Network interference positioning method and device based on multi-source data fusion

Abstract

The application provides a network interference positioning method and device based on multi-source data fusion, relates to the technical field of communication, and comprises the following steps: extracting a frequency domain feature vector of radio frequency signal data, a space feature vector of environment perception data and a topology feature vector of network topology data, and performing cross-modal alignment on the frequency domain feature vector and the space feature vector; dynamically adjusting a fusion weight based on real-time network load, fusing each feature vector based on the fusion weight, and obtaining a fusion feature tensor; performing non-line-of-sight propagation correction on the fusion feature tensor to obtain a positioning parameter set, and determining the position of an interference source based on the positioning parameter set. The method and device provided by the application use multi-source data for positioning, perform cross-modal alignment on features of different modes, dynamically adjust the fusion weight based on real-time network load, and perform targeted non-line-of-sight propagation correction, so that precise identification and rapid positioning of the interference source can be realized when the network is in a heavy load state.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a network interference localization method and apparatus based on multi-source data fusion. Background Technology

[0002] With the continuous evolution of 5G technology and future network architecture, cellular communication systems are placing higher demands on network quality, service continuity, and interference recovery capabilities. In scenarios involving dense networking, dynamic loads, and complex electromagnetic environments, high-precision location of interference sources not only directly affects network throughput and user experience but also relates to communication security and stability in critical areas. Therefore, multi-source data-driven methods integrating radio frequency characteristics and environmental awareness are gradually becoming a research hotspot, and related technologies are shifting from static rule-driven to intelligent decision-driven approaches.

[0003] Traditional interference source localization typically relies on Received Signal Strength Indication (RSSI) fingerprint matching, trilateration, or Time Difference of Arrival (TDoA) / Angle of Arrival (AoA) geometric calculation methods. The core processes are mostly based on single-mode signal features and generally employ fixed sampling frequencies, static threshold judgments, and fully connected network regression outputs. Under light load conditions, these methods possess a certain degree of interpretability and computational stability, making them suitable for low-to-mid frequency environments with static scenes and significant interference intensity. However, in high-load, high-dynamic, and obstacle-dense urban environments, due to frequent non-line-of-sight (NLOS) propagation, multipath interference, and signal-to-noise ratio (SNR) abrupt changes, the localization error of traditional methods increases dramatically, sometimes exceeding 50 meters. Simultaneously, the system lacks elastic resource scheduling capabilities, easily leading to inference lag or even task loss under increased computational pressure, failing to meet the requirements for interference source tracing within seconds.

[0004] Therefore, how to accurately identify and quickly locate interference sources under heavy load conditions has become a technical problem that the industry urgently needs to solve. Summary of the Invention

[0005] This application provides a network interference localization method and apparatus based on multi-source data fusion, which is used to solve the technical problem of how to accurately identify and quickly locate interference sources under heavy load conditions.

[0006] This application provides a network interference localization method based on multi-source data fusion, including: Acquire radio frequency signal data, environmental perception data, and network topology data; Extract the frequency domain feature vector of the radio frequency signal data, the spatial feature vector of the environmental perception data, and the topological feature vector of the network topology data, and perform cross-modal alignment on the frequency domain feature vector and the spatial feature vector; Based on the real-time network load dynamic adjustment of the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector and the topological feature vector, the feature vectors are fused based on the fusion weights to obtain the fused feature tensor; The location parameter set is obtained by performing non-line-of-sight propagation correction on the fused feature tensor, and the location of the interference source is determined based on the location parameter set.

[0007] In some embodiments, acquiring radio frequency signal data, environmental perception data, and network topology data includes: Based on the timestamps of the radio frequency signal data collected by each distributed probe node, the timestamps of the environmental perception data collected by each edge computing node, and the theoretical time difference between each distributed probe node and each edge computing node, a global timestamp alignment objective function is constructed; the theoretical time difference is calculated based on geographical location and signal transmission model. The radio frequency signal data and the environmental perception data are time-aligned with the objective of minimizing the function value of the global timestamp alignment target function.

[0008] In some embodiments, acquiring radio frequency signal data, environmental perception data, and network topology data includes: A three-dimensional spatial grid is constructed based on the geographic location, and various types of data in the radio frequency signal data are mapped to each voxel of the three-dimensional spatial grid to obtain a unified structure tensor for each voxel. Based on the data source quality of various types of data, determine the trust coefficient of each voxel; The unified structure tensor of any voxel is completed based on the unified structure tensor of the neighboring voxels of any voxel and the confidence coefficient of each neighboring voxel. Based on the unified structure tensor after each voxel is completed, the completed radio frequency signal data is obtained.

[0009] In some embodiments, the cross-modal alignment of the frequency domain feature vector and the spatial feature vector includes: An encoder is constructed to map the frequency domain feature vector and the spatial feature vector to a shared latent space, and a discriminator is constructed to determine the source of the latent space features in the shared latent space; The encoder and the discriminator are jointly optimized based on an adversarial training strategy; Cross-modal alignment of the frequency domain feature vector and the spatial feature vector is performed based on the trained encoder.

[0010] In some embodiments, the dynamic adjustment of the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector, and the topological feature vector based on real-time network load includes: Construct a weight matrix consisting of the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector, and the topological feature vector; The reward function is determined based on the real-time network load and error localization feedback value, and reinforcement learning is performed based on the reward function to update the weight matrix; If the number of consecutive detections where the error localization feedback value is greater than the preset error is greater than the preset number, a population composed of multiple weight matrix individuals is established; different weight matrix individuals represent different fusion weight selection schemes. The individuals in the population are sorted by weight matrix based on their historical positioning performance, and a predetermined proportion of individuals with the highest historical positioning performance are selected as the target population. Add perturbation to the offspring individuals of the target population to obtain offspring individuals with genetic mutation; If the decrease in the error positioning feedback value corresponding to the fusion weight selection scheme of the offspring individual is greater than the preset magnitude, the fusion weight corresponding to the offspring individual is determined as a valid fusion weight.

[0011] In some embodiments, the non-line-of-sight propagation correction of the fused feature tensor to obtain the localization parameter set includes: A first correction engine is constructed; the first correction engine is used to compare the frequency domain feature vector in the fused feature tensor with the signal feature vectors of each preset non-line-of-sight scene in the multipath fingerprint database, and correct the fused feature tensor based on the multipath deviation of the matched preset non-line-of-sight scene. A second correction engine is constructed; the second correction engine is used to construct a ray projection model based on the three-dimensional environment point cloud, and to adjust the signal propagation parameters of the mid-frequency domain feature vector of the fused feature tensor based on the ray projection model; The fused feature tensor is corrected by non-line-of-sight propagation based on the first correction engine and the second correction engine to obtain the localization parameter set.

[0012] In some embodiments, determining the location of the interference source based on the positioning parameter set includes: The location parameter set is processed based on the candidate location screening model to determine the distance confidence interval; Based on real-time network load, a target inference model is selected from the first and second inference models to process the location parameters of each candidate interference source in the distance confidence interval, thereby determining the location of the interference source; the accuracy of the first inference model is greater than that of the second inference model. If the inference time of the target inference model is longer than the preset inference time, the location of the reference base station is processed based on the weighted centroid algorithm to obtain the location of the interference source.

[0013] This application provides a network interference localization device based on multi-source data fusion, comprising: The data acquisition module is used to acquire radio frequency signal data, environmental perception data, and network topology data; The feature extraction module is used to extract the frequency domain feature vector of the radio frequency signal data, the spatial feature vector of the environmental perception data, and the topological feature vector of the network topology data, and to perform cross-modal alignment on the frequency domain feature vector and the spatial feature vector. The feature fusion module is used to dynamically adjust the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector, and the topological feature vector based on the real-time network load, and to fuse each feature vector based on the fusion weights to obtain a fused feature tensor. The positioning determination module is used to perform non-line-of-sight propagation correction on the fused feature tensor to obtain a positioning parameter set, and to determine the location of the interference source based on the positioning parameter set.

[0014] This application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the network interference localization method based on multi-source data fusion.

[0015] This application provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the network interference localization method based on multi-source data fusion.

[0016] The network interference localization method and apparatus based on multi-source data fusion provided in this application overcome the shortcomings of single-source data source localization methods, such as insufficient information dimensions and susceptibility to environmental influences, by acquiring and fusing radio frequency signal data, environmental perception data, and network topology data. It ensures the effectiveness and significance of heterogeneous information fusion by performing cross-modal alignment of features from different modes. By introducing a dynamic weight adjustment mechanism based on real-time network load, the localization system can intelligently adapt to changes in network conditions and maintain high robustness under various operating conditions. Targeted non-line-of-sight propagation correction improves the localization accuracy and stability in non-line-of-sight environments. It achieves accurate identification and rapid localization of interference sources even when the network is under heavy load. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the network interference localization method based on multi-source data fusion provided in this application.

[0020] Figure 2 This is a flowchart illustrating the dynamic weighted feature fusion method provided in this application.

[0021] Figure 3 This is a flowchart illustrating the multi-source data acquisition and preprocessing method provided in this application.

[0022] Figure 4 This is a flowchart illustrating the cross-modal feature hierarchical extraction method provided in this application.

[0023] Figure 5 This is a flowchart illustrating the dual-path correction method for non-line-of-sight propagation errors provided in this application.

[0024] Figure 6 This is a flowchart illustrating the hierarchical reasoning method for anti-overload positioning calculation provided in this application.

[0025] Figure 7 This is a schematic diagram of the network interference localization device based on multi-source data fusion provided in this application.

[0026] Figure 8This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0027] 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 should fall within the scope of protection of the present application.

[0028] It should be noted that the terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps, units, or modules is not necessarily limited to those explicitly listed, but may include other steps, units, or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0029] ‌ Figure 1 This is a flowchart illustrating the network interference localization method based on multi-source data fusion provided in this application, as shown below. Figure 1 As shown, the method includes steps 110, 120, 130 and 140.

[0030] Step 110: Acquire radio frequency signal data, environmental perception data, and network topology data.

[0031] Specifically, the network interference localization method based on multi-source data fusion provided in this application can be applied to various wireless network environments, such as cellular networks (e.g., 5G, 6G), wireless local area networks (Wi-Fi), Internet of Things (IoT), satellite communication networks, and wireless sensor networks for specific purposes. The execution entity of this method can be implemented in software, such as a network interference localization program running on a computer; or it can be implemented in hardware, such as a server or a distributed computing system (e.g., a system consisting of multiple edge computing nodes and a cloud control center).

[0032] Radio frequency (RF) signal data refers to data that directly or indirectly reflects the propagation characteristics of wireless signals. This data can be acquired by multiple distributed probe nodes, base stations, monitoring stations, or any device with RF signal acquisition capabilities deployed in the network. In one optional implementation, RF signal data may include, but is not limited to: Received Signal Strength Indicator (RSSI), Reference Received Power (RSRP), Signal-to-Noise Ratio (SNR), Channel State Information (CSI), Angle of Arrival (AoA), and Time Difference of Arrival (TDoA). The purpose of acquiring this data is to quantitatively analyze the characteristics of interference sources and their impact on the network from a physical perspective of signal propagation.

[0033] Environmental perception data refers to data describing the physical spatial environmental characteristics of the location of interference sources. This data can be collected by edge computing nodes deployed at the network edge, equipped with specific sensors. In one optional implementation, environmental perception data can specifically include: temperature data collected by temperature sensors, humidity data collected by humidity sensors, 3D point cloud data generated by LiDAR scanning, 2D image or video data captured by cameras, and building outlines and material information contained in high-precision map data. The purpose of acquiring this data is to provide prior knowledge of the physical environment for subsequent signal propagation analysis, especially for the judgment and correction of non-line-of-sight propagation.

[0034] Network topology data refers to data describing the connections and geographical distribution of various nodes (such as probes, base stations, servers, etc.) in a network. This data can typically be obtained from a network management center or network operations and maintenance database. In one optional implementation, network topology data may include: the precise geographical coordinates (latitude, longitude, altitude) of each probe node and edge computing node, a logical connection diagram between nodes, and data transmission routing information. The purpose of acquiring this data is to understand the spatial distribution of data collection points at the network structure level, providing geometric constraints for location calculation.

[0035] By acquiring data from the three different dimensions mentioned above, this application is able to construct a panoramic view that includes signal data, physical environment, and network structure, laying a solid data foundation for subsequent precise positioning.

[0036] Step 120: Extract the frequency domain feature vector of the radio frequency signal data, the spatial feature vector of the environmental perception data, and the topological feature vector of the network topology data, and perform cross-modal alignment on the frequency domain feature vector and the spatial feature vector.

[0037] Specifically, after acquiring the original multi-source data, in order to facilitate subsequent fusion processing, it is necessary to extract features from different types of data and convert them into a unified mathematical representation, namely feature vectors.

[0038] Frequency domain feature vectors are extracted from radio frequency signal data. For example, a Fast Fourier Transform (FFT) or wavelet transform can be performed on the original time-domain signal to obtain its spectrum. Values ​​that characterize the signal's energy distribution, bandwidth, center frequency shift, and other properties can then be extracted from the spectrum to form a frequency domain feature vector. This vector can effectively reflect the "fingerprint" information of the interference signal in the frequency dimension.

[0039] Spatial feature vectors are extracted from environmental perception data. For example, deep learning models such as 3D convolutional neural networks can be used to process 3D point cloud data and extract feature vectors that can characterize information such as the distribution of obstacles, visibility, and orientation of reflective surfaces in space. This vector is a mathematical abstraction of the physical environment surrounding the interference source.

[0040] Topological feature vectors are extracted from network topology data. For example, the network topology can be viewed as a graph, and graph theory algorithms can be used to calculate the degree centrality, betweenness centrality, or distance between nodes, thus constructing a topological feature vector. This vector describes the relative importance and positional relationships of the data collection points within the network structure.

[0041] After extracting their respective feature vectors, the frequency domain feature vector and the spatial feature vector belong to different modes. Their feature spaces are inconsistent, and their numerical scales and physical meanings are also different, resulting in poor fusion performance when directly concatenated. Therefore, this application introduces a cross-modal alignment step.

[0042] Cross-modal alignment aims to map feature vectors from different sources and with different structures into a unified, shared latent space. Within this shared latent space, feature vectors from different modalities become comparable. For example, the position of the frequency domain feature vector of a radio frequency signal acquired in a specific physical environment (such as a street corner), after mapping, should be highly correlated with the position of its corresponding spatial feature vector describing that street corner environment after mapping. Various techniques can be used to achieve cross-modal alignment, such as more advanced models like deep learning-based autoencoders or adversarial training networks. This step ensures the effectiveness of subsequent fusion.

[0043] Step 130: Based on the real-time network load, dynamically adjust the fusion weights corresponding to the frequency domain feature vector, spatial feature vector, and topological feature vector, and fuse each feature vector based on the fusion weights to obtain the fused feature tensor.

[0044] Specifically, considering the dynamic nature of the actual network environment, the reliability and importance of different data sources vary at different times. For example, when the network communication load is low, the system has sufficient computing and transmission resources to process high-quality, high-dimensional radio frequency signal data; however, when the network load is extremely high, the radio frequency signal data may experience significant delays or packet loss, at which point its reliability decreases, and it may be necessary to rely more on relatively stable environmental and topology data.

[0045] This application first monitors the real-time network load of the system, such as the processor utilization, memory usage, and bandwidth utilization of network links for each edge computing node and probe node. Then, based on these network load indicators, it dynamically assigns different fusion weights to the aligned frequency domain feature vectors, spatial feature vectors, and topological feature vectors. After obtaining the dynamic weights, the feature vectors are weighted and fused. The fusion method can be varied, such as weighted summation. The result obtained after fusion is referred to as the fused feature tensor in this application. This tensor is a high-dimensional mathematical entity that not only aggregates information from the signal, environment, and network dimensions, but also contains the system's judgment on the reliability of different information sources based on the current operating state. It is a highly condensed and dynamically adaptable feature representation.

[0046] Figure 2 This is a flowchart illustrating the dynamic weighted feature fusion method provided in this application, as shown below. Figure 2 As shown, the corresponding fusion weight matrix can be activated according to the current network load level (light / medium / heavy load): under heavy load, the weight of the frequency domain feature vector is increased to 0.6, the weight of the spatial feature vector is reduced to 0.2, and the weight of the topology feature vector is kept at 0.2; the dynamic feedback adjustment of the weight is realized through the gated recurrent unit (GRU) and the fusion feature tensor is output.

[0047] Step 140: Perform non-line-of-sight propagation correction on the fused feature tensor to obtain the localization parameter set, and determine the location of the interference source based on the localization parameter set.

[0048] Specifically, although the fused feature tensor contains rich information, the part originating from the radio frequency signal may still be contaminated by non-line-of-sight propagation. That is, the signal does not reach the receiving probe in a straight line, but is reflected and diffracted by buildings, which causes deviations in measurement parameters (such as distance and angle). This is the main cause of positioning errors.

[0049] First, non-line-of-sight (NLS) propagation correction is performed on the fused feature tensor. This correction process aims to identify and quantify the bias caused by NLS propagation effects and correct the corresponding feature components in the fused feature tensor. Since the fused feature tensor already contains environmental information (from spatial feature vectors), this provides a basis for judging and correcting NLS propagation. For example, the correction model can determine the possible presence of obstacles in the current signal path based on the spatial features in the tensor, and estimate the error introduced by NLS propagation by combining changes in frequency domain features (such as additional time delay and energy attenuation), and then perform inverse compensation on the features.

[0050] After correction, a purer set of positioning parameters that better reflects the actual propagation path information is obtained. This set of positioning parameters is not a direct geographic coordinate set, but a set of purified and optimized intermediate parameters used for the final positioning calculation.

[0051] Finally, based on this set of localization parameters, one or more localization algorithms are used to determine the final location of the interference source. For example, the localization parameter set can be input into a trained deep learning model that can directly regress coordinate positions.

[0052] The network interference localization method based on multi-source data fusion provided in this application overcomes the shortcomings of single-source data source localization methods, such as insufficient information dimensions and susceptibility to environmental influences, by acquiring and fusing radio frequency signal data, environmental perception data, and network topology data. It ensures the effectiveness and significance of heterogeneous information fusion by performing cross-modal alignment of features from different modes. By introducing a dynamic weight adjustment mechanism based on real-time network load, the localization system can intelligently adapt to changes in network conditions and maintain high robustness under various operating conditions. Targeted non-line-of-sight propagation correction improves localization accuracy and stability in non-line-of-sight environments. It achieves accurate identification and rapid localization of interference sources even when the network is under heavy load.

[0053] It should be noted that each implementation method of this application can be freely combined, rearranged, or executed individually, and does not need to rely on or depend on a fixed execution order.

[0054] In some embodiments, acquiring radio frequency signal data, environmental awareness data, and network topology data includes: A global timestamp alignment objective function is constructed based on the timestamps of radio frequency signal data collected by each distributed probe node, the timestamps of environmental perception data collected by each edge computing node, and the theoretical time difference between each distributed probe node and each edge computing node; the theoretical time difference is calculated based on geographical location and signal transmission model. The goal is to minimize the function value of the global timestamp alignment objective function to perform time alignment on radio frequency signal data and environmental perception data.

[0055] Specifically, Figure 3 This is a flowchart illustrating the multi-source data acquisition and preprocessing method provided in this application, as shown below. Figure 3 As shown, in the real-time acquisition and preprocessing stage of multi-source data, to ensure the stability and consistency of subsequent cross-modal feature extraction and fusion, the system deploys a high-density distributed probe node array. The distributed probe nodes are aligned with high precision timing via a synchronous clock protocol. Each node periodically acquires radio frequency signal data (including radio frequency signal strength, time difference of arrival, and angle of arrival). Simultaneously, environmental sensors are accessed through edge computing nodes to obtain temperature and humidity parameters under current climate conditions, and local obstacle point cloud maps and network topology data are constructed. To ensure the spatiotemporal consistency among these heterogeneous data, the system performs joint timestamp correction and normalization transformation to eliminate location information distortion caused by device clock drift and sampling rate differences. The core lies in designing a robust global timestamp alignment objective function, which globally minimizes the time residual between data from each node, achieving a unified mapping of data in the time domain.

[0056] Global timestamp alignment target function It can be represented as: .

[0057] in, This represents the number of distributed probe nodes. To calculate the number of nodes at the edge, For the first The timestamps of the radio frequency signal data collected by each distributed probe node. For the first The timestamps of environmental perception data collected by each edge computing node For the first The distributed probe node and the first The theoretical time difference between edge computing nodes This is a clock stability weighting factor between nodes, and its value is calculated in reverse from the standard deviation of the node's historical drift. The larger the value, the more the time error of the pair needs to be corrected.

[0058] By minimizing the global timestamp alignment objective function, a synchronous mapping of all probe data sources on a unified reference time axis can be obtained, significantly reducing cross-source time misalignment caused by clock drift. This ensures the relative consistency of each modal data in event occurrence when constructing spatiotemporally correlated data streams, enabling time alignment of RF signal data and environmental sensing data.

[0059] The network interference localization method based on multi-source data fusion provided in this application, by constructing and optimizing a global timestamp alignment objective function, can accurately estimate and compensate for clock deviations between distributed heterogeneous nodes, achieving high-precision alignment of multi-source data in the time dimension. This provides a time-consistent data foundation for subsequent multimodal feature fusion, avoiding fusion errors caused by data timestamp mismatches from the source, and ensuring that the entire localization method ultimately achieves high accuracy.

[0060] In some embodiments, acquiring radio frequency signal data, environmental awareness data, and network topology data includes: A three-dimensional spatial grid is constructed based on geographical location, and various types of data in radio frequency signal data are mapped to each voxel of the three-dimensional spatial grid to obtain a unified structure tensor for each voxel. Based on the data source quality of various types of data, determine the trust coefficient of each voxel; Based on the unified structure tensor of the neighborhood voxels of any voxel, and the confidence coefficient of each neighborhood voxel, the unified structure tensor of any voxel is completed. Based on the unified structure tensor after each voxel is completed, the completed radio frequency signal data is obtained.

[0061] Specifically, the time-aligned data also needs to undergo spatial normalization mapping and scale standardization. A three-dimensional spatial grid can be constructed based on geographic coordinates representing geographical locations, mapping various data types in the radio frequency signal data (such as radio frequency signal strength, time difference of arrival, and angle of arrival) into a unified structure tensor. ,in Index corresponding to the spatial dimension location, This is a time step. The unified structure tensor records the set of multimodal raw observations of a unit spatial voxel at this time step. A voxel is a volume element in a three-dimensional spatial grid.

[0062] Since the environmental sensor data and base station topology map are sparsely sampled, a Bayesian interpolation-based completion mechanism is needed to fill in the missing areas. Specifically, the following confidence-weighted completion function is constructed: .

[0063] in, This is the unified structure tensor after the current voxel completion. The set of voxels containing observed data within the current voxel's neighborhood. The confidence coefficient is calculated based on the quality of the data source (signal-to-noise ratio, drift, sensor temperature stability coefficient). Neighboring voxels At that time step The original value.

[0064] This mechanism not only enables reliable completion of discontinuous regions in the spatial data stream, but also provides stable input for subsequent steps.

[0065] The network interference localization method based on multi-source data fusion provided in this application transforms the original, disordered, and sparse radio frequency data points into a structured data field by constructing a three-dimensional spatial grid and mapping the data into a unified structure tensor. By introducing and utilizing a confidence coefficient, the data completion process can intelligently focus on high-quality source data, suppressing noise interference that may be caused by low-quality data. It can generate a dense, complete, and reliable radio frequency environment data, effectively overcoming the problem of information loss caused by sparse probe deployment, and providing a solid data guarantee for achieving high-precision, full-coverage interference localization.

[0066] Through time alignment and weighted completion in the above embodiments, a high-resolution, spatiotemporally consistent, and multimodal complementary raw data tensor stream can be constructed. This provides a highly structured input for the cross-modal feature hierarchical extraction stage and maintains the spatiotemporal robustness and processing efficiency of various data sources under dynamic network load scenarios. During system operation, the global timestamp alignment objective function minimization process is performed once per second, and synchronization correction is completed in the base station-level distributed edge computing unit with a latency of less than 3 milliseconds. This ensures that the spatiotemporal data fusion accuracy is maintained within an error range of 1.5 meters, effectively guaranteeing the input consistency and stability of subsequent feature hierarchical extraction and dynamic weighted fusion mechanisms.

[0067] In some embodiments, Figure 4 This is a flowchart illustrating the cross-modal feature hierarchical extraction method provided in this application, as shown below. Figure 4 As shown, in the cross-modal feature layer extraction stage, the system faces challenges such as inconsistent dimensions, large distribution differences, and dynamic fluctuations in signal-to-noise ratio among the three types of data: radio frequency, environment, and topology. Therefore, the core task of this stage is to map the heterogeneous raw data into a comparable and fusionable unified feature space representation, and to provide semantically consistent and scale-controllable input for the dynamic weighted fusion mechanism.

[0068] For point cloud data in environmental perception data, the system constructs a 3D adjacency graph. , where the set of nodes Representing spatial obstacle points, edge set The graph convolutional network is built based on Euclidean distance and occlusion probability. It takes node height, material reflectivity, and occlusion frequency as input features, and extracts encoded spatial feature vectors through multiple convolutional layers. (A real-valued space of dimension d) is used to characterize the distribution of non-line-of-sight propagation risk within a local space. A three-dimensional spatial attenuation model can be constructed using point cloud clustering algorithms.

[0069] For radio frequency signal data, the time-spectrum generated after short-time Fourier transform exhibits typical interference mode distribution characteristics. After inputting into the ResNet-18 backbone network, it outputs a frequency domain feature vector. This vector characterizes the non-stationary disturbance properties of the channel within the current time window. High-frequency components can also be decomposed using wavelet transform to capture transient interference features.

[0070] For network topology data, considering its natural graph structure and dynamic characteristics, the system constructs a directed graph based on historical load transfer sequences between base stations. In this model, node B represents the base station entity, and the edge weights are jointly determined by the signaling transfer frequency and link impedance per unit time. The graph attention network uses the load time series of each base station and its neighbor relationships as input to calculate its influence factor on neighboring nodes. and output the topological feature vector. This feature vector is used to characterize the ability of base station load to regulate interference propagation paths within a local area. During the fusion phase, this feature vector will participate in the dynamic weighted calculation as an influence factor matrix, supporting the expression of the coupling relationship between interference intensity and base station topology.

[0071] The aforementioned features can be extracted in parallel. For example, Graph Convolutional Networks (GCN), Residual Networks (ResNet), and Domain Adversarial Neural Networks (DANN) can be used to perform spatiotemporal consistency compression on features of different modalities such as radio frequency, point cloud, and topology. Before fusion, the cosine similarity of the vectors is increased to above 0.85. At the same time, feature quality monitoring and dynamic compensation strategies are introduced. When the data signal-to-noise ratio decreases, high-quality vectors are retrieved and supplemented through the historical scene database. This reduces the distribution deviation before and after feature fusion by 62%, ensuring effective information integration and supporting stable operation of subsequent inference.

[0072] In some embodiments, cross-modal alignment of frequency domain feature vectors and spatial feature vectors includes: An encoder is constructed to map frequency domain feature vectors and spatial feature vectors to a shared latent space, and a discriminator is constructed to determine the source of latent space features in the shared latent space; Joint optimization of encoder and discriminator based on adversarial training strategy; Cross-modal alignment of frequency domain feature vectors and spatial feature vectors is performed based on the trained encoder.

[0073] Specifically, to achieve alignment of the two types of features in high-dimensional space, the system introduces a mapping alignment mechanism based on domain adversarial training. An encoder is then constructed. Map any modality feature to a shared latent space and train a discriminator. Distinguish source modalities. Under optimal training conditions, make it difficult for the encoder to allow the discriminator to recognize the feature modalities, thereby achieving domain alignment. This process can be implemented using a domain adversarial neural network, with a loss function... for: .

[0074] in, and Let represent the distributions of radio frequency signal data and environmental perception data, respectively. The goal is to minimize this loss so that the two modalities exhibit similar statistical characteristics after mapping. Through this adversarial mechanism, the encoder learns modality-independent feature representations, enabling... and The cosine similarity in the shared latent space approaches 1, and the experimentally measured average cosine similarity reaches 0.87, which meets the consistency requirements for subsequent fusion.

[0075] Once the adversarial training process converges and is complete, the encoder possesses powerful modality confusion and feature extraction capabilities. During network interference localization, newly acquired frequency domain feature vectors and spatial feature vectors simply need to be input into this pre-trained encoder. The encoder outputs their corresponding aligned latent space features in the shared latent space. Since these output feature vectors reside in the same feature space, have the same dimension, and similar numerical distributions, they can be directly fused, for example, through concatenation or weighted summation, to form a fused feature tensor.

[0076] The network interference localization method based on multi-source data fusion provided in this application employs a cross-modal alignment method based on adversarial training. This method automatically learns the intrinsic relationships between heterogeneous features in an end-to-end, data-driven manner and maps them to a unified feature space. This avoids the complex and inefficient design of manual feature alignment rules, forcing the encoder to learn modality-independent, more fundamental feature representations. This significantly improves the effectiveness and robustness of subsequent feature fusion, providing key technical support for generating high-quality fused feature tensors and ultimately achieving high-precision interference localization.

[0077] In some embodiments, to further enhance the system's feature stability under weak signal-to-noise ratio (SNR) conditions, a feature quality evaluation threshold is set. When the SNR of the original input of any modality is below 15 dB, feature compensation is immediately triggered. The system retrieves the candidate vector with the smallest Euclidean distance to the current feature tensor from the historical feature database. And construct the feature replacement function: .

[0078] in, For the replaced features, The original features of the current frame data, Instead of intensity coefficients, SNR represents the actual signal-to-noise ratio of the current frame data. The lower the SNR, the higher the weighting of historical features, ensuring that even under extreme interference scenarios, the feature tensor retains discriminative power and maintains intermodal structural stability. This mechanism ultimately achieves unified scale representation and spatial alignment of heterogeneous features before dynamic weighting. In actual deployment, the system reduces the feature distribution offset before fusion by 62% under various typical interference scenarios, effectively supporting the high-precision operation of the subsequent dynamic fusion weight matrix and non-line-of-sight propagation error correction module.

[0079] In some embodiments, dynamically adjusting the fusion weights corresponding to the frequency domain feature vector, spatial feature vector, and topological feature vector based on real-time network load includes: Construct a weight matrix consisting of fusion weights corresponding to frequency domain feature vectors, spatial feature vectors, and topological feature vectors; The reward function is determined based on real-time network load and error localization feedback values, and reinforcement learning is performed based on the reward function to update the weight matrix; When the number of consecutive detections with the error localization feedback value exceeding the preset error exceeds the preset number, a population composed of multiple weight matrix individuals is established; different weight matrix individuals represent different fusion weight selection schemes. The individuals in the population are ranked according to their historical positioning performance, and a predetermined proportion of individuals with the highest historical positioning performance are selected as the target population. Add perturbations to the offspring individuals of the target population to obtain offspring individuals with genetic mutations; If the decrease in the error positioning feedback value corresponding to the fusion weight selection scheme of the offspring individual is greater than the preset magnitude, the fusion weight corresponding to the offspring individual will be determined as the effective fusion weight.

[0080] Specifically, in dynamic weighted feature fusion, to enhance robustness and improve positioning accuracy under high-load interference environments, the system designs a load-aware weighted evolution algorithm. In the dynamic weighted feature fusion stage, it not only considers the differences in representation capabilities between modes but also actively senses and adapts to changes in network operating status. In the initial stage, the weight vectors for the radio frequency feature vector, environmental feature vector, and topology feature vector are set as follows: Their sum remains constant, and the fusion tensor is defined as follows: The updates to the weights of each group follow a reinforcement learning policy with a penalty term, and its reward function... Based on the current network load level With positioning error feedback value Joint drivers are defined as follows: in, Let be the positioning error feedback value of the t-th positioning task. To determine the current network load level based on real-time network load, This is a balancing parameter. The reward function serves as the objective function for the reinforcement learning policy, encouraging smooth evolution of weights under low load conditions and enhancing sensitivity to radio frequency weights under high load conditions to ensure that anti-interference capability does not degrade. The weight matrix evolution process uses the policy gradient method to maximize the expected value of the reward function. The training process is guided by the feedback error from the gated recurrent unit (GRU) in the localization module.

[0081] If the error positioning feedback value is greater than the preset error and the number of consecutive detections exceeds the preset number (e.g., three consecutive positioning deviations are detected), then... In cases where the distance exceeds 20 meters, the genetic mutation mechanism of the weight matrix is ​​immediately triggered. During this process, the top 30% of individuals with the highest current performance ranking (based on historical positioning performance ranking) are selected as the elite population. A mutation operator is applied to the offspring individuals, and the weight perturbation rule is as follows: .

[0082] in, The weight matrix represents the first... One portion, For a uniform disturbance in the interval [-0.1, +0.1], The function is used to ensure the validity of the weights and renormalize them. The new generation of weights, after mutation, will be applied to historical datasets for batch positioning error simulation. Only when their average error decreases by more than 5% compared to the previous version will they be considered valid weights and written into the production weight buffer pool.

[0083] The entire evolutionary strategy balances responsiveness and stability. In resource-constrained scenarios, it avoids full model retraining and constructs a lightweight yet efficient feature fusion weight iteration mechanism. In actual deployment, this strategy significantly improves the convergence stability and response speed of the system under high interference and high load conditions, reducing the 90th quantile positioning error from 26.7 meters to 13.9 meters. The system can continuously maintain the dynamic optimal solution distribution of the fusion weight structure during network state abrupt changes.

[0084] The network interference localization method based on multi-source data fusion provided in this application introduces a dual mechanism of reinforcement learning and genetic perturbation, establishes a feature fusion weight evolution strategy oriented to task feedback, utilizes real-time error feedback and historical task performance to drive the adaptive update of the weight matrix, and combines an error threshold-driven mutation and verification mechanism to improve the adaptability to environmental disturbances and data quality degradation. The average localization accuracy is improved by 8.4% after fusion, while reducing the risk of long-term weight overfitting.

[0085] In some embodiments, a localization parameter set is obtained by performing non-line-of-sight propagation correction on the fused feature tensor, including: Construct a first correction engine; the first correction engine is used to compare the frequency domain feature vector in the fused feature tensor with the signal feature vectors of each preset non-line-of-sight scene in the multipath fingerprint database, and correct the fused feature tensor based on the multipath deviation of the matched preset non-line-of-sight scene. A second correction engine is constructed; the second correction engine is used to construct a ray projection model based on the 3D environment point cloud, and to adjust the signal propagation parameters of the mid-frequency domain feature vector of the fused feature tensor based on the ray projection model; The fused feature tensor is corrected by non-line-of-sight propagation based on the first and second correction engines to obtain the localization parameter set.

[0086] Specifically, Figure 5 This is a flowchart illustrating the dual-path correction method for non-line-of-sight propagation errors provided in this application, as shown below. Figure 5 As shown, in the joint correction of non-line-of-sight propagation error, in order to effectively compensate for the multipath deviation caused by obstructions during radio frequency propagation, the system introduces an adaptive ray projection and reflection path modeling mechanism to ensure that the positioning parameters still have interpretability and optimizability in complex environments.

[0087] Two correction engines can be established to perform dual-path non-line-of-sight propagation correction on the fused feature tensor.

[0088] The first correction engine is used to compare the frequency domain feature vector in the fused feature tensor with the signal feature vectors of each preset non-line-of-sight scene in the multipath fingerprint database, and correct the fused feature tensor based on the multipath deviation of the matched preset non-line-of-sight scene.

[0089] The online update of the multipath fingerprint database is triggered by a confidence score threshold of 0.7. When the confidence score of the final location solution is lower than this value, the on-site inspection process is immediately scheduled, and a drone equipped with a lidar is deployed to perform high-precision point cloud scanning of the current target area. The feature attribution labeling of new scenes is performed through semi-supervised contrastive learning, so that the feature sample set in the fingerprint database maintains dynamic expansion and scene adaptability.

[0090] The second correction engine is used to construct a ray projection model based on the 3D environment point cloud, and adjust the signal propagation parameters of the frequency domain feature vector in the fused feature tensor based on the ray projection model.

[0091] To further quantify the probability of path reachability under line-of-sight conditions, the system constructs an improved ray casting model based on 3D environmental point clouds and defines a cumulative occlusion attenuation function for each path. For the first The projection angle of each ray path. Using the candidate location of the interference source as the emission center, detection is performed in 32 uniformly distributed directions in space. Assuming the total path length is RRR, the attenuation value of each ray is defined by the following formula: .

[0092] in, For the first The number of obstacles on the ray, Indicates the first The material category of the obstruction. The unit reflection attenuation coefficient for the corresponding material is retrieved from the reflection coefficient table for the seven types of materials. Let be the radial projection width of the obstacle along the ray direction. This is the environmental attenuation power factor. If there exists a direction that satisfies... , If an occlusion threshold is set based on experience, then the direction is determined to be a non-line-of-sight path, and the secondary reflection compensation module is activated.

[0093] The signal path is reconstructed by constructing a dynamic multipath propagation function, and the line-of-sight probability is mapped to a correction factor. Used to adjust the propagation delay parameters in the Time Difference of Arrival (TDoA) and Angle of Arrival (AoA): .

[0094] in, For mapping smoothing factor, The occlusion threshold set based on experience, This is the attenuation value of the most severely occluded path in the current frame. This factor continuously models the path line-of-sight probability in a sigmoid form, and is ultimately introduced into a dynamic feedback module. By adjusting the signal propagation parameters through gain correction, the model can adapt to the reflection characteristics of various complex architectural environments in real time.

[0095] By combining the above correction strategies, the system's positioning performance in heavily obscured areas has been significantly improved. Compared with traditional path estimation methods based solely on geometric projection, the average positioning error in non-line-of-sight propagation scenarios has decreased from 35.2 meters to 7.8 meters, and the confidence score has increased by an average of 16.3%, effectively enhancing the overall system's practical stability and positioning consistency in complex geographical environments.

[0096] The network interference localization method based on multi-source data fusion provided in this application dynamically reconstructs complex propagation paths using ray projection and multipath compensation models. It introduces material reflection coefficients and point cloud obstacle information to form a real-space propagation mapping. Combined with an improved line-of-sight probability model, this significantly reduces localization errors in non-line-of-sight propagation scenarios. Furthermore, by combining semi-supervised learning and UAV-assisted point cloud sampling, the multipath fingerprint database is continuously expanded, and scene matching accuracy is improved through material reflection models. This enhances the system's adaptability to new environments (building modifications, temporary facility construction, etc.). Experimental results show that the dynamic fingerprint database update mechanism can improve scene matching confidence by 23% and reduce the sample size required for model cold start.

[0097] In some embodiments, determining the location of the interference source based on a set of positioning parameters includes: The location parameter set is processed based on the candidate location screening model to determine the distance confidence interval; Based on real-time network load, the target inference model is selected from the first and second inference models to process the location parameters of each candidate interference source in the distance confidence interval and determine the location of the interference source; the accuracy of the first inference model is greater than that of the second inference model. When the inference time of the target inference model exceeds the preset inference time, the location of the reference base station is processed based on the weighted centroid algorithm to obtain the location of the interference source.

[0098] Specifically, Figure 6 This is a flowchart illustrating the hierarchical reasoning method for overload positioning calculation provided in this application, as shown below. Figure 6 As shown, in the overload positioning solution stage, The system introduces hierarchical inference and dynamic circuit breaker mechanisms to construct a robust decision path for complex edge environments, ensuring high confidence and real-time computational capabilities even under resource-constrained conditions. The overall structure adopts a model-cascaded architecture, with the first level being a candidate location selection model. This model employs a fast filter based on an improved XGBoost library (an optimized distributed gradient boosting library). Input features include RSSI intensity interval distribution, base station channel load density, dynamic influence factor matrix features, and line-of-sight confidence scores. The goal is to eliminate significantly impossible regions from the initial location candidate space. The filter outputs a distance confidence interval. The confidence score is used to constrain the input boundary of the second-level model. Statistically, this module can eliminate an average of 70.6% of candidate locations and compress the input dimension of the second-level model to 28.3% of the original.

[0099] The second level is a conditionally configurable deep inference module that automatically switches the neural network architecture based on system load. It can employ a first inference model and a second inference model. The first inference model can be a 4-layer lightweight convolutional network (MobileNetV3), and the second inference model can be a 2-layer binary neural network (XNOR-BNN). Under light network load, the 4-layer MobileNetV3 network is activated to obtain high-precision nested features; under heavy network load, the 2-layer XNOR-BNN network is activated to reduce multiply-accumulate operations.

[0100] To avoid response delays caused by nonlinear accumulation during peak periods, the system employs a computational circuit breaker mechanism, monitoring each batch of requests with a preset inference duration of 50 milliseconds. If an inference time exceeds the preset duration, the system automatically switches to approximate solution mode, invoking the weighted centroid algorithm based on geometric constraints, defined as follows: , .

[0101] in, Indicates the predicted position of the current frame. For the first The known coordinates of the reference base stations, For its corresponding distance estimation value, For the calculated weighting factors, To eliminate the stabilizing parameters caused by weight explosion due to local minima, this geometric model can dynamically construct centroid trajectories based on historical data without training, maintaining high computational efficiency. In practice, it completes one inference cycle within 20 milliseconds. The reference base station can serve as a source of interference at historical moments.

[0102] To achieve seamless linkage between network architecture and load status, the system adopts an improved state triggering function. ,in This is the estimated load value for the current batch of models. This represents the average inference time of the previous cycle. The specific switching strategy is as follows: .

[0103] in These are the load and delay control thresholds set for the system, respectively. The output directly affects the model selection and accuracy strategy of the inference path in the current frame. Experimental results show that this mechanism can maintain a request response time of less than 100ms in 99.5% of cases under high concurrency. Its inference accuracy is 4.3% better than the single ResNet scheme under light load, and can still control the average error within 11.2 meters under heavy load, achieving the optimal balance between inference efficiency and performance.

[0104] The network interference localization method based on multi-source data fusion provided in this application introduces a hierarchical model cascading and inference circuit breaking mechanism. After quickly eliminating candidate regions in the decision tree, it dynamically selects a combination path of MobileNetV3 and binary network based on system load. In the case of inference timeout, it automatically degenerates into a weighted centroid algorithm, ensuring that 99.5% of requests are responded to within 100 milliseconds. It maintains business continuity even when edge resources are limited, and is particularly suitable for dense deployment or emergency response scenarios.

[0105] In some embodiments, the system constructs a digital twin centered on resource scheduling, mirroring key indicators such as GPU memory status, CPU thread utilization, and wireless link bandwidth in edge nodes to form a dynamic multi-dimensional resource situation map. This map is then input into a two-layer Long Short-Term Memory (LSTM) network model in a time-series manner to predict resource trends within the next 5 seconds. The LSTM structure integrates state update logic from the input gate, forget gate, and output gate, outputting a resource state prediction vector. Each dimension represents the predicted utilization rate of the corresponding resource. When the predicted value of any dimension... satisfy In such cases, the system immediately triggers a three-level response strategy to ensure the continuous operation of core computing tasks.

[0106] The core logic of resource regulation is based on a dynamic response discriminant function. To establish a strategy hierarchy, the following definitions are provided: .

[0107] in =0.85、 =0.90、 =0.95 corresponds to the first, second, and third level response thresholds, respectively. This indicates that the probe node sampling rate has dropped to its original value. 50%, This reduces the frequency of data generation at the edge. Disable the high-precision power consumption mode in the environmental sensor and switch to a low-frequency configuration to reduce the power consumption of the sensing layer. Enable a location result caching mechanism to return historical location results with a cosine similarity ≥ 0.92 to the current environment description vector in the L2 cache, thus avoiding redundant reasoning.

[0108] To maintain system continuity under high-pressure conditions, a computing power recovery mechanism and a resource reallocation scheduler are introduced, by constructing a resource recovery pool. Priority mapping function Enables dynamic migration of computing power. The priority function is based on the service intensity of the cell. User density and historical positioning accuracy scores Joint definition: .

[0109] in , , The weighting coefficients, learned through evolutionary strategy training, ensure that resources are preferentially allocated to computing units with higher load risk. The final scheduler, based on the maximum matching criterion, matches idle threads, caches, and bandwidth segments in the resource pool to those with the maximum matching criterion. The value of the cells forms an elastic load balancing mechanism across base station clusters. After the overall strategy was deployed, in multi-scenario tests, the location service response failure rate decreased by 71%, and the average energy consumption decreased by 12.4%, providing structural support for high availability and persistent computing of the system.

[0110] The network interference localization method based on multi-source data fusion provided in this application constructs a multi-source resource situation mirror and fuses it with an LSTM prediction model. It identifies resource pressure peaks 5 seconds in advance and triggers a three-level response strategy. At the same time, it realizes the redistribution of computing power across base station groups through a resource recycling pool and priority scheduling function, so that core tasks can still be guaranteed to run stably under high concurrency conditions. Resource utilization is improved by 18%, latency request rate is reduced by 62%, and localization task collaboration is supported in large-scale deployment scenarios.

[0111] Based on the above embodiments, the method for accurate location of high-load interference in operator networks that integrates multi-source data provided in this application includes the following timing steps: Step 1: Real-time acquisition and preprocessing of multi-source data: Radio frequency signal strength (RSSI), time difference of arrival (TDOA), and angle of arrival (AOA) data are synchronously acquired through distributed probe nodes. At the same time, environmental sensors are connected to obtain temperature and humidity, obstacle distribution, and base station topology data. Timestamp alignment and normalization processing are performed on heterogeneous data to eliminate clock drift between devices and generate spatiotemporally correlated raw data streams.

[0112] Step 2, Cross-modal feature layered extraction: The preprocessed data is input into the parallel feature extraction channels: The first channel decomposes the high-frequency components of the radio frequency signal through wavelet transform to capture transient interference features; the second channel uses a point cloud clustering algorithm to process the coordinates of environmental obstacles and construct a three-dimensional spatial attenuation model; the third channel parses the base station topology relationship to generate network load weight coefficients.

[0113] Step 3: Dynamic weighted feature fusion: Activate the corresponding fusion weight matrix based on the current network load level (light / medium / heavy load): In the heavy load state, increase the weight of radio frequency signal features to 0.6, reduce the weight of environmental data to 0.2, and keep the weight of topology data at 0.2; realize the dynamic feedback adjustment of weights through gated cyclic unit (GRU) and output fusion feature tensor.

[0114] Step 4, Joint Correction of NLOS Errors: Input the fused feature tensor into the dual-path correction engine: Path 1 uses multi-path fingerprint matching technology to compare real-time signal features with the pre-stored NLOS scene fingerprint database; Path 2 performs line-of-sight probability calculation, dynamically adjusts the TDOA / AOA compensation coefficients based on obstacle distribution density, and outputs the location parameter set after error correction.

[0115] Step 5, Overload-resistant localization calculation: Input the corrected localization parameter set into the lightweight convolution module, which contains 3 layers of depthwise separable convolution and channel attention mechanism. While maintaining localization accuracy, the computational complexity is reduced to 18% of that of a fully connected network, and the geographical coordinates of the interference source and confidence score are output.

[0116] Step 6, adaptive resource release: After completing the positioning, immediately release the GPU memory resources and dynamically adjust the probe sampling frequency according to the predicted data flow value of the next cycle, so that the system resource utilization rate is always lower than the preset 70% threshold.

[0117] The network interference localization method based on multi-source data fusion provided in this application not only realizes the whole process of adaptive closed-loop control from data acquisition to computing resource allocation, but also has a high environmental perception capability and task feedback adjustment mechanism, which significantly improves the positioning performance, stability and response capability of the system in complex, dynamic and high-density environments, and provides key support for high-precision positioning in fields such as urban intelligent transportation, emergency rescue, intelligent manufacturing and logistics distribution.

[0118] The apparatus provided in the embodiments of this application is described below. The apparatus described below can be referred to in correspondence with the method described above.

[0119] Figure 7 This is a schematic diagram of the network interference localization device based on multi-source data fusion provided in this application, as shown below. Figure 7 As shown, the device includes: Data acquisition module 710 is used to acquire radio frequency signal data, environmental perception data and network topology data; The feature extraction module 720 is used to extract the frequency domain feature vector of radio frequency signal data, the spatial feature vector of environmental perception data, and the topological feature vector of network topology data, and to perform cross-modal alignment on the frequency domain feature vector and the spatial feature vector. The feature fusion module 730 is used to dynamically adjust the fusion weights corresponding to the frequency domain feature vector, spatial feature vector and topological feature vector based on the real-time network load, and to fuse each feature vector based on the fusion weights to obtain the fused feature tensor. The localization determination module 740 is used to perform non-line-of-sight propagation correction on the fused feature tensor to obtain a localization parameter set, and to determine the location of the interference source based on the localization parameter set.

[0120] The network interference localization device based on multi-source data fusion provided in this application overcomes the shortcomings of single-source data source localization methods, such as insufficient information dimensions and susceptibility to environmental influences, by acquiring and fusing radio frequency signal data, environmental perception data, and network topology data. It ensures the effectiveness and significance of heterogeneous information fusion by performing cross-modal alignment of features from different modes. By introducing a dynamic weight adjustment mechanism based on real-time network load, the localization system can intelligently adapt to changes in network conditions and maintain high robustness under various operating conditions. Targeted non-line-of-sight propagation correction improves the localization accuracy and stability in non-line-of-sight environments. It achieves accurate identification and rapid localization of interference sources even when the network is under heavy load.

[0121] Figure 8 This is a schematic diagram of the structure of the electronic device provided in this application, such as... Figure 8 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communications bus 840, wherein the processor, communications interface, and memory communicate with each other via the communications bus. The processor can invoke logical commands stored in the memory to execute the methods described in the above embodiments, for example: Acquire radio frequency (RF) signal data, environmental perception data, and network topology data; extract frequency domain feature vectors from RF signal data, spatial feature vectors from environmental perception data, and topological feature vectors from network topology data, and perform cross-modal alignment on the frequency domain and spatial feature vectors; dynamically adjust the fusion weights corresponding to the frequency domain, spatial, and topological feature vectors based on real-time network load, and fuse each feature vector based on the fusion weights to obtain a fused feature tensor; perform non-line-of-sight propagation correction on the fused feature tensor to obtain a location parameter set, and determine the location of the interference source based on the location parameter set.

[0122] Furthermore, the logical commands in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several commands to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0123] The processor in the electronic device provided in this application embodiment can call logical instructions in the memory to implement the above method. Its specific implementation method is the same as the aforementioned method implementation method and can achieve the same beneficial effect, which will not be repeated here.

[0124] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0125] The specific implementation method is the same as the aforementioned method implementation method and can achieve the same beneficial effects, so it will not be repeated here.

[0126] This application provides a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0127] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown 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 this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0128] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A network interference localization method based on multi-source data fusion, characterized in that, include: Acquire radio frequency signal data, environmental perception data, and network topology data; Extract the frequency domain feature vector of the radio frequency signal data, the spatial feature vector of the environmental perception data, and the topological feature vector of the network topology data, and perform cross-modal alignment on the frequency domain feature vector and the spatial feature vector; Based on the real-time network load dynamic adjustment of the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector and the topological feature vector, the feature vectors are fused based on the fusion weights to obtain the fused feature tensor; The location parameter set is obtained by performing non-line-of-sight propagation correction on the fused feature tensor, and the location of the interference source is determined based on the location parameter set; The dynamic adjustment of the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector, and the topological feature vector based on real-time network load includes: Construct a weight matrix consisting of the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector, and the topological feature vector; The reward function is determined based on the real-time network load and error localization feedback value, and reinforcement learning is performed based on the reward function to update the weight matrix; If the number of consecutive detections where the error localization feedback value is greater than the preset error is greater than the preset number, a population composed of multiple weight matrix individuals is established; different weight matrix individuals represent different fusion weight selection schemes. The individuals in the population are sorted by weight matrix based on their historical positioning performance, and a predetermined proportion of individuals with the highest historical positioning performance are selected as the target population. Add perturbation to the offspring individuals of the target population to obtain offspring individuals with genetic mutation; If the decrease in the error positioning feedback value corresponding to the fusion weight selection scheme of the offspring individual is greater than the preset magnitude, the fusion weight corresponding to the offspring individual is determined as an effective fusion weight. The non-line-of-sight propagation correction of the fused feature tensor to obtain the localization parameter set includes: A first correction engine is constructed; the first correction engine is used to compare the frequency domain feature vector in the fused feature tensor with the signal feature vectors of each preset non-line-of-sight scene in the multipath fingerprint database, and correct the fused feature tensor based on the multipath deviation of the matched preset non-line-of-sight scene. A second correction engine is constructed; the second correction engine is used to construct a ray projection model based on the three-dimensional environment point cloud, and to adjust the signal propagation parameters of the mid-frequency domain feature vector of the fused feature tensor based on the ray projection model; The fused feature tensor is corrected by non-line-of-sight propagation based on the first correction engine and the second correction engine to obtain the localization parameter set.

2. The network interference localization method based on multi-source data fusion according to claim 1, characterized in that, The acquisition of radio frequency signal data, environmental perception data, and network topology data includes: Based on the timestamps of the radio frequency signal data collected by each distributed probe node, the timestamps of the environmental perception data collected by each edge computing node, and the theoretical time difference between each distributed probe node and each edge computing node, a global timestamp alignment objective function is constructed; the theoretical time difference is calculated based on geographical location and signal transmission model. The radio frequency signal data and the environmental perception data are time-aligned with the objective of minimizing the function value of the global timestamp alignment target function.

3. The network interference localization method based on multi-source data fusion according to claim 1, characterized in that, The acquisition of radio frequency signal data, environmental perception data, and network topology data includes: A three-dimensional spatial grid is constructed based on the geographic location, and various types of data in the radio frequency signal data are mapped to each voxel of the three-dimensional spatial grid to obtain a unified structure tensor for each voxel. Based on the data source quality of various types of data, determine the trust coefficient of each voxel; The unified structure tensor of any voxel is completed based on the unified structure tensor of the neighboring voxels of any voxel and the confidence coefficient of each neighboring voxel. Based on the unified structure tensor after each voxel is completed, the completed radio frequency signal data is obtained.

4. The network interference localization method based on multi-source data fusion according to claim 1, characterized in that, The cross-modal alignment of the frequency domain feature vector and the spatial feature vector includes: An encoder is constructed to map the frequency domain feature vector and the spatial feature vector to a shared latent space, and a discriminator is constructed to determine the source of the latent space features in the shared latent space; The encoder and the discriminator are jointly optimized based on an adversarial training strategy; Cross-modal alignment of the frequency domain feature vector and the spatial feature vector is performed based on the trained encoder.

5. The network interference localization method based on multi-source data fusion according to claim 1, characterized in that, Determining the location of the interference source based on the positioning parameter set includes: The location parameter set is processed based on the candidate location screening model to determine the distance confidence interval; Based on real-time network load, a target inference model is selected from the first and second inference models to process the location parameters of each candidate interference source in the distance confidence interval, thereby determining the location of the interference source; the accuracy of the first inference model is greater than that of the second inference model. If the inference time of the target inference model is longer than the preset inference time, the location of the reference base station is processed based on the weighted centroid algorithm to obtain the location of the interference source.

6. A network interference localization device based on multi-source data fusion, characterized in that, include: The data acquisition module is used to acquire radio frequency signal data, environmental perception data, and network topology data; The feature extraction module is used to extract the frequency domain feature vector of the radio frequency signal data, the spatial feature vector of the environmental perception data, and the topological feature vector of the network topology data, and to perform cross-modal alignment on the frequency domain feature vector and the spatial feature vector. The feature fusion module is used to dynamically adjust the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector, and the topological feature vector based on the real-time network load, and to fuse each feature vector based on the fusion weights to obtain a fused feature tensor. The localization determination module is used to perform non-line-of-sight propagation correction on the fused feature tensor to obtain a localization parameter set, and to determine the location of the interference source based on the localization parameter set; The dynamic adjustment of the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector, and the topological feature vector based on real-time network load includes: Construct a weight matrix consisting of the fusion weights corresponding to the frequency domain feature vector, the spatial feature vector, and the topological feature vector; The reward function is determined based on the real-time network load and error localization feedback value, and reinforcement learning is performed based on the reward function to update the weight matrix; If the number of consecutive detections where the error localization feedback value is greater than the preset error is greater than the preset number, a population composed of multiple weight matrix individuals is established; different weight matrix individuals represent different fusion weight selection schemes. The individuals in the population are sorted by weight matrix based on their historical positioning performance, and a predetermined proportion of individuals with the highest historical positioning performance are selected as the target population. Add perturbation to the offspring individuals of the target population to obtain offspring individuals with genetic mutation; If the decrease in the error positioning feedback value corresponding to the fusion weight selection scheme of the offspring individual is greater than the preset magnitude, the fusion weight corresponding to the offspring individual is determined as an effective fusion weight. The non-line-of-sight propagation correction of the fused feature tensor to obtain the localization parameter set includes: A first correction engine is constructed; the first correction engine is used to compare the frequency domain feature vector in the fused feature tensor with the signal feature vectors of each preset non-line-of-sight scene in the multipath fingerprint database, and correct the fused feature tensor based on the multipath deviation of the matched preset non-line-of-sight scene. A second correction engine is constructed; the second correction engine is used to construct a ray projection model based on the three-dimensional environment point cloud, and to adjust the signal propagation parameters of the mid-frequency domain feature vector of the fused feature tensor based on the ray projection model; The fused feature tensor is corrected by non-line-of-sight propagation based on the first correction engine and the second correction engine to obtain the localization parameter set.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the network interference localization method based on multi-source data fusion as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the network interference localization method based on multi-source data fusion as described in any one of claims 1 to 5.

Images