Wireless network interference positioning and regulation method and device

Abstract

This invention provides a method and apparatus for wireless network interference localization and control, relating to the field of data processing technology. The method includes: acquiring multi-source heterogeneous data within a target area; performing structured preprocessing on the multi-source heterogeneous data to generate a spatiotemporal matrix dataset with a unified timestamp; inputting the spatiotemporal matrix dataset into a dual-channel feature extractor; using a dynamic confidence entropy weighting algorithm to weight and fuse the waveform features and the signal-to-noise ratio degradation trajectory features to generate a dynamic interference fingerprint spectrum; obtaining interference identification and localization results based on the dynamic interference fingerprint spectrum; generating an optimal control strategy based on the interference identification and localization results using a pre-trained reinforcement learning decision model; executing the optimal control strategy through a software-defined network controller; and updating the dual-channel feature extractor and the reinforcement learning decision model based on a federated learning architecture.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for locating and controlling wireless network interference. Background Technology

[0002] With the rapid development of mobile communication networks, the large-scale deployment of 4G / 5G multi-standard heterogeneous networks has led to increasingly scarce wireless spectrum resources and a more prominent network interference problem. In complex scenarios such as industrial parks, along rail transit lines, and in high-density urban areas, interference sources are diverse and dynamically changing, posing a severe challenge to traditional interference management technologies.

[0003] Existing interference management methods mainly rely on static rule matching and empirical threshold judgment. These methods typically use fixed thresholds to detect anomalies in indicators such as reference signal received power, signal-to-interference-plus-noise ratio, or bit error rate. When an indicator exceeds the threshold, a manual or semi-automatic adjustment process is triggered. The extraction of interference features mainly relies on data collected by spectrum detectors, and basic Fourier transform or energy detection algorithms are used for analysis. Control strategies are mostly fixed in the form of preset templates.

[0004] The above methods have obvious shortcomings in practical applications. The static threshold mechanism and fixed control strategy are difficult to adapt to the dynamic changes in the interference environment, resulting in low interference identification accuracy, inaccurate positioning, and lagging and untargeted control response. They cannot meet the actual needs of accurate interference perception and adaptive control in complex dynamic wireless environments. Summary of the Invention

[0005] This invention provides a method and apparatus for locating and controlling interference in wireless networks, which solves the problem that existing technologies cannot meet the practical needs of accurate interference sensing and adaptive control in complex and dynamic wireless environments.

[0006] This invention provides a method for locating and controlling wireless network interference, comprising the following steps: Acquire multi-source heterogeneous data within the target area, perform structured preprocessing on the multi-source heterogeneous data, and generate a spatiotemporal matrix dataset with a unified timestamp; wherein, the multi-source heterogeneous data includes wireless measurement reports, spectrum scan data, and user equipment signaling data; The spatiotemporal matrix dataset is input into a dual-channel feature extractor, wherein the dual-channel feature extractor includes a first channel and a second channel. The first channel is used to extract waveform features of the spectrum scan data, and the second channel is used to extract signal-to-noise ratio degradation trajectory features of the wireless measurement report. The waveform features and the signal-to-noise ratio degradation trajectory features are weighted and fused using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum. Interference identification and localization results are obtained based on the dynamic interference fingerprint spectrum, wherein the interference identification and localization results include interference type and spatial distribution hotspots representing the location of interference. Based on the interference identification and localization results, an optimal control strategy is generated using a pre-trained reinforcement learning decision model. The generation of the optimal control strategy includes generating a candidate control instruction set and then screening the candidate control instruction set after simulation verification in a digital twin network environment. The optimal control strategy is executed by a software-defined network controller, and the dual-channel feature extractor and the reinforcement learning decision model are updated based on a federated learning architecture.

[0007] According to a wireless network interference localization and control method provided by the present invention, the structured preprocessing of the multi-source heterogeneous data includes: Calculate the quality weight of the data source in the multi-source heterogeneous data; wherein the quality weight is calculated based on the packet loss rate of the data source, the maximum latency deviation between the data source and the master clock, and the structural integrity score of the data source; The standardized observation vectors of the data source are weighted and summed using the quality weights to obtain the fused input data; The input data is processed for differential privacy using an asynchronous Laplace perturbation algorithm; wherein the differential privacy processing includes superimposing perturbation noise on data blocks located in user-sensitive areas, and the amplitude of the perturbation noise is inversely proportional to the quality weight.

[0008] According to the present invention, a method for locating and controlling wireless network interference includes the following step: weighting and fusing the waveform features and the signal-to-noise ratio degradation trajectory features using a dynamic confidence entropy weighting algorithm. The average information entropy of the predicted distribution of known interference categories by the first channel and the second channel within a preset time window is calculated to obtain the feature confidence entropy of the first channel and the feature confidence entropy of the second channel. Calculate the fusion weight of the first channel and the fusion weight of the second channel; wherein, the calculation rule for the fusion weight is: the lower the feature confidence entropy, the higher the corresponding fusion weight; The waveform features are weighted using the fusion weight of the first channel, and the signal-to-noise ratio degradation trajectory features are weighted using the fusion weight of the second channel. The weighted waveform features and the weighted signal-to-noise ratio degradation trajectory features are then concatenated to generate the dynamic interference fingerprint spectrum.

[0009] According to the present invention, a method for locating and controlling wireless network interference, wherein obtaining interference identification and location results based on the dynamic interference fingerprint spectrum includes: The dynamic interference fingerprint spectrum is classified and predicted using the support vector set to obtain the prediction confidence. When the prediction confidence is lower than a set threshold, a boundary migration strategy is executed; wherein, the boundary migration strategy includes: calculating the direction vector of the centroid of the current sample and the support vector set, combining the direction vector and the prediction confidence to extrapolate and expand the boundary of the support vector set, and reclassifying using the expanded support vector set.

[0010] According to the present invention, a method for locating and controlling wireless network interference, wherein generating an optimal control strategy using a pre-trained reinforcement learning decision model involves enhancing the state space of the reinforcement learning decision model, the enhancement including: Construct a disturbance factor estimation vector; wherein the construction of the disturbance factor estimation vector is based on the unit loss weight of the interference source, and the unit loss weight is determined by the normalized capacity of the affected cell in the grid neighborhood of the interference hot zone, the unit time throughput reduction of the affected cell, and the spectral overlap factor of the affected cell. The perturbation factor estimation vector is concatenated with the original network state vector to form a high-dimensional state representation, which is then used as input data into the reinforcement learning decision model.

[0011] According to a wireless network interference localization and control method provided by the present invention, the training process of the reinforcement learning decision model adopts a reward differential attribution regularization method, which includes: Calculate the gradient contribution of the action dimension output by the reinforcement learning decision model to the overall reward function; Calculate the difference between the historical average value of the action dimension and the gradient contribution, and use the difference as a difference constraint term; The final reward function is obtained by subtracting the difference constraint term from the original reward function, and the parameters of the reinforcement learning decision model are updated using the final reward function.

[0012] According to a wireless network interference localization and control method provided by the present invention, the step of executing the optimal control strategy through a software-defined network controller includes: The optimal control strategy is decomposed into base station-level executable instructions; wherein, the decomposition process includes: combining the service quality impact weight vector and the service quality improvement effect of atomic operations to select instructions that match the access network equipment standard; The base station-level executable instructions are issued, and the Shannon entropy of the feedback signal distribution after execution is monitored in real time. When the Shannon entropy increases, the trigger threshold for policy rollback is reduced.

[0013] According to a wireless network interference localization and control method provided by the present invention, the step of updating the dual-channel feature extractor and the reinforcement learning decision model based on a federated learning architecture includes: Calculate the aggregate weight of the uploaded gradient of the edge node; wherein the aggregate weight is determined based on the standard deviation of the accuracy of the edge node in the local test, and the larger the standard deviation, the lower the aggregate weight; Calculate the local drift intensity; wherein, the local drift intensity is defined as the relative entropy between the local model prediction distribution in the current round and the local model prediction distribution in the previous round; When the local drift intensity exceeds a set threshold, an active synchronization process is triggered, and the gradient of the edge node is preferentially included in the central aggregation queue.

[0014] The present invention also provides a wireless network interference location and control device, comprising the following modules: The generation module is used to acquire multi-source heterogeneous data within the target area, perform structured preprocessing on the multi-source heterogeneous data, and generate a spatiotemporal matrix dataset with a unified timestamp; wherein, the multi-source heterogeneous data includes wireless measurement reports, spectrum scan data, and user equipment signaling data; The input module is used to input the spatiotemporal matrix dataset into a dual-channel feature extractor, wherein the dual-channel feature extractor includes a first channel and a second channel, the first channel is used to extract the waveform features of the spectrum scan data, and the second channel is used to extract the signal-to-noise ratio degradation trajectory features of the wireless measurement report; The fusion module is used to perform weighted fusion of the waveform features and the signal-to-noise ratio degradation trajectory features using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum, and obtain interference identification and localization results based on the dynamic interference fingerprint spectrum, wherein the interference identification and localization results include interference type and spatial distribution hotspots characterizing the interference location; The optimization module is used to generate the optimal control strategy based on the interference identification and localization results using a pre-trained reinforcement learning decision model. The generation of the optimal control strategy includes generating a candidate control instruction set and then screening the candidate control instruction set after simulation verification in a digital twin network environment. The execution module is used to execute the optimal control strategy through a software-defined network controller and update the dual-channel feature extractor and the reinforcement learning decision model based on the federated learning architecture.

[0015] The present invention also provides an electronic 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 wireless network interference localization and control method as described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the wireless network interference localization and control method as described above.

[0017] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the wireless network interference location and control method described above.

[0018] The wireless network interference localization and control method and apparatus provided by this invention acquires multi-source heterogeneous data, including wireless measurement reports, spectrum scanning data, and user equipment signaling data, and performs structured preprocessing to generate a spatiotemporal matrix dataset with a unified timestamp. A dual-channel feature extractor extracts waveform features and signal-to-noise ratio degradation trajectory features, which are then fused using a dynamic confidence entropy weighted algorithm to generate a dynamic interference fingerprint spectrum. Based on the dynamic interference fingerprint spectrum, interference identification and localization results, including interference types and spatially distributed hot zones representing interference locations, are obtained. A reinforcement learning decision model is used to generate a candidate control instruction set, and the optimal control strategy is selected through simulation verification in a digital twin network environment. The optimal control strategy is executed by a software-defined network controller, and the model is updated based on a federated learning architecture. This achieves accurate identification, localization, and adaptive control of interference in complex dynamic wireless environments, improving the accuracy of interference identification and the effectiveness of control strategies, and enhancing the system's adaptability to dynamically changing environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention 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 invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the wireless network interference location and control method provided by the present invention; Figure 2 This is a data preprocessing architecture diagram provided by the present invention; Figure 3 This is a diagram of the dual-channel feature extractor architecture provided by the present invention; Figure 4 This is a flowchart of the reinforcement learning regulation strategy provided by the present invention; Figure 5 This is a diagram of the federated learning closed-loop optimization architecture provided by the present invention; Figure 6 This is a schematic diagram of the wireless network interference location and control device provided by the present invention; Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] The wireless network interference localization and control method provided by this invention can be executed by an electronic device with data processing capabilities. This electronic device includes, but is not limited to, servers, cloud computing platforms, edge computing nodes, network management systems, or dedicated network optimization controllers. In this embodiment, a network optimization controller deployed in an operator's network management center is used as the execution entity. This network optimization controller has a multi-source data acquisition interface, a high-performance computing unit, a model inference engine, and a network control interface.

[0023] Figure 1 This is a flowchart illustrating the wireless network interference location and control method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: Step 110: Acquire multi-source heterogeneous data within the target area, perform structured preprocessing on the multi-source heterogeneous data, and generate a spatiotemporal matrix dataset with a unified timestamp; wherein, the multi-source heterogeneous data includes wireless measurement reports, spectrum scan data, and user equipment signaling data; In this application, the target area refers to the coverage area of ​​the wireless network that needs to be monitored and controlled for interference. It can be the coverage cell of a single base station or a regional network composed of multiple base stations, such as typical scenarios like industrial parks, rail transit lines, and high-density urban commercial districts.

[0024] Multi-source heterogeneous data refers to network measurement data collected from different types of data sources, possessing different data formats and sampling characteristics. In this embodiment, the multi-source heterogeneous data includes wireless measurement reports, spectrum scan data, and user equipment signaling data.

[0025] A wireless measurement report refers to wireless channel measurement data periodically reported by user equipment or base stations, including but not limited to indicators such as reference signal received power, signal-to-interference-plus-noise ratio, and reference signal received quality. The wireless measurement report reflects the network signal quality perceived by the user equipment at its current location.

[0026] Spectrum scan data refers to radio frequency spectrum information collected by spectrum sensing devices or spectrum monitoring modules built into base stations, including but not limited to power spectral density, signal bandwidth, and spectral energy distribution at each frequency point. The spectrum scan data can intuitively present the signal distribution within the target frequency band and the frequency domain characteristics of potential interference sources.

[0027] User equipment signaling data refers to control plane messages generated during the establishment, handover, and release of radio resource control connections, including but not limited to connection request messages, measurement configuration messages, and handover command messages. This user equipment signaling data reflects network resource scheduling status and user service behavior patterns.

[0028] Methods for acquiring multi-source heterogeneous data include, but are not limited to: real-time acquisition via distributed sensor networks, reading from the database of the network management system, pulling from the base station via standardized interfaces, or acquiring from third-party data platforms. In this embodiment, the network optimization controller acquires the above three types of data in real time through data acquisition agents deployed on each base station.

[0029] Structured preprocessing refers to the process of converting, aligning, filtering out anomalies, and standardizing collected heterogeneous data from multiple sources. Because different data sources may have different sampling periods, data formats, and field definitions, structured preprocessing is necessary to convert them into a unified data representation.

[0030] More specifically, structured preprocessing includes the following steps: Time window segmentation refers to dividing a continuously acquired data stream into discrete time segments according to a preset time granularity. In this embodiment, a sliding window method is used for segmentation. The window length can be set from 100 milliseconds to 1 second, and the sliding step size can be set to half the window length to ensure data continuity between adjacent windows.

[0031] Outlier filtering refers to identifying and removing obvious errors or outliers from data. Outliers may arise from equipment malfunctions, transmission errors, or data acquisition anomalies. In this embodiment, a statistical distribution-based anomaly detection method is used. When a data point deviates from the historical mean of the data source by more than three standard deviations, it is marked as an outlier and imputed by interpolation or directly removed.

[0032] Format standardization refers to converting data from different data sources into a unified numerical representation range and data structure. In this embodiment, various measurement indicators are normalized, mapped to a numerical range of zero to one, and all data is organized into a unified key-value pair format.

[0033] A spatiotemporal matrix dataset refers to a multidimensional data structure organized with time and space dimensions as coordinate axes after structured preprocessing. The time dimension corresponds to the data acquisition time, and the space dimension corresponds to the data acquisition location or associated network node. The unified timestamp refers to assigning a unified time base to data from different data sources, eliminating time offsets caused by differences in acquisition latency.

[0034] Figure 2 This is a data preprocessing architecture diagram provided by the present invention, such as... Figure 2 As shown, after the raw data is collected, it is processed by a lightweight stream processing engine to perform time window segmentation, outlier filtering and format standardization operations in sequence, and finally generate a spatiotemporal matrix dataset.

[0035] Step 120: Input the spatiotemporal matrix dataset into a dual-channel feature extractor, wherein the dual-channel feature extractor includes a first channel and a second channel, the first channel is used to extract waveform features of the spectrum scan data, and the second channel is used to extract signal-to-noise ratio degradation trajectory features of the wireless measurement report; A dual-channel feature extractor refers to a feature extraction module that includes two parallel processing channels, each employing a corresponding feature extraction algorithm for different types of input data. The purpose of using a dual-channel structure is that spectrum scan data and wireless measurement reports have different data characteristics and semantic information. Using a unified feature extraction method is insufficient to fully extract the discriminative features of each type of data. Therefore, it is necessary to design dedicated processing channels to perform feature extraction separately.

[0036] The first channel is used to extract waveform features from the spectrum scan data. These waveform features refer to the signal morphological characteristics of the spectrum scan data in the time and frequency domains, including but not limited to the signal envelope shape, pulse width, repetition period, and spectrum occupancy pattern. These waveform features can characterize the physical layer properties of interference signals and help distinguish different types of interference sources.

[0037] In this embodiment, the feature extraction process of the first channel includes: firstly, performing a short-time Fourier transform on the spectral scan data to generate a time-spectrum representation; then, using a convolutional neural network to extract features from the time-spectrum and outputting a waveform feature vector. The convolutional neural network may employ a deformable convolutional structure to enhance its ability to represent irregular interference waveforms.

[0038] The second channel is used to extract the signal-to-noise ratio (SNR) degradation trajectory features from the wireless measurement report. These SNR degradation trajectory features refer to the dynamic pattern of the signal-to-interference-plus-noise ratio (SNR) over time, reflecting the degree of interference's impact on user experience and the time-varying characteristics of the interference.

[0039] In this embodiment, the feature extraction process of the second channel includes: firstly, constructing a network topology graph structure based on the cell identifier and user identifier in the wireless measurement report, where nodes represent cells or users, and edges represent the adjacency relationship between cells or the association relationship between users and serving cells; then, using a graph neural network to extract features from the topology graph structure, capturing the propagation mode of signal-to-noise ratio degradation on the network topology, and outputting a signal-to-noise ratio degradation trajectory feature vector.

[0040] Figure 3 This is a diagram of the dual-channel feature extractor architecture provided by the present invention, as shown below. Figure 3 As shown, the spectrum scanning data is input into the spatiotemporal graph convolutional network channel for processing, and the wireless measurement report data is input into the adaptive attention channel for processing. The output features of the two channels are fused in the cross-channel feature fusion module to finally generate a dynamic interference fingerprint spectrum.

[0041] Step 130: The waveform features and the signal-to-noise ratio degradation trajectory features are weighted and fused using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum. Interference identification and localization results are obtained based on the dynamic interference fingerprint spectrum. The interference identification and localization results include interference type and spatial distribution hotspots representing the location of interference. The dynamic confidence entropy weighted algorithm is an algorithm that dynamically adjusts the fusion weights based on the confidence level of each channel's features. The purpose of using dynamic weighting instead of fixed weights is that the feature representation capabilities of the two channels may differ under different interference scenarios. In some scenarios, waveform features are more discriminative, while in others, signal-to-noise ratio degradation trajectory features are more discriminative. Therefore, it is necessary to dynamically adjust the contribution of each channel according to the actual situation.

[0042] Weighted fusion refers to combining the waveform features and the signal-to-noise ratio degradation trajectory features according to their respective fusion weights to form a unified feature representation. The weighted fusion methods include, but are not limited to, weighted summation, weighted concatenation, and attention weighting. In this embodiment, weighted concatenation is used, that is, the feature vectors of the two channels are multiplied by their respective fusion weights and then concatenated.

[0043] A dynamic interference fingerprint refers to a high-dimensional feature representation formed after fusion, which can uniquely characterize the interference state within the target area at the current time. The dynamic interference fingerprint includes multi-dimensional information such as the frequency domain characteristics, time domain characteristics, spatial distribution characteristics, and impact characteristics on network performance of the interference.

[0044] Interference identification and localization results refer to the interference analysis conclusions obtained based on the dynamic interference fingerprint spectrum. The interference identification and localization results include the interference type and spatial distribution hotspots characterizing the interference location.

[0045] Interference type refers to the classification and determination of the nature of the interference source, including but not limited to co-channel interference, adjacent-channel interference, intermodulation interference, blocking interference, and interference from external sources. Different types of interference require different control strategies for handling.

[0046] Spatial distribution hotspots refer to geographical areas or sets of network nodes with a high degree of interference impact, used to characterize the spatial location information of the interference. These spatial distribution hotspots can be represented using heat maps, grid maps, or lists of affected cells, etc.

[0047] In this embodiment, the process of obtaining interference identification and localization results includes: inputting the dynamic interference fingerprint spectrum into a pre-trained classifier model, the classifier model outputting the probability distribution of each interference type, and selecting the category with the highest probability as the interference type determination result; at the same time, generating a spatial distribution map of interference intensity based on the feature intensity of each spatial location in the dynamic interference fingerprint spectrum, and marking the area where the interference intensity exceeds a preset threshold as a spatial distribution hotspot.

[0048] Step 140: Based on the interference identification and localization results, generate the optimal control strategy using a pre-trained reinforcement learning decision model. The generation of the optimal control strategy includes generating a candidate control instruction set and then screening the candidate control instruction set after simulation verification in a digital twin network environment. A reinforcement learning decision model is a policy network model trained using reinforcement learning algorithms, capable of outputting corresponding regulatory actions based on the current network state and interference conditions. Reinforcement learning is a machine learning method that optimizes decision-making policies by interacting with the environment and based on reward signals; it is suitable for handling optimization problems with temporal decision-making characteristics.

[0049] Pre-training refers to the process of training the reinforcement learning decision model offline using historical data or a simulation environment before actual deployment. Through pre-training, the reinforcement learning decision model can learn the mapping relationship between disturbance states and optimal control actions.

[0050] The process of generating the optimal control strategy includes two stages: first, generating a set of candidate control instructions, and then simulating and verifying the set of candidate control instructions in a digital twin network environment before screening.

[0051] The candidate control instruction set refers to a set of possible control schemes output by the reinforcement learning decision model. Each control scheme includes one or more control instructions, which include, but are not limited to, power adjustment instructions, frequency switching instructions, beamforming instructions, and resource scheduling instructions.

[0052] A digital twin network environment refers to a simulation environment formed by digitally modeling a real wireless network, capable of simulating the response behavior of a real network to control commands. The purpose of simulation verification in a digital twin network environment is to pre-assess the potential effects and risks of control strategies before they are implemented on the real network, thus avoiding network performance degradation caused by inappropriate control.

[0053] The simulation verification process includes: applying each control scheme in the candidate control instruction set to the digital twin network environment in sequence, performing multiple rounds of simulation using the Monte Carlo simulation method, and statistically analyzing the performance of each scheme in the simulation environment, including indicators such as the improvement in signal-to-noise ratio, service interruption duration, and power consumption changes.

[0054] Screening refers to selecting the optimal control scheme from the candidate control instruction set based on simulation verification results. The optimal control strategy should meet preset optimization objectives, including but not limited to improving the signal-to-noise ratio to a preset threshold, reducing the service interruption rate to a preset threshold, and keeping the power consumption increase within an acceptable range.

[0055] Figure 4 This is a flowchart of the reinforcement learning regulation strategy provided by the present invention, such as... Figure 4 As shown, the interference identification results are input into the reinforcement learning decision model to generate a set of control instructions. After verification by digital twin simulation, the optimal strategy is selected.

[0056] Step 150: Execute the optimal control strategy through a software-defined network controller, and update the dual-channel feature extractor and the reinforcement learning decision model based on the federated learning architecture.

[0057] A software-defined network controller (SDB) is a centralized network control device implemented using software-defined networking (SDN) technology. It enables unified configuration and management of network devices through a standardized southbound interface. The advantages of using a SDB to execute control strategies include: unified control of heterogeneous network devices and support for coordinated control of 4G and 5G multi-standard devices.

[0058] The process of executing the optimal control strategy includes: parsing the optimal control strategy into specific network configuration parameters, sending the configuration parameters to the corresponding base station equipment through the network configuration protocol, and adjusting the radio frequency parameters or resource scheduling parameters according to the configuration parameters.

[0059] Federated learning architecture refers to a distributed machine learning framework that allows multiple participants to collaboratively train a shared model without sharing the original data. The advantages of using a federated learning architecture for model updates are: each edge node can use local data for model training without uploading sensitive network measurement data to a central server, thus protecting data privacy while enabling continuous model optimization.

[0060] The update process includes: each edge node uses locally collected data to train the dual-channel feature extractor and the reinforcement learning decision model locally, and calculates the update gradient of the model parameters; each edge node uploads the update gradient to the central server; the central server aggregates the gradients uploaded by each edge node and updates the global model parameters; the central server distributes the updated global model parameters to each edge node.

[0061] Figure 5 This is a diagram of the federated learning closed-loop optimization architecture provided by the present invention, such as... Figure 5 As shown, multiple edge nodes train local models, update the global model through an adaptive confidence weight gradient aggregation mechanism, and then distribute the updated model to each edge node.

[0062] This application achieves accurate identification, localization, and adaptive control of wireless network interference. By acquiring multi-source heterogeneous data and performing structured preprocessing, the problem of effective data fusion is solved. A dual-channel feature extractor extracts spectral waveform features and signal-to-noise ratio degradation trajectory features, enabling effective characterization of the multi-dimensional characteristics of interference. Feature fusion using a dynamic confidence entropy weighting algorithm improves the discrimination capability of interference fingerprints. An optimal control strategy is generated through a reinforcement learning decision model combined with digital twin simulation verification, making control decisions more accurate and effective. A software-defined network controller executes the control strategy and updates the model based on a federated learning architecture, achieving coordinated control of multiple device types and continuous system evolution and optimization.

[0063] Optionally, the multi-source heterogeneous data undergoes structured preprocessing, including: Calculate the quality weight of the data source in the multi-source heterogeneous data; wherein the quality weight is calculated based on the packet loss rate of the data source, the maximum latency deviation between the data source and the master clock, and the structural integrity score of the data source; The standardized observation vectors of the data source are weighted and summed using the quality weights to obtain the fused input data; The input data is processed for differential privacy using an asynchronous Laplace perturbation algorithm; wherein the differential privacy processing includes superimposing perturbation noise on data blocks located in user-sensitive areas, and the amplitude of the perturbation noise is inversely proportional to the quality weight.

[0064] In this application, a data source refers to the entity that provides multi-source heterogeneous data, with each data source corresponding to an independent data acquisition channel or data provider. In this embodiment, wireless measurement reports from different base stations and spectrum scanning data collected by different spectrum monitoring devices are considered as different data sources.

[0065] Quality weight is a numerical indicator used to quantify the reliability of a data source. The higher the quality weight, the more reliable the data from that data source, and the greater its influence should be given in the subsequent data fusion process.

[0066] The quality weight is calculated based on the packet loss rate of the data source, the maximum latency deviation between the data source and the master clock, and the structural integrity score of the data source.

[0067] Packet loss rate refers to the ratio of the number of data packets lost during data transmission to the total number of data packets sent. A higher packet loss rate indicates that the transmission channel of the data source is less reliable and the data integrity is worse.

[0068] The master clock refers to the unified time reference used by the system, typically provided by a high-precision clock source. The maximum time delay deviation refers to the maximum time difference between the sampling time of the data source and the master clock. A larger time delay deviation indicates poorer time synchronization accuracy of the data source and greater difficulty in aligning it with other data sources.

[0069] Structural integrity score is a score obtained by evaluating the completeness of fields in the data reported by the data source. The higher the integrity score, the more complete the data fields reported by the data source and the fewer missing key fields.

[0070] The weighting is based on a quality function composed of the following indicators: ; in, This represents the packet loss rate of the i-th source. This represents the maximum time delay deviation between the data stream and the master clock. The structural integrity of the system is scored, where τ is the expected synchronization window of the system. This is a hyperparameter used to balance the contributions of the three metrics to the overall quality. This weighting mechanism naturally suppresses the influence of data sources with large latency jitter or severe data gaps in dynamic acquisition environments, thereby improving the accuracy and robustness of the final generated interference feature matrix in the subsequent ST-GCN input stage.

[0071] In this application, a standardized observation vector refers to a vector representation obtained after standardizing the raw data reported by the data source. Standardization includes operations such as numerical normalization and format standardization, making data from different data sources comparable.

[0072] Weighted summation refers to the process of multiplying the standardized observation vectors of each data source by their corresponding quality weights and then performing vector addition.

[0073] The formula for calculating the fused input data is: ; in, Indicates the first i Data sources at timestamps t The standardized observation vectors below, Its corresponding quality weight, This is the high-quality input data after fusion.

[0074] By using quality-weighted fusion, high-quality data sources contribute more to the final fusion result, while the influence of low-quality data sources is naturally suppressed, thereby improving the accuracy and robustness of the generated interference feature matrix in subsequent feature extraction stages.

[0075] In this application, differential privacy processing refers to a data privacy protection technology that adds random noise to data, making it impossible for attackers to infer any sensitive information about a single user from the published data. The purpose of applying differential privacy processing in this invention is that wireless measurement reports and user equipment signaling data may contain sensitive information such as user location and behavioral patterns, requiring privacy protection processing.

[0076] The asynchronous Laplace perturbation algorithm is a differential privacy implementation method that selectively adds Laplace noise. Unlike traditional methods that uniformly add noise to all data, the asynchronous Laplace perturbation algorithm only adds noise to data blocks located in user-sensitive areas, and the noise amplitude is adaptively adjusted according to data quality weights.

[0077] User sensitive segments refer to data fields or data fragments that are directly related to sensitive information such as user identity, location, and behavior.

[0078] The amplitude of the disturbance noise is inversely proportional to the quality weight, and the specific disturbance term is defined as follows: ; in, μ For privacy control factors, b is the Laplace distribution scaling parameter. Lap(b) Indicates that it follows the parameter. b A random variable with a Laplace distribution.

[0079] This design enhances privacy by preserving availability from high-quality data sources, while compressing interfering information through perturbation of weak-quality data sources, thus achieving adaptive adjustment of information strength during preprocessing. The overall strategy strengthens the discriminative power of data fusion, the privacy compliance of processing, and the efficiency of subsequent interference identification by the graph neural network.

[0080] This invention realizes quality-weighted multi-source data fusion and differential privacy protection, which is a key step in realizing the transformation from raw heterogeneous data into structured, highly robust input tensors.

[0081] Optionally, the step of using a dynamic confidence entropy weighting algorithm to weight and fuse the waveform features and the signal-to-noise ratio degradation trajectory features includes: The average information entropy of the predicted distribution of known interference categories by the first channel and the second channel within a preset time window is calculated to obtain the feature confidence entropy of the first channel and the feature confidence entropy of the second channel. Calculate the fusion weight of the first channel and the fusion weight of the second channel; wherein, the calculation rule for the fusion weight is: the lower the feature confidence entropy, the higher the corresponding fusion weight; The waveform features are weighted using the fusion weight of the first channel, and the signal-to-noise ratio degradation trajectory features are weighted using the fusion weight of the second channel. The weighted waveform features and the weighted signal-to-noise ratio degradation trajectory features are then concatenated to generate the dynamic interference fingerprint spectrum.

[0082] In this application, the preset time window refers to the historical time range used to calculate the confidence entropy statistic. In this embodiment, it is set to the most recent T time steps, which is called the fusion period window.

[0083] Known interference categories refer to the set of interference types that have been defined during the model training phase, including but not limited to C categories such as co-channel interference, adjacent-channel interference, intermodulation interference, blocking interference, and external interference.

[0084] The prediction distribution refers to the probability values ​​of each category output by each channel when classifying and predicting input features. The average information entropy refers to the arithmetic mean of the prediction distribution entropy values ​​at each time step within a preset time window.

[0085] The formula for calculating feature confidence entropy is: ; in, Let be the output probability of channel k for category c at time step t, where C is the total number of categories and T is the fusion period window length. A lower entropy value indicates a more concentrated output and higher confidence level for that channel, corresponding to a higher weight. The larger.

[0086] This mechanism enables adaptive weighting during feature fusion, preventing interference from broadband interference coverage in the spectral channel or feature degradation caused by sparse topology in the graph neural network channel, thereby improving the accuracy and robustness of interference identification.

[0087] The calculation rule for the fusion weight is as follows: the lower the feature confidence entropy, the higher the corresponding fusion weight. The design basis for this rule is that a low confidence entropy indicates that the prediction of that channel is more certain, meaning that the features extracted by that channel have a stronger discriminative ability for the current sample, and therefore should be given a higher weight in the fusion process.

[0088] The confidence entropy weighting function is defined as follows: ; in, The fusion weight for the k-th channel is... This represents the feature confidence entropy within the current time window of the k-th channel. and These are the initial feature vectors extracted from the spectral waveform channel and the graph convolution channel, respectively, and F is the fused feature finally input to the CCA-GRU module.

[0089] In this application, adaptive weight adjustment can be achieved during the feature fusion process, avoiding interference to the recognition results when the spectral channel encounters broadband interference coverage or when the graph neural network channel suffers feature degradation due to topological graph sparsity, thereby improving the accuracy and robustness of interference recognition.

[0090] Optionally, obtaining the interference identification and localization results based on the dynamic interference fingerprint spectrum includes: The dynamic interference fingerprint spectrum is classified and predicted using the support vector set to obtain the prediction confidence. When the prediction confidence is lower than a set threshold, a boundary migration strategy is executed; wherein, the boundary migration strategy includes: calculating the direction vector of the centroid of the current sample and the support vector set, combining the direction vector and the prediction confidence to extrapolate and expand the boundary of the support vector set, and reclassifying using the expanded support vector set.

[0091] In this application, the support vector set refers to the set of key samples used in a support vector machine classifier to define the classification decision boundary. In this embodiment, an incremental support vector machine is used as the classifier. The incremental support vector machine can dynamically update the support vector set based on newly arrived samples without retraining the entire model, making it suitable for handling online learning scenarios.

[0092] The classification prediction refers to using the dynamic interference fingerprint as input and employing a support vector machine to calculate its distance from the decision boundaries of each category, thereby determining its category.

[0093] The prediction confidence refers to the degree of certainty the classifier has about the current prediction result, and is defined as the probability value that a sample is classified as the most likely class. .

[0094] Set threshold η This is a critical value used to determine whether the prediction confidence is high enough. When the prediction confidence is lower than the set threshold, it indicates that the current sample may belong to a new type of interference that the classifier has not learned before, or it may be located in a fuzzy region of the classification boundary, requiring special handling strategies.

[0095] Boundary shifting strategy refers to a mechanism that adaptively expands the classification decision boundary when a classifier encounters low-confidence samples. The purpose of this strategy is to enhance the classifier's adaptability to unknown types of interference and avoid misclassifying novel interference as known types.

[0096] The boundary migration strategy includes the following processing steps: Calculate the direction vector of the centroid of the current sample relative to the set of support vectors. The centroid refers to the geometric center point of all support vectors in the set. The direction vector d... It refers to the unit vector pointing from the centroid to the current sample, representing the direction of the current sample relative to the known sample distribution.

[0097] The boundary of the support vector set is extrapolated and expanded by combining the direction vector and the predicted confidence. This extrapolation refers to extending the classification boundary outwards by a certain distance along the direction vector to cover the region where the current low-confidence samples are located. The formula for calculating the expanded support vector set is: ; Where S is the original set of support vectors, δ is the maximum prediction confidence of the current sample, δ is the boundary expansion coefficient, and d is the direction vector between the current sample and the centroid of the support vector.

[0098] This strategy guides the model to expand its recognition capabilities in low-confidence regions by extrapolating the boundary at a limited scale. It also combines new samples synthesized by GAN to gradually fill the gaps in the classifier's interference type distribution space, thereby improving the model's adaptation speed and generalization performance to unknown interference.

[0099] This invention establishes a feature fusion and decision expansion mechanism with confidence-driven as its core, which is a key technical support for building a highly reliable interference fingerprint generation and classification system.

[0100] Optionally, generating the optimal control policy using a pre-trained reinforcement learning decision model involves enhancing the state space of the reinforcement learning decision model, the enhancement including: Construct a disturbance factor estimation vector; wherein the construction of the disturbance factor estimation vector is based on the unit loss weight of the interference source, and the unit loss weight is determined by the normalized capacity of the affected cell in the grid neighborhood of the interference hot zone, the unit time throughput reduction of the affected cell, and the spectral overlap factor of the affected cell. The perturbation factor estimation vector is concatenated with the original network state vector to form a high-dimensional state representation, which is then used as input data into the reinforcement learning decision model.

[0101] In this application, the disturbance factor estimation vector Φ refers to the vector representation used to quantify the degree of impact of each disturbance source on network performance.

[0102] The purpose of constructing the perturbation factor estimation vector is to dynamically fuse the potential influence mapping of the interference source on the basis of the original 48-dimensional features of the state space, so as to construct a high-dimensional state representation with the ability to perceive the propagation trend of perturbation, thereby improving the response accuracy of the dual-delay deep deterministic policy gradient policy network to local abnormal scenarios.

[0103] The construction of the disturbance factor estimation vector is based on the unit loss weight of the interference source. The unit loss weight refers to the degree of unit loss caused by a single interference source to the efficiency of spectrum resource utilization.

[0104] The unit loss weight is determined by the normalized capacity of the affected cell in the grid neighborhood of the interference hot zone, the decrease in throughput per unit time of the affected cell, and the spectral overlap factor of the affected cell.

[0105] Interference hotspots refer to spatially distributed hotspots identified in interference identification and localization results. The grid neighborhood refers to the set of neighboring grid points for each grid point after the interference hotspots are divided into regular grids.

[0106] State enhancement is expressed as: ; Where s is the original network state vector, and s′ is the high-dimensional state representation. Let i be the set of affected cells within the i-th grid neighborhood of the interference hotspot. Let j be the normalized capacity of cell j. Let $\frac{j}{k}$ be the decrease in throughput of cell $j$ per unit time under the influence of interference source $k$. This is the spectral overlap factor.

[0107] This fusion strategy guides the policy network to prioritize the perception of the dynamic evolution of disturbances in key areas through the aggregation calculation of local disturbance effects, thereby improving the sensitivity and accuracy of the disturbance response in action decision-making.

[0108] Optionally, the training process of the reinforcement learning decision model employs a reward differential attribution regularization method, which includes: Calculate the gradient contribution of the action dimension output by the reinforcement learning decision model to the overall reward function; Calculate the difference between the historical average value of the action dimension and the gradient contribution, and use the difference as a difference constraint term; The final reward function is obtained by subtracting the difference constraint term from the original reward function, and the parameters of the reinforcement learning decision model are updated using the final reward function.

[0109] In this application, the action dimension refers to each component in the control action vector output by the reinforcement learning decision model. Each action dimension corresponds to a specific control parameter, such as power adjustment amplitude, frequency switching target, beam direction angle, etc.

[0110] The overall reward function refers to the scalar signal used to evaluate the quality of actions during reinforcement learning training. In this embodiment, the overall reward function integrates signal-to-noise ratio improvement and system stability constraints, and is designed as follows: R=α×ΔSINR-β×Outage_duration-γ×ΔPower ; Where α, β, and γ are weighting coefficients, ΔSINR is the signal-to-noise ratio improvement, Outage_duration is the service interruption duration, and ΔPower is the power consumption change.

[0111] Gradient contribution refers to the partial derivative of the overall reward function with respect to each action dimension. It reflects the degree to which small changes in each action dimension affect the overall reward, denoted as . .

[0112] The historical average contribution value of the action refers to the moving average of the gradient contribution of the i-th action dimension during the historical training process, which is used to reflect the long-term average importance of the action dimension.

[0113] The differential constraint term refers to the difference between the current gradient contribution and the historical average value, used to constrain the update magnitude of each action dimension. The purpose of introducing the reward differential attribution regularization term is to address the slow convergence speed and policy oscillation issues in reinforcement learning under multi-objective weight configurations, constructing a penalty-sensitive attribution matrix, and applying a dynamic adjustment coefficient to each action dimension.

[0114] The final reward function expression is: ; in, Let A be the original reward function, and A be the current set of actions. Let i be the gradient contribution of the i-th action to the overall reward. λ is the historical average contribution value of this action, and λ is the adjustment coefficient.

[0115] This formula suppresses policy disturbances caused by short-term reward fluctuations by differentially constraining the historical average value of actions with the current gradient response, enabling the policy network to form a stable and balanced action response pattern in complex interference scenarios, and ensuring convergence equilibrium between SINR improvement and energy consumption control and service interruption constraints.

[0116] In this application, the state enhancement and reward regularization mechanism described above, along with the dual-delay deep deterministic gradient control strategy, demonstrates strong anti-interference robustness and optimization stability in the face of 20 types of interference superposition in the digital twin environment. It is the core algorithmic support for ensuring the effectiveness of dynamic control strategies for multi-standard access networks.

[0117] Optionally, executing the optimal control strategy through a software-defined network controller includes: The optimal control strategy is decomposed into base station-level executable instructions; wherein, the decomposition process includes: combining the service quality impact weight vector and the service quality improvement effect of atomic operations to select instructions that match the access network equipment standard; The base station-level executable instructions are issued, and the Shannon entropy of the feedback signal distribution after execution is monitored in real time. When the Shannon entropy increases, the trigger threshold for policy rollback is reduced.

[0118] In the application, decomposition refers to the process of converting abstract control strategies into specific network configuration instructions. The base station-level executable instructions refer to configuration instructions for a single base station device, including but not limited to radio frequency parameter configuration instructions, resource scheduling configuration instructions, and beam management configuration instructions.

[0119] The decomposition process involves combining the service quality impact weight vector and the service quality improvement effect of atomic operations to select instructions that match the access network equipment standard.

[0120] To improve the cross-protocol consistency and multi-device compatibility of control strategies in multi-standard network environments, this invention introduces a cross-standard instruction mapping algorithm based on multi-dimensional quality of service driving, and constructs a decoupling mapping mechanism between policy intent and network configuration parameters in the software-defined network controller.

[0121] By constructing a QoS influence weight vector By combining the target influence domain and key performance index sensitivity of each atomic operation in the control strategy, the mapping calculation from the abstract strategy intent space to the actual configuration instruction space is completed. The core mapping function is defined as follows: ; Where a is the policy action vector, and C is the set of executable configuration instructions. This represents the QoS improvement effect produced by the policy action in dimension k through the configuration instruction c. This refers to the weight vector obtained through reinforcement learning during the policy training phase.

[0122] This mapping mechanism dynamically evaluates the potential impact of actions on multiple key performance indicators such as rate, latency, and stability, and automatically selects the most suitable instruction combination for the current standard equipment. This ensures that the reconfiguration of the radio resource control layer in the 4G system and the beamforming control in the 5G system are executed under the same objective, thus ensuring the equivalence and coordination of multi-standard instruction strategies.

[0123] Sending down refers to sending the base station-level executable instructions to the target base station equipment through the southbound interface of the software-defined network controller.

[0124] In this embodiment, the NETCONF / YANG protocol is used as the southbound interface protocol. Feedback signals refer to the status information returned by the base station equipment after executing configuration commands. The Shannon entropy is a measure of the uncertainty in the distribution of the feedback signal, denoted as... .

[0125] The higher the Shannon entropy, the more dispersed the distribution of the feedback signal and the greater the uncertainty, indicating that the system's response to the control strategy is more unstable.

[0126] Policy rollback refers to a protection mechanism that restores the network configuration to its pre-control state when the control policy is ineffective. To avoid local anomaly propagation or policy divergence during instruction execution, this invention designs a policy rollback threshold adjustment algorithm driven by execution feedback entropy. This algorithm senses the uncertainty of the issued instruction response in real time, quantifies the degree of system-level feedback disturbance, and dynamically corrects the rollback trigger threshold τ.

[0127] This dynamic threshold is expressed as: ; in, The initial default failure threshold is δ, which is an adjustment factor. The Shannon entropy of the feedback signal distribution. This is the baseline entropy value during the training phase.

[0128] This mechanism senses the fluctuation in the network's response to the control strategy by detecting changes in the entropy of the feedback signal. When system stability decreases and the entropy value increases, it automatically lowers the tolerance threshold τ, quickly triggering a conservative strategy backoff, thereby effectively suppressing the negative effects of cross-system control link mismatch or interference response lag. This mechanism demonstrates good robustness and execution reliability in multi-system heterogeneous network control systems and is a key technical guarantee for realizing SDN-driven network instruction set coordinated control.

[0129] Optionally, updating the dual-channel feature extractor and the reinforcement learning decision model based on the federated learning architecture includes: Calculate the aggregate weight of the uploaded gradient of the edge node; wherein the aggregate weight is determined based on the standard deviation of the accuracy of the edge node in the local test, and the larger the standard deviation, the lower the aggregate weight; Calculate the local drift intensity; wherein, the local drift intensity is defined as the relative entropy between the local model prediction distribution in the current round and the local model prediction distribution in the previous round; When the local drift intensity exceeds a set threshold, an active synchronization process is triggered, and the gradient of the edge node is preferentially included in the central aggregation queue.

[0130] In this application, edge nodes refer to computing nodes deployed at the network edge, which in this embodiment correspond to edge computing servers or computing units on the base station side in each region. Each edge node is responsible for using local data to train the model and uploading the trained gradients to the central server.

[0131] In the context of federated learning architecture, to achieve distributed closed-loop optimization of the interference identification model, this invention designs an asynchronous gradient aggregation mechanism based on adaptive confidence weights. The core of this mechanism is to integrate the local stability index of edge node model updates and dynamically adjust the contribution of each node's gradient to the global model update.

[0132] To address the sensitivity of traditional federated averaging methods to nodes with high variance, a confidence coefficient is introduced. To reflect the local model robustness of the i-th edge node, the aggregation update formula is defined as follows: ; in, The gradient uploaded for the i-th edge node. This represents the standard deviation of the node's accuracy in the most recent three rounds of local testing, used to measure the stability of the model's output, with λ being the confidence adjustment coefficient. Through this weighting mechanism, the central node tends to absorb gradient contributions from those with stable local training results, thereby enhancing the generalization ability of the aggregated model in complex interference environments.

[0133] To further improve the global model's response efficiency to dynamic network changes, a model synchronization scheduling algorithm driven by local drift intensity is designed. This algorithm uses the node's local drift rate as a synchronization priority indicator to dynamically determine whether it should participate in the global update in advance. Let the local drift intensity be... Defined as the KL divergence between the local model prediction distribution of the current round and the previous round: ; in, Let be the average predicted probability of the i-th node for the j-th type of interference at time t, and C be the total number of interference types. If the local accuracy drops below a set threshold, an active synchronization process is triggered, prioritizing the inclusion of the gradient of that node into the central aggregation queue to prevent the model from accumulating errors under local heterogeneous distribution over a long period of time.

[0134] In this application, the use of homomorphic encryption mechanism to ensure that each round of gradient upload maintains a dynamic balance between model update efficiency and discrimination accuracy while preserving the privacy and security of the entire network is a key strategy foundation for supporting the long-term evolution and high-frequency closed-loop iteration of the interference identification system.

[0135] The wireless network interference location and control device provided by the present invention is described below. The wireless network interference location and control device described below can be referred to in correspondence with the wireless network interference location and control method described above.

[0136] Figure 6 This is a schematic diagram of the wireless network interference location and control device provided by the present invention, as shown below. Figure 6 As shown, it includes: The generation module 610 is used to acquire multi-source heterogeneous data within the target area, perform structured preprocessing on the multi-source heterogeneous data, and generate a spatiotemporal matrix dataset with a unified timestamp; wherein, the multi-source heterogeneous data includes wireless measurement reports, spectrum scan data, and user equipment signaling data; The input module 620 is used to input the spatiotemporal matrix dataset into a dual-channel feature extractor, wherein the dual-channel feature extractor includes a first channel and a second channel, the first channel is used to extract the waveform features of the spectrum scan data, and the second channel is used to extract the signal-to-noise ratio degradation trajectory features of the wireless measurement report; The fusion module 630 is used to perform weighted fusion of the waveform features and the signal-to-noise ratio degradation trajectory features using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum, and obtain interference identification and localization results based on the dynamic interference fingerprint spectrum, wherein the interference identification and localization results include interference type and spatial distribution hotspots characterizing the interference location; The optimization module 640 is used to generate an optimal control strategy based on the interference identification and localization results using a pre-trained reinforcement learning decision model. The generation of the optimal control strategy includes generating a candidate control instruction set and then screening the candidate control instruction set after simulation verification in a digital twin network environment. The execution module 650 is used to execute the optimal control strategy through a software-defined network controller and update the dual-channel feature extractor and the reinforcement learning decision model based on the federated learning architecture.

[0137] In this application, multi-source heterogeneous data, including wireless measurement reports, spectrum scanning data, and user equipment signaling data, are acquired and preprocessed in a structured manner to generate a spatiotemporal matrix dataset with a unified timestamp. Waveform features and signal-to-noise ratio degradation trajectory features are extracted using a dual-channel feature extractor, and fused using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum. Based on the dynamic interference fingerprint spectrum, interference identification and localization results, including interference types and spatial distribution hotspots representing interference locations, are obtained. A reinforcement learning decision model is used to generate a candidate control instruction set, and the optimal control strategy is screened through simulation verification in a digital twin network environment. The optimal control strategy is executed by a software-defined network controller, and the model is updated based on a federated learning architecture. This achieves accurate identification, localization, and adaptive control of interference in complex dynamic wireless environments, improves the accuracy of interference identification and the effectiveness of control strategies, and enhances the system's adaptability to dynamically changing environments.

[0138] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 7 As shown, the electronic device may include a processor 710, a communications interface 720, a memory 730, and a communication bus 740, wherein the processor 710, communications interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a wireless network interference localization and control method. This method includes: acquiring multi-source heterogeneous data within a target area; performing structured preprocessing on the multi-source heterogeneous data; and generating a spatiotemporal matrix dataset with a unified timestamp; wherein the multi-source heterogeneous data includes wireless measurement reports, spectrum scan data, and user equipment signaling data. The spatiotemporal matrix dataset is input into a dual-channel feature extractor, wherein the dual-channel feature extractor includes a first channel and a second channel. The first channel is used to extract waveform features of the spectrum scan data, and the second channel is used to extract signal-to-noise ratio degradation trajectory features of the wireless measurement report. The waveform features and the signal-to-noise ratio degradation trajectory features are weighted and fused using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum. Interference identification and localization results are obtained based on the dynamic interference fingerprint spectrum, wherein the interference identification and localization results include interference type and spatial distribution hotspots representing the location of interference. Based on the interference identification and localization results, an optimal control strategy is generated using a pre-trained reinforcement learning decision model. The generation of the optimal control strategy includes generating a candidate control instruction set and then screening the candidate control instruction set after simulation verification in a digital twin network environment. The optimal control strategy is executed by a software-defined network controller, and the dual-channel feature extractor and the reinforcement learning decision model are updated based on a federated learning architecture.

[0139] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a 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 instructions 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 the present invention. 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.

[0140] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the wireless network interference localization and control method provided by the above methods. The method includes: acquiring multi-source heterogeneous data within a target area, performing structured preprocessing on the multi-source heterogeneous data, and generating a spatiotemporal matrix dataset with a unified timestamp; wherein the multi-source heterogeneous data includes wireless measurement reports, spectrum scanning data, and user equipment signaling data. The spatiotemporal matrix dataset is input into a dual-channel feature extractor, wherein the dual-channel feature extractor includes a first channel and a second channel. The first channel is used to extract waveform features of the spectrum scan data, and the second channel is used to extract signal-to-noise ratio degradation trajectory features of the wireless measurement report. The waveform features and the signal-to-noise ratio degradation trajectory features are weighted and fused using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum. Interference identification and localization results are obtained based on the dynamic interference fingerprint spectrum, wherein the interference identification and localization results include interference type and spatial distribution hotspots representing the location of interference. Based on the interference identification and localization results, an optimal control strategy is generated using a pre-trained reinforcement learning decision model. The generation of the optimal control strategy includes generating a candidate control instruction set and then screening the candidate control instruction set after simulation verification in a digital twin network environment. The optimal control strategy is executed by a software-defined network controller, and the dual-channel feature extractor and the reinforcement learning decision model are updated based on a federated learning architecture.

[0141] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the wireless network interference localization and control method provided by the above methods. The method includes: acquiring multi-source heterogeneous data within a target area, performing structured preprocessing on the multi-source heterogeneous data, and generating a spatiotemporal matrix dataset with a unified timestamp; wherein the multi-source heterogeneous data includes wireless measurement reports, spectrum scan data, and user equipment signaling data. The spatiotemporal matrix dataset is input into a dual-channel feature extractor, wherein the dual-channel feature extractor includes a first channel and a second channel. The first channel is used to extract waveform features of the spectrum scan data, and the second channel is used to extract signal-to-noise ratio degradation trajectory features of the wireless measurement report. The waveform features and the signal-to-noise ratio degradation trajectory features are weighted and fused using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum. Interference identification and localization results are obtained based on the dynamic interference fingerprint spectrum, wherein the interference identification and localization results include interference type and spatial distribution hotspots representing the location of interference. Based on the interference identification and localization results, an optimal control strategy is generated using a pre-trained reinforcement learning decision model. The generation of the optimal control strategy includes generating a candidate control instruction set and then screening the candidate control instruction set after simulation verification in a digital twin network environment. The optimal control strategy is executed by a software-defined network controller, and the dual-channel feature extractor and the reinforcement learning decision model are updated based on a federated learning architecture.

[0142] 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.

[0143] 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.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.

Claims

1. A method for locating and controlling interference in wireless networks, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data within the target area, perform structured preprocessing on the multi-source heterogeneous data, and generate a spatiotemporal matrix dataset with a unified timestamp; wherein, the multi-source heterogeneous data includes wireless measurement reports, spectrum scan data, and user equipment signaling data; The spatiotemporal matrix dataset is input into a dual-channel feature extractor, wherein the dual-channel feature extractor includes a first channel and a second channel. The first channel is used to extract waveform features of the spectrum scan data, and the second channel is used to extract signal-to-noise ratio degradation trajectory features of the wireless measurement report. The waveform features and the signal-to-noise ratio degradation trajectory features are weighted and fused using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum. Interference identification and localization results are obtained based on the dynamic interference fingerprint spectrum, wherein the interference identification and localization results include interference type and spatial distribution hotspots representing the location of interference. Based on the interference identification and localization results, an optimal control strategy is generated using a pre-trained reinforcement learning decision model. The generation of the optimal control strategy includes generating a candidate control instruction set and then screening the candidate control instruction set after simulation verification in a digital twin network environment. The optimal control strategy is executed by a software-defined network controller, and the dual-channel feature extractor and the reinforcement learning decision model are updated based on a federated learning architecture.

2. The wireless network interference localization and control method according to claim 1, characterized in that, The multi-source heterogeneous data undergoes structured preprocessing, including: Calculate the quality weight of the data source in the multi-source heterogeneous data; wherein the quality weight is calculated based on the packet loss rate of the data source, the maximum latency deviation between the data source and the master clock, and the structural integrity score of the data source; The standardized observation vectors of the data source are weighted and summed using the quality weights to obtain the fused input data; The input data is processed for differential privacy using an asynchronous Laplace perturbation algorithm; wherein the differential privacy processing includes superimposing perturbation noise on data blocks located in user-sensitive areas, and the amplitude of the perturbation noise is inversely proportional to the quality weight.

3. The wireless network interference localization and control method according to claim 1, characterized in that, The step of using a dynamic confidence entropy weighting algorithm to weight and fuse the waveform features and the signal-to-noise ratio degradation trajectory features includes: The average information entropy of the predicted distribution of known interference categories by the first channel and the second channel within a preset time window is calculated to obtain the feature confidence entropy of the first channel and the feature confidence entropy of the second channel. Calculate the fusion weight of the first channel and the fusion weight of the second channel; wherein, the calculation rule for the fusion weight is: the lower the feature confidence entropy, the higher the corresponding fusion weight; The waveform features are weighted using the fusion weight of the first channel, and the signal-to-noise ratio degradation trajectory features are weighted using the fusion weight of the second channel. The weighted waveform features and the weighted signal-to-noise ratio degradation trajectory features are then concatenated to generate the dynamic interference fingerprint spectrum.

4. The wireless network interference localization and control method according to claim 1, characterized in that, The process of obtaining interference identification and localization results based on the dynamic interference fingerprint spectrum includes: The dynamic interference fingerprint spectrum is classified and predicted using the support vector set to obtain the prediction confidence. When the prediction confidence is lower than a set threshold, a boundary migration strategy is executed; wherein, the boundary migration strategy includes: calculating the direction vector of the centroid of the current sample and the support vector set, combining the direction vector and the prediction confidence to extrapolate and expand the boundary of the support vector set, and reclassifying using the expanded support vector set.

5. The wireless network interference localization and control method according to claim 1, characterized in that, The method of generating optimal control strategies using a pre-trained reinforcement learning decision model involves enhancing the state space of the reinforcement learning decision model, the enhancement including: Construct a disturbance factor estimation vector; wherein the construction of the disturbance factor estimation vector is based on the unit loss weight of the interference source, and the unit loss weight is determined by the normalized capacity of the affected cell in the grid neighborhood of the interference hot zone, the unit time throughput reduction of the affected cell, and the spectral overlap factor of the affected cell. The perturbation factor estimation vector is concatenated with the original network state vector to form a high-dimensional state representation, which is then used as input data into the reinforcement learning decision model.

6. The wireless network interference localization and control method according to claim 1, characterized in that, The training process of the reinforcement learning decision model employs a reward differential attribution regularization method, which includes: Calculate the gradient contribution of the action dimension output by the reinforcement learning decision model to the overall reward function; Calculate the difference between the historical average value of the action dimension and the gradient contribution, and use the difference as a difference constraint term; The final reward function is obtained by subtracting the difference constraint term from the original reward function, and the parameters of the reinforcement learning decision model are updated using the final reward function.

7. The wireless network interference localization and control method according to claim 1, characterized in that, The execution of the optimal control strategy through a software-defined network controller includes: The optimal control strategy is decomposed into base station-level executable instructions; wherein, the decomposition process includes: combining the service quality impact weight vector and the service quality improvement effect of atomic operations to select instructions that match the access network equipment standard; The base station-level executable instructions are issued, and the Shannon entropy of the feedback signal distribution after execution is monitored in real time. When the Shannon entropy increases, the trigger threshold for policy rollback is reduced.

8. The wireless network interference localization and control method according to claim 1, characterized in that, The update of the dual-channel feature extractor and the reinforcement learning decision model based on the federated learning architecture includes: Calculate the aggregate weight of the uploaded gradient of the edge node; wherein the aggregate weight is determined based on the standard deviation of the accuracy of the edge node in the local test, and the larger the standard deviation, the lower the aggregate weight; Calculate the local drift intensity; wherein, the local drift intensity is defined as the relative entropy between the local model prediction distribution in the current round and the local model prediction distribution in the previous round; When the local drift intensity exceeds a set threshold, an active synchronization process is triggered, and the gradient of the edge node is preferentially included in the central aggregation queue.

9. A wireless network interference location and control device, characterized in that, include: The generation module is used to acquire multi-source heterogeneous data within the target area, perform structured preprocessing on the multi-source heterogeneous data, and generate a spatiotemporal matrix dataset with a unified timestamp; wherein, the multi-source heterogeneous data includes wireless measurement reports, spectrum scan data, and user equipment signaling data; The input module is used to input the spatiotemporal matrix dataset into a dual-channel feature extractor, wherein the dual-channel feature extractor includes a first channel and a second channel, the first channel is used to extract the waveform features of the spectrum scan data, and the second channel is used to extract the signal-to-noise ratio degradation trajectory features of the wireless measurement report; The fusion module is used to perform weighted fusion of the waveform features and the signal-to-noise ratio degradation trajectory features using a dynamic confidence entropy weighting algorithm to generate a dynamic interference fingerprint spectrum, and obtain interference identification and localization results based on the dynamic interference fingerprint spectrum, wherein the interference identification and localization results include interference type and spatial distribution hotspots characterizing the interference location; The optimization module is used to generate the optimal control strategy based on the interference identification and localization results using a pre-trained reinforcement learning decision model. The generation of the optimal control strategy includes generating a candidate control instruction set and then screening the candidate control instruction set after simulation verification in a digital twin network environment. The execution module is used to execute the optimal control strategy through a software-defined network controller and update the dual-channel feature extractor and the reinforcement learning decision model based on the federated learning architecture.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the wireless network interference localization and control method according to any one of claims 1 to 8.

Images