effectiveness of the above method, multiple simulation experiments were conducted in this embodiment, and the experimental performance was analyzed, as follows:

[0106] Experimental performance evaluation indicators:

[0107] Signal-to-noise ratio (SNR) is defined as:

[0108] (7)

[0109] in, For signal power, This represents noise power.

[0110] The performance estimation criterion is the Joint Root Mean Square Error (RMSE), defined as follows:

[0111] (8)

[0112] in, For the first The Monte Carlo Trial Precise estimates of the coordinates of each source. Indicates the number of information sources. This indicates the number of Monte Carlo trials.

[0113] Experimental results:

[0114] Figure 1 A schematic diagram of the uniform linear array structure involved in this invention is shown.

[0115] Figure 2 A schematic diagram of the four base stations and three signal sources involved in this invention is shown.

[0116] Figure 3 A schematic diagram of the deep learning model structure involved in this invention is shown.

[0117] Figure 4 In the number of snapshots Signal-to-noise ratio The coordinates of the three sources are as follows: The locations of the four base stations are as follows: Monte Carlo trials The models being compared in the context of the scenario Figure. From Figure 3 As can be seen from this, all models The performance of the method of the present invention gradually decreases as the number of snapshots increases. Except for the small snapshot count of 50, the performance of the method of the present invention is consistently better than that of the method of the present invention at other snapshot counts. and At the same time, it can maintain good estimation accuracy even with a small number of snapshots.

[0118] Figure 5 In signal-to-noise ratio Quick shot number The coordinates of the three sources are as follows: With the base station location unchanged, the number of Monte Carlo trials... The models being compared in the context of the scenario Figure. From Figure 4 As can be seen from this, all models It gradually decreases as the signal-to-noise ratio increases, but compared to the model and The performance of these methods is inferior to that of the present invention, and due to the limitation of low signal-to-noise ratio, The inability to accurately estimate the positions of the three signal sources demonstrates that the method of this invention can also accurately locate them directly even under low signal-to-noise ratio conditions.

[0119] Example 2

[0120] This embodiment provides a direct localization system based on a multi-scale feature fusion neural network, used to implement the direct localization method based on a multi-scale feature fusion neural network disclosed in Embodiment 1. The system includes: a signal preprocessing module, used to acquire time-domain signals received by a multi-site uniform linear array, and complete linear regression label generation and covariance matrix dimension reshaping; a multi-scale feature extraction module, including a multi-scale sampling convolution sub-module, a multi-scale dilated convolution sub-module, and a feature fusion sub-module, used to extract local features and global features of the covariance matrix data, respectively, and complete selective fusion of multi-scale features to obtain deep signal features; a localization model construction and estimation module, used to construct a direct localization model, extract deep features, and estimate the coordinates of the signal source through linear regression; and a model training module, used to perform end-to-end training of the direct localization model using a direct training strategy to generate a trained direct localization model.

[0121] Example 3

[0122] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the direct localization method based on a multi-scale feature fusion neural network disclosed in Embodiment 1.

[0123] Example 4

[0124] This embodiment provides an electronic device, including a processor and a memory. The memory stores a computer program. When the processor executes the computer program, it implements the steps of the direct localization method based on a multi-scale feature fusion neural network disclosed in Embodiment 1. The hardware environment of the electronic device is configured as follows: GPU is NVIDIA GeForce RTX 4090, CPU is AMD EPYC 7K62 48-Core 2.6GHz Base; software environment is configured as follows: deep learning framework is PyTorch 2.6, CUDA version is 12.8, and Python version is 3.12.

[0125] The foregoing has shown and described 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 embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A direct localization method based on a multi-scale feature fusion neural network, characterized in that, Includes the following steps: S1. Obtain the time-domain signal received by the multi-site uniform linear array, preprocess the time-domain signal, reshape the covariance matrix dimension of the preprocessed time-domain signal, and generate linear regression labels for the corresponding source coordinates. S2. For the reshaped covariance matrix data, local and global features of the data are extracted by multi-scale sampling convolutional neural network and multi-scale dilated convolutional neural network, respectively. Then, the global and local features are selectively fused by multi-scale to obtain the deep features of the entire signal. Step S2, which involves the selective fusion of global and local features at multiple scales, is a multi-scale selective feature fusion method, comprising the following steps: Global information aggregation and channel modeling: Global average pooling is performed on all branch feature maps to be fused, and then the dependencies between channels are modeled using convolutional layers with shared weights to obtain unnormalized channel descriptors. Multi-scale weight generation: All the channel descriptors are spliced ​​together in the spatial dimension to form an integrated weight tensor. The ReLU activation function and the Softmax function are applied in sequence to generate normalized weights. The normalized weights satisfy that the sum of the weights of each branch under each channel is 1. Weighted feature reconstruction and projection: The normalized weights are used to perform element-wise multiplication and summation on each original feature branch to achieve adaptive feature fusion and obtain fused features; the fused features are residually connected with the original input features, and then feature mapping is performed through a 1×1 convolutional layer to obtain the deep features of the entire signal; S3. Based on the extracted deep features, the coordinates of the signal source are estimated by linear regression.

2. The direct localization method based on a multi-scale feature fusion neural network according to claim 1, characterized in that, The specific preprocessing process in step S1 is as follows: using the number of base stations in multiple sites as the base, the ideal covariance matrix is ​​obtained from the signals received by each base station. The real part, imaginary part, and phase of the ideal covariance matrix are extracted respectively to generate a three-dimensional tensor of a preset dimension. Then, the real part, imaginary part, and phase of the covariance matrix of each site are concatenated in sequence to form a three-dimensional tensor that adapts to the dimension of the network input data.

3. The direct localization method based on a multi-scale feature fusion neural network according to claim 1, characterized in that, The process of extracting local features using a multi-scale sampling convolutional neural network in step S2 includes the following steps: Feature tensor preprocessing: Obtain the input feature tensor, the dimensions of which include batch size, number of channels, feature map height and width; divide the input feature tensor into three branches according to the number of channels; Deep feature extraction: Each branch consists of two convolutional layers, an upsampling module, and a downsampling module. The convolutional kernels include a depth convolutional kernel of a preset size and a regular convolutional kernel. The number of channels remains unchanged before and after convolution. Downsampling, depth convolution, and upsampling operations are performed on each branch in sequence to restore the original data dimension. After concatenating the branches along the number of channels, the data is processed by a convolutional kernel of a preset size and then multiplied element-wise with the original input feature tensor to obtain local features.

4. The direct localization method based on a multi-scale feature fusion neural network according to claim 3, characterized in that, The downsampling uses adaptive max pooling with downsampling factors of 4, 8, and 16; the depthwise convolution kernel has a size of 3×3, a stride of 1, and padding of 1; each group has 22 channels; and the ReLU activation function is used during convolution.

5. The direct localization method based on a multi-scale feature fusion neural network according to claim 1, characterized in that, The process of extracting global features through a multi-scale dilated convolutional neural network in step S2 includes the following steps: after processing the covariance matrix, the height and width of the feature tensor are transformed to the preset size. According to the formula for calculating the receptive field of dilated convolution, the dilation rates of the dilated convolutional layers are set to 2, 4 and 8, respectively, and the convolutional kernel is set to the preset size. At the same time, two-dimensional batch normalization and ReLU activation function are used in the dilated convolution process to obtain global features after processing.

6. The direct localization method based on a multi-scale feature fusion neural network according to claim 5, characterized in that, The size of the dilated convolution kernel is 3×3. The multi-scale dilated convolutional neural network has 3 dilated convolution branches, and the number of input feature channels and the number of output feature channels are both 3.

7. The direct localization method based on a multi-scale feature fusion neural network according to claim 1, characterized in that, In step S3, the coordinates of the signal sources are estimated by linear regression. The specific process is as follows: the final feature tensor is expanded into a one-dimensional vector, and then processed by multiple linear layers and activation functions in sequence. The last linear layer directly outputs the vector of coordinate estimates to obtain the coordinate values ​​of each signal source.

8. The direct localization method based on a multi-scale feature fusion neural network according to claim 7, characterized in that, In the regression prediction module used for the linear regression, the number of input feature channels is 3072, and the number of output categories is 6; the number of linear layers is set to 2048, 1024, 512, 256, and 6 respectively, and the linear layers use the ReLU activation function.

9. The direct localization method based on a multi-scale feature fusion neural network according to any one of claims 1-8, characterized in that, Steps S2 and S3 constitute the direct positioning model.

10. The direct localization method based on a multi-scale feature fusion neural network according to claim 9, characterized in that, The training steps of the direct localization model employ a direct training strategy for end-to-end training, specifically including: Construction of the training dataset: Based on a uniform linear array antenna structure, the signal-to-noise ratio (SNR) parameter is discretized within a preset range with a fixed step size. The parameter ranges for the base station location and source coordinates are set, and the parameter ranges are divided into a preset number of grids. A fixed number of point coordinates are set in each grid, and each grid contains only one source. All non-repeating multi-source grid combinations are traversed to simulate different multi-source spatial distribution scenarios. For each scenario defined by a specific SNR and multi-source coordinates, the corresponding ideal covariance matrix is ​​calculated, independent training samples are generated, and the results are integrated to obtain a multi-source ideal covariance matrix dataset. Direct localization model training: All parameters of the direct localization model were trained iteratively on the multi-source ideal covariance matrix dataset. During training, the AdamW optimizer was used, an adaptive learning rate scheduler was set, and an early stopping strategy was configured to complete model training. The direct localization model uses the mean squared error (MSE) as the loss function during training and the linear regression labels of the corresponding source coordinates as the true labels.

11. The direct localization method based on a multi-scale feature fusion neural network according to claim 10, characterized in that, The model training parameters were configured as follows: 100 training epochs, batch size 256, validation set ratio 0.1; the initial learning rate of the AdamW optimizer was... The weight decays to The earliest stopping round is 15 rounds.

Images