A direct positioning method for a cellular-free system and related products

Abstract

This application discloses a direct positioning method and related products for cellular-free systems. The method includes: acquiring channel state information; inputting the channel state information into a spatial spectrogram estimation model to obtain a spatial spectrogram output by the spatial spectrogram estimation model; detecting and processing the main lobe and grating lobe of the spatial spectrogram to obtain a detected spatial spectrogram; and inputting the detected spatial spectrogram into a source location estimation model to obtain a source location positioning result output by the source location estimation model. This application not only retains the theoretical advantage of high accuracy in direct positioning but also effectively alleviates the problems of computational redundancy caused by grid search and the degradation of positioning performance caused by grating lobe interference. The final output source location has higher accuracy and robustness. Even if some nodes fail, the distributed architecture can still maintain reliable positioning capability, thereby achieving efficient, stable, and real-time high-precision positioning in complex wireless environments.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a direct positioning method and related products for non-cellular systems. Background Technology

[0002] Cellular-free systems evolved from distributed antenna systems, with the core concept of being "user-centric," completely eliminating the boundary limitations of fixed cells in traditional cellular networks. By deploying a large number of low-power, densely distributed access points (APs), the system can provide users with a continuous and seamless coverage experience and achieve smooth handover during user movement. This architecture not only significantly improves spatial diversity gain but also greatly enhances system channel capacity and spectrum utilization efficiency.

[0003] In terms of positioning, the non-cellular system embeds positioning functionality in a distributed manner into each access node, allowing each node to independently estimate the location of the target information source. Thanks to this decentralized processing mechanism, even if some nodes fail due to faults or interference, the remaining nodes can still work together to maintain network connectivity and continue providing positioning services, thus giving the system excellent robustness and high positioning accuracy.

[0004] Direct Position Determination (DPD), as an advanced positioning paradigm, eliminates the need to estimate intermediate parameters such as angle of arrival (AoA) and time of arrival (ToA). Instead, it directly extracts the radiation source location information from the raw received signal, constructs a cost function (such as MUSIC, MVDR, or maximum likelihood function) that is only related to the target location, and solves for the target location through optimization algorithms. Compared to traditional two-step positioning methods, DPD avoids intermediate parameter estimation errors and pairing errors in the process of associating with the signal source, theoretically enabling higher-precision positioning.

[0005] However, Direct Positioning Devices (DPDs) typically require discretizing the target area into a dense grid and employing a grid search strategy to find the optimal solution for the cost function. While this mechanism is conceptually clear, it faces two major challenges: First, it has high computational complexity, especially in large-scale or high-precision scenarios where the number of grids increases dramatically, leading to a significant increase in computational latency and making it difficult to meet real-time requirements. Second, when DPDs are applied to non-cellular systems, the spatial sampling discontinuities or non-uniform array structures introduced by a large number of distributed nodes easily generate a large number of grating lobes in the constructed spatial spectrum. These grating lobes have similar energies to the main lobe, severely interfering with the accurate determination of the target location, weakening spatial resolution, and ultimately causing a significant decrease in positioning accuracy. Summary of the Invention

[0006] To address the aforementioned issues, this application provides a direct positioning method and related products for non-cellular systems, aiming to reduce the computational complexity of direct positioning and improve positioning accuracy.

[0007] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides a method for direct positioning in a non-cellular system, the method comprising: Obtain channel state information; The channel state information is input into the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model; The main lobe and grating lobe of the spatial spectrum are detected and processed to obtain the detected spatial spectrum; The detected spatial spectrum is input into the source location estimation model to obtain the source location positioning result output by the source location estimation model.

[0008] Optionally, the step of inputting the channel state information into the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model specifically includes: The channel state information is normalized to obtain normalized channel state information; The normalized channel state information is input into the first convolutional layer of the spatial spectrum estimation model, an activation function is applied, and max pooling is performed to obtain the first feature map tensor. The first feature map tensor is input into the second convolutional layer of the spatial spectral estimation model, an activation function is applied, and max pooling is performed to obtain the second feature map tensor. The second feature map tensor is input into the third convolutional layer of the spatial spectral estimation model, and an activation function is applied to obtain the third feature map tensor. The third feature map tensor is flattened into a one-dimensional tensor and sequentially input into the first fully connected layer, the LSTM layer, and the second fully connected layer in the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model.

[0009] Optionally, the normalization process for the channel state information to obtain normalized channel state information specifically includes: The channel state information is transformed into real and imaginary data. The real part data and the imaginary part data are each subjected to independent data normalization processing; The normalized real and imaginary data are constructed into a complex number form as the normalized channel state information.

[0010] Optionally, the step of detecting the main lobe and grating lobe of the spatial spectrum to obtain the detected spatial spectrum specifically includes: The main lobe and grating lobe of the spatial spectrum are detected to find target main lobes and target grating lobes that meet preset requirements. The values ​​of the spatial spectrum located within the preset regions of the target main lobe and target grating lobe are retained, and the values ​​of the spatial spectrum not located within the preset regions of the target main lobe and target grating lobe are set to 0, thus obtaining the detected spatial spectrum.

[0011] Optionally, the step of inputting the detected spatial spectrum into the source location estimation model to obtain the source location localization result output by the source location estimation model specifically includes: The detected spatial spectrum is normalized to obtain a normalized detected spatial spectrum. The normalized detected spatial spectrum is input into the first convolutional layer of the source location estimation model, and an activation function is applied, followed by max pooling, to obtain the fourth feature map tensor. The fourth feature map tensor is input into the second convolutional layer of the source location estimation model, an activation function is applied, and max pooling is performed to obtain the fifth feature map tensor. The fifth feature map tensor is input into the third convolutional layer of the source location estimation model, an activation function is applied, and max pooling is performed to obtain the sixth feature map tensor. The sixth feature map tensor is flattened into a one-dimensional tensor and then sequentially input into the Transform layer, global pooling layer, and fully connected layer in the source location estimation model to obtain the source location localization result output by the source location estimation model.

[0012] Optionally, the spatial spectrum estimation model is a neural network model constructed using a smoothing loss function and an Adam optimizer; the source location estimation model is a neural network model constructed using a crossover loss function and an Adam optimizer.

[0013] A second aspect of this application provides a direct positioning device for a cellular system, the direct positioning device for a cellular system comprising: The acquisition module is used to acquire channel state information; The spatial spectrum determination module is used to input the channel state information into the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model; The processing module is used to detect and process the main lobe and grating lobe of the spatial spectrum to obtain the detected spatial spectrum. The localization module is used to input the detected spatial spectrum into the source location estimation model to obtain the source location localization result output by the source location estimation model.

[0014] A third aspect of this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the direct positioning method for a non-cellular system provided in the first aspect.

[0015] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the direct positioning method for non-cellular systems provided in the first aspect.

[0016] The fifth aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the direct positioning method for non-cellular systems provided in the first aspect.

[0017] Compared with the prior art, this application has the following beneficial effects: This application includes acquiring channel state information; inputting the channel state information into a spatial spectrum estimation model to obtain a spatial spectrum output by the spatial spectrum estimation model; performing detection processing on the main lobe and grating lobe of the spatial spectrum to obtain a detected spatial spectrum; and inputting the detected spatial spectrum into a source location estimation model to obtain a source location positioning result output by the source location estimation model.

[0018] This application generates a spatial spectrum reflecting the spatial energy distribution of a signal by acquiring Channel State Information (CSI) from distributed access points and inputting it into a spatial spectrum estimation model. Addressing the issue of excessive grating lobes in direct positioning within a non-cellular architecture, this application further detects and processes the main lobe and grating lobes of the spatial spectrum, significantly improving its resolution and signal-to-noise ratio. The detected spatial spectrum is then fed into a source location estimation model, where the target location is calculated, avoiding errors introduced by intermediate parameter estimation and data correlation in traditional two-step methods. This not only retains the theoretical advantages of high accuracy in direct positioning but also effectively mitigates the computational redundancy and performance degradation caused by grating lobe interference from grid search. The final output source location exhibits higher accuracy and robustness; even with partial node failure, the distributed architecture maintains reliable positioning capabilities, thus achieving efficient, stable, and real-time high-precision positioning in complex wireless environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of a non-cellular system; Figure 2 A flowchart illustrating a direct positioning method for a cellular-free system provided in this application embodiment; Figure 3 A structural diagram of the spatial spectrum estimation model provided in the embodiments of this application; Figure 4 A schematic diagram of a spatial spectrum provided in an embodiment of this application; Figure 5 A schematic diagram of the detection of the main lobe and grating lobe provided for an embodiment of this application; Figure 6 A structural diagram of the source location estimation model provided in the embodiments of this application; Figure 7 A schematic diagram showing the actual location in the spatial spectrum provided in the embodiments of this application; Figure 8 A schematic diagram of the localization results in a spatial spectrum provided in an embodiment of this application; Figure 9 A schematic diagram illustrating a direct positioning method for a cellular-free system provided in an embodiment of this application; Figure 10 This is a structural diagram of a direct positioning device for a cellular-free system provided in an embodiment of this application. Detailed Implementation

[0021] Previous implementations included a method based on dual-antenna drone signal reception using a CNN-LSTM network for direction of arrival (DOA) estimation. This method extracts features from the amplitude and phase of the dual-channel signals, achieving near 100% accuracy in classifying and recognizing the angle of near-range drone signals. Another method combines a convolutional neural network (CNN) and a long short-term memory (LSTM) network for Wi-Fi fingerprint indoor positioning. Even using only a single Wi-Fi router, it achieved an average positioning error of 2.5 meters, with the error not exceeding 4 meters in 80% of cases.

[0022] This application provides a method for generating direct location data in a cellular system and related products. The method includes: acquiring channel state information; inputting the channel state information into a spatial spectrum estimation model to obtain a spatial spectrum output by the spatial spectrum estimation model; performing detection processing on the main lobe and grating lobe of the spatial spectrum to obtain a detected spatial spectrum; and inputting the detected spatial spectrum into a source location estimation model to obtain a source location positioning result output by the source location estimation model.

[0023] This application focuses on near-field localization, while related technologies are more often applied to far-field DOA estimation. This application employs a spatial spectrogram estimation model for spatial spectrum estimation, addressing a regression problem; whereas existing technologies use CNN-LSTM networks for target source angle estimation, which is a classification problem. This application relies on a distributed node system, providing more robust localization performance, maintaining a certain level of localization capability even if some nodes fail; in contrast, existing technologies typically rely on a single node or device. This application uses a spatial spectrogram estimation model to generate and optimize the spatial spectrogram, thereby achieving accurate localization; while existing technologies may use indirect methods such as CNN-LSTM networks for Wi-Fi fingerprint matching for localization.

[0024] Explanation of relevant professional terms: Direct positioning: The target position can be directly calculated from the raw received signal without estimating the intermediate parameters required for positioning.

[0025] Cellular non-cellular systems: Cellular non-cellular systems are a new type of wireless access network that abandons the traditional cellular cell division and adopts large-scale distributed access points with global coherent cooperation.

[0026] Spectrum: Generally refers to power spectrum, which represents the spatial distribution of power spectrum. This invention considers the signal spectrum in the XOY plane.

[0027] Main lobe and grating lobe: The main lobe is the lobe with the highest radiation intensity or gain in the antenna pattern (radiation pattern). The grating lobe refers to a false strong beam that appears in space with the same amplitude as the main lobe.

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0029] Figure 1 This is a schematic diagram of a non-cellular system, such as... Figure 1As shown, in a cellular-free scenario, the system no longer relies on the cell division and central base station architecture of traditional cellular networks. Instead, a large number of distributed access points (APs) jointly provide services to users within the area. These access points are connected to a centralized processing unit via fronthaul links, enabling coordinated signal reception, joint processing, and resource sharing. Under this architecture, user terminals maintain uplinks with multiple access points simultaneously, eliminating strict cell boundaries and handover mechanisms. This effectively alleviates problems such as co-channel interference, edge effects, and uneven coverage found in traditional cellular systems. For this scenario, this application proposes a direct positioning method for cellular-free systems. Figure 2 A flowchart of a direct positioning method for a cellular-free system provided in this application embodiment is shown below. Figure 2 The method for direct positioning in a cellular-free system, as shown, includes: S201: Obtain channel state information.

[0030] This application does not limit the method for obtaining channel state information; an example is provided: In the field of wireless communication, channel state information (CSI) characterizes the inherent channel properties of a communication link. Essentially, it describes the fading characteristics of a signal after multipath propagation, specifically reflected in the values ​​of each element in the channel gain matrix H. It encompasses key physical characteristics of multipath propagation processes such as signal scattering, environmental fading, and distance attenuation. CSI can reflect the fine-grained channel propagation characteristics of a communication link and, due to its strong correlation with spatial location, is often used as a core positioning parameter in wireless positioning technology.

[0031] Assume the system has a total of The antenna, the first The positions of the antennas are as follows: The source location is There are a total of The frequency, the first A frequency is defined as The signal modeling in this application is as follows: ………………………(1) in, To receive the signal, model it as follows Matrix; This is the channel gain matrix, with the following dimensions: , can be represented as: …………(2) in, To transmit the signal, model it as follows: Matrix; For signal noise, model as Matrix; This refers to the number of antennas. This refers to the number of snapshots; The number of signal frequencies; in formula (2) This is channel state information; its first m element Indicates the first The channel state information measured by each antenna can be represented as: ………………..(3) in, The channel fading coefficient; ; The speed of light; The wavelength is [wavelength]. Wireless channels typically include multipath effects, in which case... This is an overlay of channel state information from multiple paths.

[0032] S202: Input the channel state information into the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model.

[0033] The spatial spectrum estimation model in this application aims to model the complex nonlinear mapping relationship between channel state information (CSI) and the spatial spectrum. The core function of the spatial spectrum estimation model is to automatically extract high-dimensional features related to the spatial location of the radiation source from the original CSI and generate a continuous spectrum that accurately reflects the distribution of signal energy in the spatial angular or positional domains. To achieve this goal, the model needs sufficient expressive power to capture the influence of multipath effects, near-field spherical wavefronts, distributed node geometry, and other factors on the spatial spectrum morphology.

[0034] It is important to emphasize that this application does not impose mandatory restrictions on the specific network architecture or training strategy of the spatial spectrum estimation model, as long as it can effectively learn the aforementioned nonlinear relationships and output high-quality spatial spectra. For example, in a typical implementation, a deep neural network (such as a convolutional neural network CNN, U-Net, or Transformer) can be used as the basic architecture, combined with a smoothing loss function (such as L2 loss, perceptual loss, or spectral correlation loss) to avoid excessive oscillations or spurious spikes in the output spectrum; at the same time, the Adam optimizer is used for end-to-end training to balance convergence speed and generalization performance. In addition, the model can also introduce physical priors (such as array manifold constraints) or adopt semi-supervised learning strategies to further improve robustness under low signal-to-noise ratio or sparse sampling conditions.

[0035] Through this flexible and powerful modeling approach, the spatial spectral estimation model of this application can not only adapt to the deployment characteristics of distributed and heterogeneous nodes in non-cellular systems, but also provide high-fidelity, low-noise spectral input for subsequent main lobe / grating lobe detection and precise positioning, thus laying a key foundation for achieving high-precision direct positioning.

[0036] S203: Detect and process the main lobe and grating lobe of the spatial spectrum to obtain the detected spatial spectrum.

[0037] This step aims to accurately identify and locate the main lobe and grating lobes in the spatial spectrum, providing a reliable basis for subsequent applications such as signal source localization and parameter estimation. The main lobe typically represents the main propagation direction of the signal, while the grating lobe is a phenomenon of abnormally increased sidelobe gain caused by array antenna design or insufficient signal sampling. In the non-cellular system considered in this application, the distributed antenna deployment and discontinuous spatial sampling by the array elements result in a severe grating lobe problem. This application does not impose strict limitations on the specific execution of this process, but rather selects appropriate algorithms and technologies based on actual needs and application scenarios. The following are several possible optimization and extension methods: The Cell-Averaging Constant False Alarm Rate (CA-CFAR) algorithm is used to detect the presence of target signals in noisy environments. In this context, the spatial spectrum can be viewed as a two-dimensional search space, and the CA-CFAR algorithm automatically adjusts the threshold to distinguish the main lobe and grating lobe from background noise. This method calculates the average power level of the reference cell and sets a dynamic threshold accordingly to ensure a constant false alarm probability under different signal-to-noise ratio conditions.

[0038] Machine learning / deep learning-based methods can train a classification model (e.g., a convolutional neural network, CNN) to directly learn and identify the positions of the main lobe and grating lobes from the spatial spectral map. The advantage of this approach is its ability to handle complex spatial spectral morphologies, especially in the presence of multipath effects or interference from multiple signal sources.

[0039] Physically guided filtering and transformation, based on the design parameters of the array antenna (such as element spacing and array shape), applies specific mathematical transformations (such as Fourier transform and beamforming transform) to enhance the main lobe characteristics while suppressing the influence of the grating lobe. This requires a deep understanding of the basic principles of antenna arrays and adjustments based on specific engineering practices.

[0040] S203 goes beyond using a single technique for main lobe and grating lobe detection; it encourages the flexible selection or combination of different algorithm strategies based on specific circumstances to achieve optimal processing results. Through detailed analysis and processing of the spatial spectrum, the accuracy and reliability of subsequent signal processing tasks can be effectively improved.

[0041] S204: Input the detected spatial spectrum into the source location estimation model to obtain the source location positioning result output by the source location estimation model.

[0042] The source location estimation model in this application aims to establish a high-dimensional nonlinear mapping relationship between the detected spatial spectrum and the actual physical location of the source. Since non-cellular systems typically operate in the near-field region, signal propagation exhibits spherical wave characteristics and is affected by factors such as multipath propagation, occlusion, and uneven node distribution. Traditional parameterization methods based on the far-field plane wave assumption (such as MUSIC and MVDR) are difficult to apply directly. Therefore, this application adopts a data-driven approach, using a deep neural network to regress the two-dimensional or three-dimensional coordinates of the target source end-to-end from the detected spatial spectrum.

[0043] It should be noted that this application does not impose mandatory restrictions on the specific structure or training details of the source location estimation model, as long as it can effectively learn the mapping from the spatial spectrum to location coordinates. In a typical implementation, a neural network consisting of fully connected layers, convolutional layers, or attention mechanisms can be constructed, and the cross-entropy loss function (suitable for discretized grid positioning scenarios) or mean squared error (MSE, suitable for continuous coordinate regression) can be used as the optimization objective; at the same time, the Adam optimizer is used for efficient training to balance convergence speed and generalization ability. The source location estimation model can not only make full use of the high-quality spatial spectrum that has been denoised and de-correlated in the previous steps, but also achieve high-precision, low-latency, and strong generalization direct positioning capability in complex wireless environments, which is significantly better than the traditional two-step method that relies on intermediate parameter estimation, providing reliable location awareness support for integrated non-cellular communication and sensing systems.

[0044] In non-cellular system scenarios, this application fully leverages the powerful nonlinear fitting capabilities of deep learning neural networks to construct an end-to-end direct localization framework. First, a spatial spectrum estimation model learns the complex nonlinear mapping relationship between Channel State Information (CSI) collected from distributed access points and the spatial energy distribution of the target source, thereby reconstructing a spatial spectrum reflecting the direction or location characteristics of the signal source with high accuracy. Subsequently, the spatial spectrum, after main lobe enhancement and grating lobe suppression detection, is input into the source location estimation model. This model further learns and optimizes the nonlinear correlation between the optimized spectrum and the actual physical location of the source, directly outputting the target's two-dimensional or three-dimensional coordinates. This application eliminates the dependence on intermediate parameters in traditional two-step localization, avoiding parameter estimation errors and data association errors. It is particularly suitable for complex electromagnetic environments in non-cellular systems characterized by near-field, multipath density, and non-uniform node distribution. Through the collaborative work of two deep neural networks, not only is an end-to-end mapping from the original CSI to the final location achieved, but localization accuracy, robustness, and real-time performance are also significantly improved, providing an efficient and scalable technical path for the fusion of non-cellular communication and sensing.

[0045] The above describes the main technical solution of this application. Further implementations of the main technical solution are now introduced. Details are as follows: Regarding S202, which inputs the channel state information into the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model, this application provides an optional embodiment: The channel state information is normalized to obtain normalized channel state information.

[0046] This application does not limit the normalization processing method. Considering that the channel state information is in the form of a virtual exponent, this application provides an example: The channel state information is transformed into real and imaginary data.

[0047] The real part data and the imaginary part data are each processed independently for data normalization.

[0048] Formula (4) is used to perform independent data normalization on the real and imaginary data, mapping the features of any distribution to a standard normal distribution with a mean of 0 and a variance of 1. This eliminates the influence of differences in feature dimensions and numerical ranges, and makes the feature distribution more consistent with the training characteristics of the neural network. ………………………(4) in, These are the standardized sample values. To obtain the mean of this feature dimension in the training set, The standard deviation of this feature dimension in the training set.

[0049] The normalized real and imaginary data are constructed into a complex number form as the normalized channel state information.

[0050] Normalization of channel state information can preserve its amplitude and phase characteristics, eliminating the influence of the difference in dimensions between the real and imaginary parts.

[0051] The normalized channel state information is input into the first convolutional layer of the spatial spectral estimation model, an activation function is applied, and max pooling is performed to obtain the first feature map tensor.

[0052] The first feature map tensor is input into the second convolutional layer of the spatial spectral estimation model, an activation function is applied, and max pooling is performed to obtain the second feature map tensor.

[0053] The second feature map tensor is input into the third convolutional layer of the spatial spectral estimation model, and an activation function is applied to obtain the third feature map tensor.

[0054] The third feature map tensor is flattened into a one-dimensional tensor and sequentially input into the first fully connected layer, the LSTM layer, and the second fully connected layer in the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model.

[0055] Figure 3 A structural diagram of the spatial spectrum estimation model provided in the embodiments of this application, such as... Figure 3 As shown, the standardized channel state information data is imported into a three-layer spatial spectrum estimation model. The specific steps are as follows: The obtained channel state information data X is input into the first convolutional layer conv1, which uses the ReLU activation function and performs max pooling.

[0056] The input continues into the second convolutional layer conv2, which also uses the ReLU activation function and performs max pooling.

[0057] The input continues to the third convolutional layer conv3, which also uses the ReLU activation function and performs max pooling.

[0058] The 3D feature map tensor output by the convolutional layer is flattened into a 1D tensor and input into the fully connected layer Fc1. The output of the fully connected layer Fc1 is used as the input to the subsequent LSTM layer.

[0059] Construct a bidirectional LSTM layer, take the one-dimensional tensor output of the fully connected layer as input, and use its output as input to the fully connected layer Fc2. The output of the fully connected layer Fc2 is the final localization result Y.

[0060] Through the aforementioned multi-layered convolutional and recurrent neural network structure, the spatial spectrogram estimation model of this application can fully mine the potential features in the channel state information, provide high-quality spatial spectrograms for subsequent processing, and significantly improve the direct positioning accuracy and robustness in non-cellular systems.

[0061] Regarding the training method for the spatial spectrum estimation model, this application also provides an optional embodiment: Using the smoothed L1 loss function as the loss function, its theoretical formula is: ……………(5) in, ; This is a predicted value; This is the actual value; this framework Use the default value of 1.

[0062] The Adam (Adaptive Moment Estimation) optimizer is used to iteratively update the model parameters, and the model training process can be rapidly converged and effectively approximated to the global optimum through fine-tuning of the learning rate.

[0063] Repeat the training process until the loss function converges to a stable region and no longer shows a significant decrease, at which point training is stopped.

[0064] After the model is trained until the loss converges and reaches a stable state, a test dataset is selected. This dataset is preprocessed according to the model input requirements and then input into the trained spatial spectrum estimation model. The model then performs inference operations on the test data to generate the corresponding spatial spectrum. The system verifies the effectiveness of the model's output spatial spectrum from multiple core evaluation dimensions, including spatial spectrum resolution, target localization accuracy, spectral peak identification, and noise suppression capability. It quantitatively analyzes the degree of agreement between the spatial spectrum and the real-world spatial spectrum distribution, thereby evaluating the model's effectiveness and reliability in the spatial spectrum estimation task. The obtained model is then used as the spatial spectrum estimation model.

[0065] Figure 4 A schematic diagram of the spatial spectrum provided in the embodiments of this application, such as Figure 4 As shown, the spatial spectrum map, output by the spatial spectrum estimation model, represents the visualization of the target signal source energy distribution on a two-dimensional plane (X and Y directions). The map is presented as a heatmap, with the horizontal axis X (unit: meters) and the vertical axis Y (unit: meters) representing the spatial coordinates of the location area. The color intensity corresponds to the signal energy strength at that location. The color bars on the right represent the normalized spatial spectrum value range (0.03–0.09), where the red area represents the location with the strongest energy, most likely the location of the target signal source.

[0066] As can be seen from the figure, there is a significant high-energy main lobe at coordinates (10, 10), which appears as a bright orange-red spot, indicating that this location is a potential source of the main radiation. At the same time, several weaker energy responses are also distributed in other regions, some of which may be pseudo-peaks (i.e., grating lobes) caused by multipath reflection or interference. Although these low-intensity regions are not the dominant signal source, they reflect the spatial characteristics of a complex propagation environment.

[0067] This spatial spectrum serves as a crucial input for subsequent processing, enabling the identification of the main lobe and grating lobe, peak detection, and precise location. Further analysis of the spatial spectrum using a source location estimation model effectively suppresses spurious responses and accurately pinpoints the true source location, achieving high-precision, robust, cellless direct localization. This figure visually demonstrates the ability of the proposed method to generate high-quality spatial spectra in near-field environments, validating its effectiveness and feasibility in practical applications.

[0068] Regarding the detection processing of the main lobe and grating lobe of the spatial spectrum in S203 to obtain the detected spatial spectrum, this application provides an optional embodiment: The main lobe and grating lobe of the spatial spectrum are detected to find target main lobes and target grating lobes that meet preset requirements. The values ​​of the spatial spectrum located within the preset regions of the target main lobe and target grating lobe are retained, and the values ​​of the spatial spectrum not located within the preset regions of the target main lobe and target grating lobe are set to 0, thus obtaining the detected spatial spectrum.

[0069] In a specific implementation, the Cell-Averaging Constant False Alarm Rate (CA-CFAR) algorithm can be used to detect the main lobe and grating lobe of the spatial spectrum, and identify the main lobe and grating lobe that meet the requirements. Figure 5 This is a schematic diagram of the detection of the main lobe and grating lobe provided in an embodiment of this application, as shown below. Figure 5 In the two-dimensional scene shown, the CUT cell, i.e., the red area in the figure, is the smallest cell to be detected. Since spectral peaks in spatial spectra often exhibit "diffusion" in practical applications—meaning a peak doesn't occupy only one cell or has side lobes around it—a guard cell of a certain width is needed, i.e., the green area in the figure. The reference cell, i.e., the yellow area in the figure, is used to determine whether the CUT cell is a peak point. If the average value of the reference cells is greater than a set threshold, the CUT cell is a peak point; if the average value of the reference cells is less than the set threshold, the CUT cell is not a peak point. The threshold can be obtained by selecting the value with the best peak point detection effect through multiple experiments. In this way, the main lobe and grating lobe can be detected. Then, the values ​​of the spatial spectra near the main lobe and grating lobe are retained, while the values ​​of the spatial spectra at locations that do not meet the requirements are set to 0, constructing a new spatial spectra.

[0070] Regarding S204, which inputs the detected spatial spectrum into the source location estimation model to obtain the source location positioning result output by the source location estimation model, this application provides an optional embodiment: The detected spatial spectrum is normalized to obtain a normalized detected spatial spectrum.

[0071] This application does not limit the normalization processing method of this embodiment. For example, formula (4) can be used to normalize the detected spatial spectrum to obtain the normalized detected spatial spectrum.

[0072] The normalized detected spatial spectrum is input into the first convolutional layer of the source location estimation model, and an activation function is applied, followed by max pooling, to obtain the fourth feature map tensor.

[0073] The fourth feature map tensor is input into the second convolutional layer of the source location estimation model, and an activation function is applied, followed by max pooling, to obtain the fifth feature map tensor.

[0074] The fifth feature map tensor is input into the third convolutional layer of the source location estimation model, an activation function is applied, and max pooling is performed to obtain the sixth feature map tensor.

[0075] The sixth feature map tensor is flattened into a one-dimensional tensor and sequentially input into the Transform layer (including normalization layer 1, multi-head self-attention layer, residual connection layer 1, normalization layer 2, feedforward neural network layer, residual connection layer 2), global pooling and fully connected layers in the source location estimation model to obtain the source location localization result output by the source location estimation model.

[0076] Figure 6 This is a structural diagram of the source location estimation model provided in the embodiments of this application, as shown below. Figure 6 As shown, the spatial spectrum data is normalized using formula (4) and mapped to a standard normal distribution with a mean of 0 and a variance of 1.

[0077] The normalized spatial spectral data is input into a three-layer convolutional neural network to extract its local features. The specific steps are as follows: The obtained spatial spectral data X is input into the first convolutional layer conv1, which uses the ReLU activation function and performs max pooling.

[0078] The input continues into the second convolutional layer conv2, which also uses the ReLU activation function and performs max pooling.

[0079] The input continues to the third convolutional layer conv3, which also uses the ReLU activation function and performs max pooling.

[0080] The 3D feature map tensor output by the convolutional layer is flattened into a 1D tensor and input into the normalization layer 1 (LayerNorm1), the multihead attention layer (MultiheadAttention), the residual connection layer 1 (ResidualConnection1), the normalization layer 2 (LayerNorm2), the feedforward neural network layer (Feed Forward Net), and the residual connection layer 2 (Residual Connection2).

[0081] The output of the Transform layer is subjected to global pooling, and the result of global pooling is input into the fully connected layer Fc. The output of the fully connected layer Fc is the final localization result Y.

[0082] This ensures an efficient conversion from spatial spectral maps to precise location estimates, improving positioning accuracy and robustness.

[0083] Regarding the training method of the source location estimation model, this application provides an optional embodiment: The CrossEntropyLoss function is used as the loss function, and its core formula is as follows: ………(6) in, For the first The sample corresponds to the first The original output of the class model; For the first The logit value corresponding to the true class of each sample; The number of samples; This represents the number of categories; in the cross-entropy loss function, the reduction variable is set to mean, which is the default parameter.

[0084] By using the Adam optimizer to iteratively update model parameters and through fine-grained learning rate scheduling, the model training convergence speed is accelerated, while the model's ability to approach the global optimum is improved.

[0085] The training process is executed iteratively until the loss function enters a stable region and the value does not decrease significantly, at which point the training is terminated.

[0086] Input the test dataset, compare the estimated location output by the model with the actual physical location of the information source in the test set, calculate the spatial relative distance between the two, and quantitatively evaluate the model's positioning accuracy and generalization ability, thereby verifying the effectiveness and reliability of the model in real-world scenarios. The obtained model is then used as the information source location estimation model.

[0087] Figure 7 This is a schematic diagram showing the actual location in the spatial spectrum provided in the embodiments of this application. Figure 8 This is a schematic diagram illustrating the localization results in a spatial spectral map provided in an embodiment of this application. For example... Figure 7 As shown in the figure, the red marker indicates the actual physical location of the signal source, located at coordinates (10 m, 10 m). This location corresponds to the main lobe peak point in the original spatial spectrum generated by the MUSIC algorithm, representing the location where the real signal energy is strongest.

[0088] Figure 8The prediction results output by the source location estimation model of this application are shown, such as... Figure 8 As shown, a significant high-value response also appears near the coordinates (10m, 10m), and its predicted peak position highly coincides with the actual position, indicating that the model can accurately identify and regress the true spatial coordinates of the source. By comparison, it can be seen that the source location estimation model constructed in this application, after processing the optimized spatial spectrum, can effectively learn the nonlinear mapping relationship between the spatial spectrum and the true position, achieving high-precision, low-bias direct positioning. Experimental results verify the feasibility and superiority of this application in near-field non-cellular systems, with the positioning error controlled within an acceptable range.

[0089] Figure 9 This is a schematic diagram of a direct positioning method for a cellular-free system provided in an embodiment of this application, as shown below. Figure 9 As shown, a centralized processing architecture is adopted, with a single central processing unit (CPU) uniformly receiving Channel State Information (CSI) collected from multiple distributed access points (APs). After receiving the CSI data, signal modeling is performed, and the CSI is input into a CNN-LSTM-based spatial spectrum estimation model. Utilizing its powerful nonlinear feature extraction capabilities, a spatial spectrum reflecting the spatial energy distribution of the target source is generated. Peak point detection is performed on the generated spatial spectrum, i.e., identifying and distinguishing between the main lobe and the grating lobe. By suppressing spurious responses and enhancing the true main lobe signal, the resolution and reliability of the spatial spectrum are improved. The processed high-quality spatial spectrum is then fed into a CNN-Transformer-based source location estimation model. This model combines the local feature extraction capabilities of convolutional neural networks with the global context modeling advantages of Transformers to achieve end-to-end mapping from the spatial spectrum to the target's physical coordinates, ultimately outputting a high-precision source location result.

[0090] It achieves closed-loop processing from raw CSI to direct positioning, avoiding intermediate parameter estimation errors in the traditional two-step method, and significantly improving positioning accuracy, robustness and real-time performance. It is suitable for sensing and positioning tasks of non-cell cooperative systems in complex electromagnetic environments.

[0091] This application rationally decomposes the entire localization process into two stages: spectral estimation and location estimation, and designs dedicated deep learning models to implement each stage. In the spectral estimation stage, considering that the spatial location characteristics of the target source are implicit in the channel state information (CSI), a CNN-LSTM neural network architecture is used to perform end-to-end modeling of the CSI. Utilizing its powerful nonlinear fitting capability, a high-fidelity spatial spectrum is directly learned from the original CSI to generate, thereby completely avoiding the large amount of redundant computation caused by the grid search reliance in traditional direct localization methods, significantly reducing system complexity and improving real-time processing performance.

[0092] In the location estimation stage, to address the challenge of dense raster lobes and difficulty in main lobe identification in the spatial spectral map due to the distributed node structure and near-field effect in cellular-free systems, a CNN-Transformer hybrid architecture is further introduced to transform the localization task into an image recognition problem. By modeling the spatial spectral map as a classification task for each row and column (e.g., if the true location in a 20×20 spectral map is in row 5 and column 10, then the corresponding row category is 5 and the column category is 10), the model can effectively focus on the main lobe region, suppress raster lobe interference, and accurately regress the target coordinates. This strategy not only enhances the ability to distinguish the main lobe but also improves the overall localization robustness.

[0093] In summary, this application achieves efficient spectral reconstruction through CNN-LSTM and completes high-precision location decoding by combining it with CNN-Transformer, forming a complete, low-complexity, and high-precision cellular-free direct positioning solution. While avoiding intermediate parameter estimation errors, it significantly improves the positioning accuracy and practicality of the system in complex wireless environments.

[0094] Figure 10 A structural diagram of a direct positioning device for a cellular-free system provided in this application embodiment is shown below. Figure 10 As shown, based on the direct positioning method for a non-cellular system provided in the preceding embodiments, this application also provides a corresponding direct positioning device for a non-cellular system, comprising: The acquisition module is used to acquire channel state information; The spatial spectrum determination module is used to input the channel state information into the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model; The processing module is used to detect and process the main lobe and grating lobe of the spatial spectrum to obtain the detected spatial spectrum. The localization module is used to input the detected spatial spectrum into the source location estimation model to obtain the source location localization result output by the source location estimation model.

[0095] As an optional embodiment, the spatial spectrum determination module specifically includes: The normalization unit is used to normalize the channel state information to obtain normalized channel state information.

[0096] The first convolutional unit is used to input the normalized channel state information into the first convolutional layer of the spatial spectrum estimation model, apply an activation function, and perform max pooling to obtain the first feature map tensor.

[0097] The second convolutional unit is used to input the first feature map tensor into the second convolutional layer in the spatial spectral estimation model, apply an activation function, and perform max pooling to obtain the second feature map tensor.

[0098] The third convolutional unit is used to input the second feature map tensor into the third convolutional layer in the spatial spectral estimation model, and obtain the third feature map tensor by using an activation function.

[0099] The connection unit is used to flatten the third feature map tensor into a one-dimensional tensor and input it sequentially into the first fully connected layer, the LSTM layer and the second fully connected layer in the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model.

[0100] As an optional embodiment, the normalization unit is specifically used for: The channel state information is transformed into real and imaginary data. The real part data and the imaginary part data are each subjected to independent data normalization processing; The normalized real and imaginary data are constructed into a complex number form as the normalized channel state information.

[0101] As an optional embodiment, the processing module is specifically used for: The main lobe and grating lobe of the spatial spectrum are detected to find target main lobes and target grating lobes that meet preset requirements. The values ​​of the spatial spectrum located within the preset regions of the target main lobe and target grating lobe are retained, and the values ​​of the spatial spectrum not located within the preset regions of the target main lobe and target grating lobe are set to 0, thus obtaining the detected spatial spectrum.

[0102] As an optional embodiment, the positioning module is specifically used for: The detected spatial spectrum is normalized to obtain a normalized detected spatial spectrum. The normalized detected spatial spectrum is input into the first convolutional layer of the source location estimation model, and an activation function is applied, followed by max pooling, to obtain the fourth feature map tensor. The fourth feature map tensor is input into the second convolutional layer of the source location estimation model, an activation function is applied, and max pooling is performed to obtain the fifth feature map tensor. The fifth feature map tensor is input into the third convolutional layer of the source location estimation model, an activation function is applied, and max pooling is performed to obtain the sixth feature map tensor. The sixth feature map tensor is flattened into a one-dimensional tensor and sequentially input into the Transform layer (including normalization layer 1, multi-head self-attention layer, residual connection layer 1, normalization layer 2, feedforward neural network layer, residual connection layer 2), global pooling and fully connected layers in the source location estimation model to obtain the source location localization result output by the source location estimation model.

[0103] As an optional embodiment, the spatial spectrum estimation model is a neural network model constructed using a smoothing loss function and an Adam optimizer; the source location estimation model is a neural network model constructed using a crossover loss function and an Adam optimizer.

[0104] This application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a direct positioning method for non-cellular systems.

[0105] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a direct positioning method for non-cellular systems.

[0106] This application provides a computer program product, including a computer program that, when executed by a processor, implements a direct positioning method for non-cellular systems.

[0107] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and equipment embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and equipment embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0108] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A direct positioning method for a cellular-free system, characterized in that, The method includes: Obtain channel state information; The channel state information is input into the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model; The main lobe and grating lobe of the spatial spectrum are detected to find target main lobe and target grating lobe that meet the preset requirements. The values ​​of the spatial spectrum located within the preset regions of the target main lobe and target grating lobe are retained, and the values ​​of the spatial spectrum not located within the preset regions of the target main lobe and target grating lobe are set to 0 to obtain the detected spatial spectrum. The detected spatial spectrum is input into the source location estimation model to obtain the source location positioning result output by the source location estimation model.

2. The direct positioning method for a non-cellular system according to claim 1, characterized in that, The step of inputting the channel state information into the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model specifically includes: The channel state information is normalized to obtain normalized channel state information; The normalized channel state information is input into the first convolutional layer of the spatial spectrum estimation model, an activation function is applied, and max pooling is performed to obtain the first feature map tensor. The first feature map tensor is input into the second convolutional layer of the spatial spectral estimation model, an activation function is applied, and max pooling is performed to obtain the second feature map tensor. The second feature map tensor is input into the third convolutional layer of the spatial spectral estimation model, and an activation function is applied to obtain the third feature map tensor. The third feature map tensor is flattened into a one-dimensional tensor and sequentially input into the first fully connected layer, the LSTM layer, and the second fully connected layer in the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model.

3. The direct positioning method for a non-cellular system according to claim 2, characterized in that, The normalization process for the channel state information to obtain normalized channel state information specifically includes: The channel state information is transformed into real and imaginary data. The real part data and the imaginary part data are each subjected to independent data normalization processing; The normalized real and imaginary data are constructed into a complex number form as the normalized channel state information.

4. The direct positioning method for a non-cellular system according to claim 1, characterized in that, The step of inputting the detected spatial spectrum into the source location estimation model to obtain the source location localization result output by the source location estimation model specifically includes: The detected spatial spectrum is normalized to obtain a normalized detected spatial spectrum. The normalized detected spatial spectrum is input into the first convolutional layer of the source location estimation model, and an activation function is applied, followed by max pooling, to obtain the fourth feature map tensor. The fourth feature map tensor is input into the second convolutional layer of the source location estimation model, an activation function is applied, and max pooling is performed to obtain the fifth feature map tensor. The fifth feature map tensor is input into the third convolutional layer of the source location estimation model, an activation function is applied, and max pooling is performed to obtain the sixth feature map tensor. The sixth feature map tensor is flattened into a one-dimensional tensor and then sequentially input into the Transform layer, global pooling layer, and fully connected layer in the source location estimation model to obtain the source location localization result output by the source location estimation model.

5. The direct positioning method for a non-cellular system according to claim 1, characterized in that, The spatial spectrum estimation model is a neural network model constructed using a smoothing loss function and an Adam optimizer; the source location estimation model is a neural network model constructed using a crossover loss function and an Adam optimizer.

6. A direct positioning device for a non-cellular system, characterized in that, The direct positioning device for the non-cellular system includes: The acquisition module is used to acquire channel state information; The spatial spectrum determination module is used to input the channel state information into the spatial spectrum estimation model to obtain the spatial spectrum output by the spatial spectrum estimation model; The processing module is used to detect the main lobe and grating lobe of the spatial spectrum, find the target main lobe and target grating lobe that meet the preset requirements, retain the values ​​of the spatial spectrum located within the preset regions of the target main lobe and target grating lobe, and set the values ​​of the spatial spectrum not located within the preset regions of the target main lobe and target grating lobe to 0, so as to obtain the detected spatial spectrum. The localization module is used to input the detected spatial spectrum into the source location estimation model to obtain the source location localization result output by the source location estimation model.

7. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the direct positioning method for a non-cellular system according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the direct positioning method for non-cellular systems as described in any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the direct positioning method for non-cellular systems as described in any one of claims 1-5.

Images