Radiation source target positioning method and device based on deep learning in urban environment

Abstract

The invention provides a radiation source target positioning method based on deep learning in an urban environment. The radiation source target positioning method comprises the following steps: constructing a simulated geography-electromagnetic data set and an actually acquired geography-electromagnetic data set; on the basis of an encoder-decoder network structure, constructing a neural network model used for radiation source target positioning; constructing a loss function; carrying out the training of a minimization loss function on the neural network model through employing the simulation geographic-electromagnetic data set, obtaining a trained positioning model, and completing the testing and verification of the trained positioning model through employing the actual collection geographic-electromagnetic data set; and inputting geographical-electromagnetic data acquired by the unmanned aerial vehicle into the verified positioning model, and outputting a radiation source target positioning result. According to the method provided by the invention, under the condition that high-cost acquisition is limited, the generalization ability and the actual deployability of the model are guaranteed; and the positioning precision in an urban environment with complex shielding and remarkable multipath reflection is improved.

Description

Technical Field

[0001] This invention relates to the field of passive positioning, and in particular to a method and device for locating radiation source targets based on deep learning in an urban environment. Background Technology

[0002] In recent years, the widespread use of various electronic devices in land, sea, air, and space has made the electromagnetic environment increasingly complex. In this environment, efficiently and reliably acquiring the location information of radiation sources has become crucial. Radiation source localization technology has wide applications, especially in fields such as radio signal processing, target detection, and wireless communication. Based on whether the sensors in the localization system emit electromagnetic signals, radiation source localization can be divided into active localization and passive localization. Active localization, also known as proactive localization, requires localization sensors (e.g., radar, sonar) to actively emit electromagnetic signals to determine the target's location. This makes the transmitter's location easily exposed, thus becoming a key target for enemy electronic countermeasures or firepower. Furthermore, active localization systems are expensive to manufacture and consume a lot of energy, resulting in high operating costs. Passive localization, on the other hand, achieves localization by passively receiving and processing the electromagnetic signals radiated by the target radiation source itself through localization sensors, offering greater concealment and security. Therefore, passive localization has very wide applications in both military and civilian fields.

[0003] However, in urban environments, due to the widespread obstruction of buildings, radiation sources are often not within the line-of-sight range of a drone swarm, making it impossible for vision-based methods to directly achieve positioning. Furthermore, the propagation of wireless signals is affected to varying degrees at different scales, making it difficult to achieve accurate positioning by considering only a single scale factor. Summary of the Invention

[0004] To address the problems existing in the prior art, a method and device for locating radiation source targets in urban environments based on deep learning are provided. Based on real-time geographic-electromagnetic (spatial-spectrum) combined data collected by an electromagnetic spectrum payload carried by an unmanned aerial vehicle, the method achieves accurate passive location of radiation source targets by performing real-time modeling and inference on known urban environmental information.

[0005] The first aspect of this invention proposes a deep learning-based method for locating radiation source targets in urban environments, comprising: Construct simulated geo-electromagnetic datasets and actual geo-electromagnetic datasets; Based on the encoder-decoder network structure, a neural network model for radiation source target localization is constructed. The encoder is designed with dilated convolutional modules with different dilation rates to extract features at different scales. The decoder is designed with a dual-head task output layer consisting of a path loss prediction head and a localization head to output radiation source location information. Construct the loss function; The neural network model was trained by minimizing the loss function using a simulated geographic-electromagnetic dataset to obtain a trained positioning model. The trained positioning model was then tested and validated using a real-world geographic-electromagnetic dataset. The geographic-electromagnetic data collected by the UAV is input into the validated positioning model, and the positioning results of the radiation source target are output.

[0006] As a preferred embodiment, the method for constructing the simulated geo-electromagnetic dataset includes: Import publicly available urban geographic information into high-precision electromagnetic simulation software; The electromagnetic properties of city buildings are set, as are the antenna parameters and location of the radiation source and the receiving antenna parameters. The signal propagation of the radiation source is simulated at a fixed height, and spatial and gray-scale quantization is performed to obtain a rasterized image representing the global path loss map of the radiation source. The geometric shape of the city buildings represented by the rasterized gray-scale image from the overhead view is saved as urban geographic information. The corresponding spatial coordinates of the radiation source are saved, and the simulation process is repeated to obtain the simulated geo-electromagnetic dataset.

[0007] As a preferred embodiment, the method for constructing the actual geographic-electromagnetic dataset includes: Drones collect spatial-spectrum data and correlate them with urban geographic environment at a fixed altitude; The spatial data and intensity data of the spatial-spectral data are pre-quantized separately to obtain spatial spectral data represented by a rasterized image. The associated urban geographic environment is quantified to obtain a quantized rasterized image representation of the urban geographic environment; The images are merged to obtain a real-world geographic-electromagnetic dataset.

[0008] As a preferred embodiment, the encoder is designed as follows: Construct a simplified path loss model for urban environments; Based on the path loss model, dilated convolutional modules with different porosity are constructed. These modules include low-porosity, medium-porosity, and high-porosity modules. The low-porosity module describes the propagation characteristics of short-range free space through local feature extraction. The medium-porosity module reflects the signal shadowing effect caused by building shading by modeling the mid-scale space distribution of building walls. The high-porosity module is used to capture features from the building cluster scale to the regional scale.

[0009] As a preferred embodiment, the void ratio of the low void ratio module is set to S1, the void ratio of the medium void ratio module is set to S2, and the void ratio of the high void ratio module is set to S3, wherein S1 < S2 < S3.

[0010] As a preferred embodiment, the decoder includes a path loss prediction output layer. In the path loss prediction output layer, two upsampling operations are first performed through two deconvolution modules to gradually restore the high-level features compressed by the encoder to the input resolution, preserving the macroscopic propagation characteristics in the urban environment. Then, the number of channels is reduced to 1 through a convolution module. Finally, the path loss is constrained to the range of [-1,1] through an activation function for normalized loss gain prediction.

[0011] As a preferred embodiment, the decoder includes a positioning output layer. In the positioning output layer, firstly, the output results after two upsampling operations are input into two convolutional modules to extract positioning-sensitive features; then, the number of channels is compressed to 1 by a downsampling module, and a normalized probability distribution heatmap is generated by an activation function; next, normalized network coordinates are pre-generated and decoupled from the input resolution; finally, the centroid is calculated by a centroid calculation module using a weighted average of the heatmap and the pre-stored network coordinates, which is the positioning result of the radiation source target.

[0012] As a preferred embodiment, the loss function includes using mean squared error to quantify the path loss between the predicted path loss and the actual path loss, specifically:

[0013] in, H and W These represent the spatial dimensions of the gain tensor (height and width of the feature map), respectively. This represents the predicted path loss value output by the model. This represents the corresponding actual path loss value; The coefficient used to adjust the weight of the loss term.

[0014] As a preferred embodiment, the loss function further includes positioning error, specifically:

[0015] Where, in the formula, N Indicates the number of samples in the batch; M This indicates the total dimensions of the positioning output; This represents the model's predicted location value; This represents the corresponding actual location value; This represents the loss weighting coefficient for that item.

[0016] A second aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in the first aspect, corresponding to the deep learning-based radiation source target localization method in an urban environment.

[0017] Compared with existing technologies, the beneficial effects of adopting the above technical solution are as follows: (1) This invention proposes a training strategy that integrates simulation data and real-world data, which ensures the generalization ability and practical deployability of the model under the condition of high cost and limited data collection. (2) This invention abandons the traditional "two-step" positioning process and uses an end-to-end neural network for direct positioning, which simplifies the system process and improves computational efficiency and real-time positioning. (3) By rationally designing the network sensing domain, the present invention can fully cover urban spatial areas and improve the positioning accuracy in urban environments with complex obstruction and significant multipath reflection. (4) By introducing a visual spatial spectrum modeling method and combining it with the geometric distribution of urban buildings, we have realized the modeling of different scale information on the propagation law of electromagnetic signals, which helps to improve the robustness of positioning in complex scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart of the deep learning-based radiation source target localization method proposed in this invention for urban environments.

[0019] Figure 2 This is a schematic diagram of a multi-scale feature extraction mechanism and a dual-task output network structure according to an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the algorithm verification results based on the constructed dataset in an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram illustrating the algorithm verification results of an embodiment of the present invention, in which random noise perturbation was added to both the training and test sets.

[0021] Figure 5 This is a schematic diagram of an electronic device according to an embodiment of this application.

[0022] Figure 6 This is a schematic diagram of the structure of a computer system suitable for implementing the electronic devices of the present application embodiments. Detailed Implementation

[0023] The embodiments of this application are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar modules or modules having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application. Rather, the embodiments of this application include all variations, modifications, and equivalents falling within the spirit and scope of the appended claims.

[0024] To address the problems in existing technologies, this invention proposes a deep learning-based method for locating radiation source targets in urban environments. This method enables precise location of radiation source targets in urban environments using only the location of the UAV itself and the electromagnetic signal intensity information collected by its payload. Please refer to... Figure 1 The specific plan is as follows: S100, Construct simulated geographic-electromagnetic datasets and actual geographic-electromagnetic datasets.

[0025] In this embodiment, the global path loss image corresponding to any radiation source location in the urban geographic environment is first calculated. Then, during training, samples are taken from the path loss image to obtain simulated spatial-spectrum sensing data for radiation source location estimation. The dataset mainly includes two types: simulated geographic-electromagnetic dataset and actual geographic-electromagnetic dataset. The simulated geographic-electromagnetic dataset is mainly used for training the subsequent localization model, i.e., as the training set, while the actual geographic-electromagnetic dataset is mainly used for testing and validating the trained model, i.e., as the test set.

[0026] To construct a large-scale simulated geo-electromagnetic dataset, diverse urban geographic information was first collected from large-scale open-source world mapping software (such as OpenStreetMap). This software contains crowdsourced labeled shape and height information of urban buildings, which has a crucial impact on urban electromagnetic spectrum propagation. The simulation process is as follows: 1) Import urban geographic information into high-precision electromagnetic simulation software.

[0027] 2) Set the electromagnetic property parameters of urban buildings in the software.

[0028] 3) Set the antenna parameters and location of the radiation source, as well as the receiving antenna parameters, in the software, and perform simulation calculations on the signal propagation of the radiation source at a fixed height. Then, perform spatial and gray-scale quantization, and store the resulting rasterized image representing the global path loss map of the radiation source. .

[0029] 4) Save the geometric shape of the city buildings represented by the rasterized grayscale image from an overhead view. This serves as urban geographic information. Non-zero grayscale areas represent the shape and height of city buildings, while zero grayscale areas represent open areas without buildings. The spatial coordinates of the radiation source from this simulation are saved. .

[0030] In order to standardize the simulated geo-electromagnetic dataset, in this embodiment, an image covering a 256m×256m area is extracted from the real map and its spatial resolution is set to 256×256 for storage.

[0031] To construct a geo-electromagnetic dataset collected by a swarm of drones, we first assume that the drones collect spatial-spectral data at a fixed altitude. Each spatial-spectral data sample is denoted as: (1) in, In the urban geographical environment i The first collection j The spatial coordinates of each sample This indicates the signal strength received by the electromagnetic spectrum load. (Tag) This represents the true spatial coordinates of the radiation source corresponding to this set of space-spectrum data.

[0032] Assuming that samples were collected from the same radiation source location within the same urban geographical environment... L If there are 10 samples, then the corresponding spatial-spectral data can be represented as: (2) The corresponding urban geographical environment can be represented by raster images. express.

[0033] To unify image processing and In this embodiment, The spatial information and signal strength data are quantized and preprocessed separately to obtain spatial spectrum data in the form of a raster image. In the spatial quantization process, a simplified assumption is made that the urban area is a 256m × 256m square, the viewing angle is a fixed-height overhead view, and each unit distance is 1m × 1m, thus obtaining a spatial resolution of 256 × 256. In the grayscale quantization part, an 8-bit quantization method is used.

[0034] Similarly, urban geographic environment data is also quantified in the same way to obtain... For multiple signal strength values ​​falling on the same grid point, the average value is used for fusion. After the above preprocessing, a real geo-electromagnetic dataset with an image representation can be obtained: (3) In the formula, N This indicates the number of geographic-electromagnetic data pairs.

[0035] S200: Based on the encoder-decoder network structure, a neural network model for radiation source target localization is constructed.

[0036] Since signal propagation in urban environments is affected by different structural levels, it is necessary to extract features at different scales of the urban environment in order to accurately characterize the signal propagation characteristics and achieve precise location of the signal source. Therefore, this embodiment proposes an encoder structure based on multi-scale feature extraction. By designing dilated convolution modules with different dilation rates, the extraction of features at different scales is achieved.

[0037] In regular convolution, the kernel acts as a filter to extract features from the input image. Dilated convolution introduces the concept of "dilation rate," which is simply the spacing between pixels within the kernel. If the dilation rate is 1, it behaves like regular convolution; however, if the dilation rate is greater than 1, dilated convolution is achieved. This spacing helps expand the kernel's "field of view," meaning it can see more pixels and extract features at a larger scale. Compared to regular convolution, dilated convolution expands the receptive field by increasing the dilation rate without increasing the kernel size, thus preventing the model from becoming too large. Furthermore, the behavior of the kernel can be altered by adjusting the dilation rate, meaning the model can be customized to meet the needs of the task, making it more flexible.

[0038] Before building the encoder structure, a simplified path loss model needs to be built first; (4) In the formula, the free space path loss term This indicates signal attenuation caused by distance between the transmitter and receiver. d It is Euclidean distance. n It is a path loss index under specific conditions. It is the frequency scaling factor; building penetration loss term. This indicates the signal loss caused by the penetration of a building. s It refers to the thickness of the material. It is the attenuation coefficient. It is the angle correction factor; multipath reflection term. This indicates the effects of multiple reflection paths, where Indicates the first i The reflection coefficient of the path, This represents its distance-weighted contribution value.

[0039] Based on a simplified path loss model, this embodiment designs dilated convolutional modules with different dilation rates to extract features of the urban environment at different scales, as detailed below: The first category is the low-voidity module (voidity = 1). This model describes the near-range free-space propagation characteristics through local feature extraction. Based on the receptive field calculation formula: (5) In the formula, Indicates the receptive domain of the current layer. This indicates the sensory domain of the next higher level. It is the equivalent convolution kernel size after the holes are filled. k It is the original kernel size. d It is the void ratio. From the 1st floor to the 2nd floor i The cumulative product of layer strides. With a 3×3 convolution kernel, the effective receptive field spans approximately 3×3 pixel kernels. This configuration can capture street-level geometric details in urban environments, including the interaction of building facades and signal propagation patterns in narrow alleyways.

[0040] The second category is the porosity module (porosity = 2, 5). This module reflects the signal shadowing effect caused by building shading by modeling the mesoscale spatial distribution of building walls. According to the receptive field calculation formula (5), for a 3×3 convolution kernel, the effective receptive field spans approximately 7×7 to 13×13 pixels.

[0041] The third type is the high porosity module (porosity = 9). This module can capture features from the building cluster scale to the regional scale, such as signal propagation across blocks and the continuity of main roads. This configuration can extract wireless signal attenuation patterns that reflect the macro-layout characteristics of the city, including multipath reflection effects of building clusters and path loss variations caused by terrain. According to the receptive field calculation formula (5), the effective receptive field can be expanded to 25×25 pixels with a 3×3 convolution kernel. This configuration can extract radio signal attenuation patterns that reflect the macro-layout characteristics of the city, including multipath reflection effects of building clusters and path loss variations caused by terrain.

[0042] Based on the multi-scale features of the urban environment extracted by the network through shared coding, this embodiment designs a dual-head task output layer with a path loss prediction head and a localization head during the decoding stage.

[0043] Specifically, in the path loss prediction output layer, two upsampling operations are first performed using two deconvolution modules to gradually restore the high-level features compressed by the encoder to the input resolution (256×256), preserving the macroscopic propagation characteristics in the urban environment. Then, a 3×3 convolutional module reduces the number of channels to 1. The 3×3 kernel is chosen because its local connectivity forces spatial smoothness in the path loss prediction, conforming to the continuity of propagation characteristics between adjacent areas in the urban environment. Finally, the path loss is constrained to the range [-1,1] using the Tanh activation function for predicting the normalized loss gain.

[0044] In the localization output layer, to integrate the advantages of deep learning and the physics propagation model, this invention employs a two-stage localization strategy of probability heatmap and centroid calculation. First, the output results after two upsampling operations are input into two convolutional modules with 3×3 kernels to extract localization-sensitive features. Then, the number of channels is compressed to 1 through a joint convolution and heatmap module, and a normalized probability distribution heatmap is generated using a softmax activation function. Second, to improve the model's generalization ability, normalized network coordinates are pre-generated and decoupled from the input resolution. Finally, the centroid is calculated using a mass calculation module through a weighted average of the heatmap and the pre-stored network coordinates; this centroid represents the radiation source location information. The specific calculation formula is shown below: (6) In the formula, In the heat map, the first... i,j The probability value of each pixel. These represent the normalized grid coordinates.

[0045] S300, Construct the loss function.

[0046] To ensure model convergence and achieve higher accuracy, this embodiment uses mean squared error to quantify the path loss error and positioning error between the predicted path loss and the actual path loss, without loss of generality.

[0047] In the path loss error term, the error is calculated independently for each spatial location of the signal strength, while the spatial dimension of the feature map is used to calculate the mean squared error to ensure that the loss is comparable for inputs at different resolutions. Furthermore, through... Adjust the contribution of the gain task to the total loss to avoid a single task dominating the optimization direction. The specific loss function model is shown below: (7) In the formula, H and W These represent the spatial dimensions of the gain tensor (height and width of the feature map), respectively. This represents the predicted path loss value output by the model. This represents the corresponding actual path loss value; The coefficient used to adjust the weight of this loss term is here. Set to 1.

[0048] S2.4 Construct the loss function for the positioning error term; In the positioning error term, the penalty for large errors is strengthened by calculating the squared difference between the predicted and actual values. Meanwhile, to avoid batch size affecting the magnitude of the loss, the sample size is included in the loss function calculation. The specific loss function model is shown below: (8) In the formula, N Indicates the number of samples in the batch; M This indicates the total dimensions of the positioning output; This represents the model's predicted location value; This represents the corresponding actual location value; This represents the loss weighting coefficient for this item, which is set to 1 here.

[0049] Finally, the total loss function is expressed as: (9) S400. The neural network model is trained by minimizing the loss function using a simulated geographic-electromagnetic dataset to obtain a trained positioning model. The trained positioning model is then tested and validated using a real-world geographic-electromagnetic dataset.

[0050] The specific process includes: (1) Training parameter settings: During the training process, the parameters used in this embodiment are ( )and( The Adam optimizer. The initial learning rate is set to... The number of training epochs is reduced to 0.9 times the original size every 20 training epochs. The total number of training epochs is set to 200, and the batch size is 32. In one embodiment, the overall code framework and model implementation are built based on Python and PyTorch.

[0051] (2) Model training: After initializing the number of training iterations, learning rate, and optimizer, the urban geographic environment images are trained using a simulated geo-electromagnetic dataset. and the global electromagnetic signal distribution map Obtained by spatial random sampling The input is fed into the proposed autoencoder network, and the total loss function is calculated. The network is trained by minimizing the total loss function. When the maximum number of training iterations is reached, if the loss function has converged to a small value, training is stopped; otherwise, the number of training iterations and network parameter settings are modified until the loss function converges.

[0052] (3) Testing and verification: Input the actual geographic-electromagnetic dataset into the trained localization model for verification. If the verification is successful, save the structure of the proposed autoencoder and the network parameters as the final localization model; otherwise, retrain.

[0053] S500 inputs the geo-electromagnetic data collected by the UAV into the verified positioning model and outputs the positioning results of the radiation source target.

[0054] The localization model is obtained by training the constructed autoencoder network based on multi-scale feature extraction. The geo-electromagnetic data sampled by the UAV swarm is input into the localization model to generate the predicted coordinate information of the radiation source, thereby achieving the effect of accurate localization of the radiation source target.

[0055] This invention proposes a radiation source target localization method in urban environments. By constructing a geo-electromagnetic dataset combining simulation and real-world data collection, and utilizing an image-based representation incorporating urban geographic structure and spatial-spectral data, a code-decode neural network structure is employed to directly estimate the spatial location of radiation sources from input images, achieving passive localization in complex urban environments. Through perceptual domain optimization, multi-scale feature extraction, spatial upsampling and downsampling structure design, and a two-stage localization mechanism involving probabilistic heatmaps and centroid calculation, the implicit correlation between geographic information and electromagnetic features is effectively extracted, achieving efficient estimation of radiation source locations.

[0056] Please refer to Figure 3 The results of algorithm validation based on the constructed dataset are shown; the black area represents urban buildings, the red marker in the center indicates the actual location of the radiation source, and the blue marker in the center indicates the location predicted by the model based on the spatial sampling points indicated by the green marker. Experiments on the test set show that the proposed method achieves an average positioning accuracy of 2.8 pixels, which is equivalent to a positioning error of 2.8 meters within a 256m × 256m urban geographic area.

[0057] Please refer to Figure 4 The results show the algorithm verification results with random noise perturbations added to both the training and test sets to evaluate the robustness of the proposed algorithm. The meanings of the elements are as follows: Figure 2 The results are identical. Experiments on the test set show that, even with added noise, the proposed method achieves an average positioning accuracy of 4.6 pixels within a 256m × 256m urban geographic area, equivalent to a positioning error of 4.6 meters. Compared to the noise-free case, the positioning error only increases by 1.8 meters, demonstrating the strong robustness of the proposed positioning method.

[0058] Reference Figure 5 As shown, an electronic device 100 according to an embodiment of this application includes: a memory 101 and a processor 102, wherein the memory 101 stores a computer program that can be loaded by the processor 102 and executed to correspond to the deep learning-based radiation source target localization method in an urban environment as described in the first aspect.

[0059] Figure 6 A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown.

[0060] It should be noted that, Figure 6The computer system 200 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0061] like Figure 6 As shown, the computer system 200 includes a Central Processing Unit (CPU) 201, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 202 or programs loaded from Storage Section 208 into Random Access Memory (RAM) 203, such as performing the methods described in the above embodiments. The RAM 203 also stores various programs and data required for system operation. The CPU 201, ROM 202, and RAM 203 are interconnected via a bus 204. An Input / Output (I / O) interface 205 is also connected to the bus 204.

[0062] The following components are connected to I / O interface 205: an input section 206 including a keyboard, mouse, etc.; an output section 207 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 208 including a hard disk, etc.; and a communication section 209 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 209 performs communication processing via a network such as the Internet. A drive 210 is also connected to I / O interface 205 as needed. A removable storage medium 211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 210 as needed so that computer programs read from it can be installed into storage section 208 as needed.

[0063] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 209, and / or installed from removable storage medium 211. When the computer program is executed by central processing unit (CPU) 201, it performs various functions defined in the system of this application.

[0064] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0065] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0066] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0067] In another aspect, this application also provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the deep learning-based radiation source target localization method in an urban environment as described in the above embodiments.

[0068] In another aspect, this application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to implement the deep learning-based radiation source target localization method in an urban environment as described in the above embodiments.

[0069] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0070] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0071] For those skilled in the art, the specific meanings of the above terms in this invention can be understood according to the specific circumstances; the accompanying drawings in the embodiments are used to clearly and completely describe the technical solutions in the embodiments of this invention. Obviously, the described embodiments are some embodiments of this invention, but not all embodiments. Generally, the components of the embodiments of this invention described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0072] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A deep learning-based method for locating radiation source targets in urban environments, characterized in that, include: Construct simulated geo-electromagnetic datasets and actual geo-electromagnetic datasets; Based on the encoder-decoder network structure, a neural network model for radiation source target localization is constructed. The encoder is designed with dilated convolutional modules with different dilation rates to extract features at different scales. The decoder is designed with a dual-head task output layer consisting of a path loss prediction head and a localization head to output radiation source location information. Construct the loss function; The neural network model was trained by minimizing the loss function using a simulated geographic-electromagnetic dataset to obtain a trained positioning model. The trained positioning model was then tested and validated using a real-world geographic-electromagnetic dataset. The geographic-electromagnetic data collected by the UAV is input into the validated positioning model, and the positioning results of the radiation source target are output.

2. The method for locating radiation source targets in an urban environment based on deep learning according to claim 1, characterized in that, The method for constructing the simulated geo-electromagnetic dataset includes: Import publicly available urban geographic information into high-precision electromagnetic simulation software; The electromagnetic properties of city buildings are set, as are the antenna parameters and location of the radiation source and the receiving antenna parameters. The signal propagation of the radiation source is simulated at a fixed height, and spatial and gray-scale quantization is performed to obtain a rasterized image representing the global path loss map of the radiation source. The geometric shape of the city buildings represented by the rasterized gray-scale image from the overhead view is saved as urban geographic information. The corresponding spatial coordinates of the radiation source are saved, and the simulation process is repeated to obtain the simulated geo-electromagnetic dataset.

3. The method for locating radiation source targets in an urban environment based on deep learning according to claim 1, characterized in that, The method for constructing the real-world geographic-electromagnetic dataset includes: Drones collect spatial-spectrum data and correlate them with urban geographic environment at a fixed altitude; The spatial data and intensity data of the spatial-spectral data are pre-quantized separately to obtain spatial spectral data represented by a rasterized image. The associated urban geographic environment is quantified to obtain a quantized rasterized image representation of the urban geographic environment; The images are merged to obtain a real-world geographic-electromagnetic dataset.

4. The method for locating radiation source targets in an urban environment based on deep learning according to claim 1, characterized in that, The specific design process of the encoder is as follows: Construct a simplified path loss model for urban environments; Based on the path loss model, dilated convolutional modules with different porosity are constructed. These modules include low-porosity, medium-porosity, and high-porosity modules. The low-porosity module describes the propagation characteristics of short-range free space through local feature extraction. The medium-porosity module reflects the signal shadowing effect caused by building shading by modeling the mid-scale space distribution of building walls. The high-porosity module is used to capture features from the building cluster scale to the regional scale.

5. The method for locating radiation source targets in an urban environment based on deep learning according to claim 4, characterized in that, The void ratio of the low void ratio module is set to S1, the void ratio of the medium void ratio module is set to S2, and the void ratio of the high void ratio module is set to S3, wherein S1 < S2 < S3.

6. The method for locating radiation source targets in an urban environment based on deep learning according to claim 1 or 4, characterized in that, The decoder includes a path loss prediction output layer. In the path loss prediction output layer, two upsampling operations are first performed through two deconvolution modules to gradually restore the high-level features compressed by the encoder to the input resolution, preserving the macroscopic propagation characteristics in the urban environment. Then, the number of channels is reduced to 1 through a convolution module. Finally, the path loss is constrained to the range of [-1,1] through an activation function for normalized loss prediction.

7. The method for locating radiation source targets in an urban environment based on deep learning according to claim 6, characterized in that, The decoder includes a positioning output layer. In the positioning output layer, firstly, the output results after two upsampling operations are input into two convolutional modules to extract positioning-sensitive features; then, the number of channels is compressed to 1 by the downsampling module, and a normalized probability distribution heatmap is generated by the activation function; next, normalized network coordinates are pre-generated and decoupled from the input resolution; finally, the centroid is calculated by the centroid calculation module using the weighted average of the heatmap and the pre-stored network coordinates, which is the positioning result of the radiation source target.

8. The method for locating radiation source targets in an urban environment based on deep learning according to claim 1, characterized in that, The loss function includes the path loss between the predicted path loss and the actual path loss, quantified using mean squared error, specifically: in, H and W These represent the spatial dimensions of the gain tensor, i.e., the height and width of the feature map, respectively; This represents the predicted path loss value output by the model. This represents the corresponding actual path loss value; The coefficient used to adjust the weight of the loss term.

9. The method for locating radiation source targets in an urban environment based on deep learning according to claim 8, characterized in that, The loss function also includes the positioning error, specifically: in, N Indicates the number of samples in the batch; M This indicates the total dimensions of the positioning output; This represents the model's predicted location value; This represents the corresponding actual location value; This represents the loss weighting coefficient for that item.

10. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed, corresponding to the deep learning-based radiation source target localization method in an urban environment as described in any one of claims 1 to 9.

Images