An artificial intelligence and federated learning integrated distributed device group cooperative wireless channel sensing system and method
By integrating artificial intelligence and federated learning into a distributed device group collaborative wireless channel awareness system, the problems of privacy protection, communication overhead, and real-time performance of distributed device groups are solved, achieving efficient and secure channel awareness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUIZHOU TECHNICIAN COLLEGE (HUIZHOU SENIOR TECH SCHOOL)
- Filing Date
- 2026-03-02
- Publication Date
- 2026-06-16
Smart Images
Figure CN122227296A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless channel sensing technology, specifically a distributed device group collaborative wireless channel sensing system and method integrating artificial intelligence and federated learning. Background Technology
[0002] Wireless channel sensing, as a core supporting technology for ensuring communication quality, directly impacts the overall network performance through its sensing accuracy, processing efficiency, and operational security. Current mainstream wireless channel sensing technologies are mostly based on centralized or simple distributed architectures, which are difficult to match the collaborative communication needs of distributed device clusters, mainly due to the following technical problems: Traditional centralized channel sensing systems require each device to upload raw channel data such as CSI, signal power, and latency to a central server for unified processing and model training. This can easily lead to the leakage of sensitive device data and fails to meet the privacy protection requirement of data being usable but not visible. In particular, in multi-entity collaborative scenarios, the issues of cross-border data transmission and compliance become more prominent.
[0003] Distributed device clusters are widely distributed and numerous, generating massive amounts of raw channel data. Centralized transmission would consume a large amount of wireless spectrum resources, leading to a surge in communication overhead. At the same time, the link delay between data transmission and central processing cannot meet the real-time channel awareness requirements of high-speed mobile devices such as autonomous vehicles.
[0004] The deployment environments of different distributed devices vary, such as densely populated urban areas, suburbs, and indoors, resulting in significant heterogeneity in the collected channel data, such as different channel fading characteristics and interference types. Existing globally unified models are unable to accurately match the local channel characteristics of each device, leading to a decrease in sensing accuracy. Summary of the Invention
[0005] The purpose of this invention is to provide a distributed device group collaborative wireless channel sensing system and method that integrates artificial intelligence and federated learning, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a distributed device group collaborative wireless channel sensing system integrating artificial intelligence and federated learning, comprising a sensing prediction module, a data preprocessing module, a three-dimensional clustering module, a feature learning module, an encryption aggregation module, a resource optimization module, and a security protection module; Preferably, the perception prediction module is specifically as follows: Each distributed device has a built-in sensing and prediction unit that integrates a MIMO multi-antenna receiving module, a high-precision signal analysis unit, and a lightweight channel prediction submodule. The collected channel data covers multi-dimensional core features including channel state information, received signal strength, delay spread, and Doppler frequency shift, and simultaneously collects core network state parameters including the bandwidth utilization and packet loss rate of the device's access link. The integrated perception and prediction unit is equipped with a lightweight AI algorithm, which uses a two-layer shallow neural network combined with an attention mechanism to quickly complete the initial local channel perception, including channel type identification, interference signal judgment, and signal-to-noise ratio estimation. The prediction submodule uses an LSTM network to make short-term predictions of channel gain and interference intensity in the next 5 to 20 ms based on historical channel data and device movement trajectory. It generates a binary data packet of perception results and prediction trends, which is then transmitted to the data preprocessing module via the internal bus after being marked with device ID and timestamp. The system also integrates a load adaptive mechanism to detect the device's computing load and wireless link quality in real time, dynamically adjust the data acquisition frequency and feature acquisition dimensions, and automatically start the core feature priority acquisition mode when the link bandwidth utilization exceeds 70%, retaining key features and eliminating redundant dimensions.
[0007] Preferably, the data preprocessing module is as follows: Each distributed device has a dedicated data processing unit deployed locally. In accordance with the spectrum sensing requirements, the data processing is carried out according to the process of spectrum status judgment, data cleaning, feature alignment, enhancement and compression. A spectrum occupancy detection subunit is set up to quickly determine the current spectrum busyness based on the sub-Nyquist sampling principle; data cleaning adopts the 3σ criterion combined with the gradient anomaly detection algorithm to remove abnormal data caused by noise and electromagnetic interference; for channel mutation data in high-speed mobile scenarios, an adaptive adjustment mechanism for mutation threshold is set up; feature alignment is based on the network time synchronization protocol to calibrate the timestamps of each device, and a standard feature lookup table is used to achieve unified matching of channel feature dimensions between different devices. Data augmentation employs WGAN-GP generative adversarial network, focusing on expanding data in typical wireless communication scenarios, including high-speed mobility, strong interference, and scarce spectrum. Adaptive compression dynamically adjusts the compression strategy based on spectrum status judgment results. When the spectrum is busy, an efficient quantization coding algorithm is used to control the data accuracy loss to within 5% and increase the compression ratio to over 40%. When the spectrum is idle, the compression ratio is appropriately reduced to ensure data integrity, and the coding format is adapted to the federated learning parameter transmission format.
[0008] Preferably, the three-dimensional clustering module is as follows: A decentralized clustering strategy is adopted, and network load status parameters are introduced to construct a three-dimensional clustering system of channel characteristics, device capabilities, and network load; The three-dimensional feature extraction covers three core features: channel features, device capabilities, and network load. Among them, the network load parameter comes from the network status parameters collected by the perception and prediction module, such as base station load and spectrum utilization. Based on resource management requirements, the clustering weight of the three-dimensional features is dynamically allocated through the Voter-Model election mechanism. When the network load is too high, the weight of device capabilities is increased to prioritize the intra-cluster collaboration of low-latency devices. When the channel environment is complex, the weight of channel features is increased to ensure the homogeneity of channel features of devices in the same cluster.
[0009] The dynamic update of the cluster structure is triggered by the following conditions: the device location offset exceeds 50 meters, the channel feature similarity is lower than the threshold, and the network load fluctuation exceeds 30%. If any condition is met, re-clustering is triggered. The cluster head node integrates link quality monitoring function to coordinate the allocation of communication resources within the cluster in real time.
[0010] Preferably, the feature learning module is specifically as follows: Based on the standardized data output by the data preprocessing module ("spectrum judgment-cleaning-alignment-enhancement-compression"), and combined with the device clusters divided by the three-dimensional clustering module, a feature learning module is constructed to achieve deep extraction of channel features and model training. Specifically, a prediction feedback deep model that integrates CNN and LSTM is adopted. Spatial domain feature extraction is achieved through a CNN module, which employs a 3-layer convolutional layer combined with a 2-layer max pooling layer structure and introduces an attention mechanism to extract the spatial distribution differences and spatial interference distribution features of multi-antenna received signals. It also integrates scene features including base station coverage and terrain features. Temporal domain feature and prediction fusion is accomplished by an LSTM module, which uses a 2-layer hidden layer design with 128 hidden units to extract the channel fading time-series variation law and Doppler frequency shift state evolution trend. It also incorporates the short-term channel prediction results of the local prediction submodule and achieves deep fusion of historical sensing data and future prediction data through an attention gating mechanism. Federated training optimization is performed iteratively on the preprocessed local labeled dataset. L2 regularization constraint (regularization coefficient of 0.001) is introduced to ensure the compatibility between the local model and the global model within the cluster. A dynamic early stopping mechanism based on channel prediction error is adopted. By dynamically adjusting the stopping threshold, training is stopped when the validation set loss does not decrease for three consecutive rounds and the prediction error exceeds 2dB.
[0011] Preferably, the encryption aggregation module is specifically as follows: A collaborative strategy of layered adaptive encryption, link adaptive dynamic weighted aggregation, and asynchronous update is constructed to achieve a dynamic balance between security and efficiency. Layered adaptive encryption refers to the dynamic selection of encryption schemes based on link quality for model parameters trained locally on each device. When link bandwidth is sufficient and packet loss rate is less than 5%, Paillier homomorphic encryption algorithm is used for full-dimensional encryption. When link resources are scarce, it switches to lightweight BFV homomorphic encryption algorithm, which only performs high-strength encryption on core parameters and uses symmetric encryption for non-core parameters. The encryption key is generated locally on the device and shared only with the cluster head node currently elected by the 3D clustering module. When the cluster head node is updated, the key is synchronized to the new cluster head node through a secure key handover mechanism with AES-256 encryption. Before the handover, the identity of the new cluster head must be verified by SHA-256 hash. Dynamic weighted aggregation optimizes the aggregation weight by the cluster head node based on communication scenario requirements. Data quality score, model local verification accuracy, and link contribution jointly determine the aggregation weight, generating global model parameters within the cluster. Cross-cluster parameter fusion is achieved through a secure multi-party computation protocol between clusters to build a system-level global channel awareness model. The asynchronous update mechanism supports parameter retransmission after offline devices reconnect. It is designed with parameter caching and incremental update mechanisms. During the device's offline period, the cluster head node caches the global model incremental parameters. After the device reconnects, only the difference between the local model and the cached incremental parameters needs to be transmitted, reducing the communication overhead during reconnection.
[0012] Preferably, the resource optimization module is specifically as follows: Channel sensing and prediction results are directly used to drive radio resource scheduling, forming a closed-loop optimization across the entire link. The refined channel assessment receives the channel state assessment results output by the global model, including channel capacity, interference type and intensity, link stability level, and transmission delay budget. It also integrates short-term channel prediction results from the perception and prediction module to generate a complete channel assessment report showing the current state and future trends. Dynamic resource scheduling, based on the assessment report and in accordance with wireless resource management specifications, uses a genetic algorithm to optimize spectrum resource allocation. Simultaneously, based on the Doppler frequency shift trend predicted by the channel, it adjusts the transmission power and modulation / coding scheme in advance. When high-speed mobile scenarios are predicted, LDPC coding is prioritized to ensure low latency; when strong interference scenarios are predicted, Turbo coding is switched to. The feedback adjustment mechanism feeds back the scheduling results to the perception prediction module and the 3D clustering module through a low-latency control channel, dynamically adjusting the data acquisition frequency, clustering weights, and model training parameters.
[0013] Preferably, the security protection module is as follows: By integrating homomorphic encryption, differential privacy, blockchain, and wireless anti-interference technologies, a full-process security protection system adapted to wireless communication scenarios is formed. Identity authentication and behavior tracing adopt a consortium blockchain architecture, implementing identity authentication and operation behavior tracing for all participating nodes, including the perception and prediction module and the data preprocessing module. When a node joins, it must complete key authentication and permission registration. All interactive operations generate immutable blockchain transaction records. At the same time, the blockchain data adopts a lightweight storage solution to adapt to edge devices. Anti-interference and data verification are achieved by combining real-time power spectral density analysis with a deep learning interference identification model. Interference detection is performed on the received signal every 10ms to identify malicious interference and data poisoning attacks, triggering channel switching or power enhancement strategies. At the same time, the perception data output by the perception and prediction module and the model parameters transmitted by the encryption aggregation module are verified for integrity using the SHA-256 hash verification algorithm. If the verification fails, data access or parameter aggregation is rejected. The privacy protection mechanism introduces an adaptive ε-differential privacy mechanism during the parameter aggregation process of the encrypted aggregation module. The ε value is dynamically adjusted according to the wireless link quality, and controllable Laplace noise is injected. Combined with the local training characteristics of federated learning, it achieves the dual goals of data usability without visibility and privacy protection.
[0014] Malicious interference response strategy: Upon detecting malicious interference, prioritize switching to backup channels. There are 3-5 backup channels pre-selected, ordered by signal strength. If no backup channel is available, activate the power enhancement strategy to increase transmission power by 5-10dBm, not exceeding the 50dBm limit. Data poisoning attack response: After identifying abnormal parameters, reduce the aggregation weight of the device to 50% of its original weight. If abnormal parameters are uploaded twice consecutively, temporarily remove the device from the aggregation privileges. Re-evaluate after 10 minutes.
[0015] The consortium blockchain nodes include core cluster head nodes (consensus nodes) and ordinary device nodes (ledger nodes). The consensus mechanism adopts PBFT (Practical Byzantine Fault Tolerance), and the block generation interval is 2 seconds. The lightweight storage solution adopts "transaction data sharding + hash digest storage", storing only the transaction digests of its own related transactions on ordinary device nodes, while the complete data is stored on the core cluster head nodes.
[0016] The scaling parameter of Laplace noise corresponds to the value of ε: when ε=0.1, the scaling parameter is 0.02; when ε=0.05, the scaling parameter is 0.04. Noise is only injected into non-critical bits of the model parameters, such as the third decimal place and beyond.
[0017] This invention also provides a distributed device group cooperative wireless channel sensing method integrating artificial intelligence and federated learning. Based on the above system, the specific steps are as follows: Perception preprocessing stage: Each device collects core channel features and network status parameters through its built-in sensing unit, completes local preliminary perception using a lightweight AI algorithm, and achieves short-term channel prediction by combining an LSTM model; the generated perception results and predicted trend binary data packets are standardized according to the process of spectrum judgment, cleaning, alignment, enhancement, and compression. The standardization process uses Z-score normalization to calculate the mean μ and standard deviation σ for each channel feature dimension. μ and σ are derived from nearly 1,000 historical valid data points on the device. The processing formula is simplified to standardized value = (original value - μ) / σ, ensuring that the data distribution meets the model input requirements (mean ≈ 0, standard deviation ≈ 1).
[0018] Three-dimensional clustering stage: A decentralized strategy is adopted to extract three-dimensional features of channel features, device capabilities, and network load, dynamically allocate clustering weights to achieve accurate device clustering, complete the adaptive update of cluster structure based on multiple triggering conditions, and coordinate intra-cluster communication resources by the cluster head node; Fusion learning phase: Based on preprocessed data and device cluster division, channel spatiotemporal dynamic features are extracted through CNN-LSTM fusion model, local model training is completed by combining federated learning, and regularization constraints and dynamic early stopping mechanism are adopted to ensure model compatibility and training efficiency. The parameter interaction frequency between the local model and the global model within the cluster is "upload parameters once every 20 rounds of local iteration". After receiving all parameters from devices within the cluster, the cluster head node completes weighted aggregation within 100ms and feeds back the updated incremental parameters of the global model. After receiving the parameters, the local model only updates the differences, without needing to cover all parameters, thus shortening the update time.
[0019] In the aggregation and optimization phase: encryption schemes are dynamically selected based on link quality, cluster head nodes are weighted and aggregated to generate a global model, and cross-cluster fusion is used to build a system-level model; spectrum resource scheduling is driven based on global perception results, and collection, clustering and training parameters are dynamically optimized through feedback adjustment, while a full-process security protection mechanism is integrated to ensure data and collaboration security.
[0020] The key configurations of the genetic algorithm are: population size = 50, maximum number of iterations = 100, selection strategy is roulette wheel selection, crossover probability = 0.8, mutation probability = 0.05; the quantitative standards for the channel evaluation report are: link stability level is divided into 5 levels, level 1: latency fluctuation ≤ 5ms; level 5: latency fluctuation ≥ 20ms; channel capacity is divided into "≤ 10Mbps (low), 10-50Mbps (medium), ≥ 50Mbps (high)"; interference intensity is defined as "≤ -80dBm (weak), -80~-60dBm (medium), ≥ -60dBm (strong)".
[0021] The beneficial effects of this invention are as follows: 1. This invention utilizes a federated learning architecture, enabling devices to complete model training locally, transmitting only encrypted model parameters rather than the original data. The layered adaptive encryption strategy of the encryption aggregation module combines Paillier homomorphic encryption and lightweight encryption technology to dynamically adapt to different link states; the security protection module incorporates an adaptive ε-differential privacy mechanism and consortium blockchain identity authentication, ensuring data integrity through hash verification and enabling traceability of operational behavior; it avoids the risks of cross-border data transmission and sensitive information leakage, fully meeting the privacy protection needs and compliance requirements in multi-entity collaborative scenarios.
[0022] 2. This invention improves operational efficiency through multi-level optimization. The data preprocessing module adopts WGAN-GP data augmentation and adaptive compression strategies, which can increase the compression rate to over 40% even when the spectrum is busy, while keeping the accuracy loss within 5%. The three-dimensional clustering module clusters devices in a decentralized manner, with cluster head nodes coordinating resources to reduce cross-device communication conflicts. The asynchronous update mechanism reduces the communication overhead when offline devices reconnect through parameter caching and incremental transmission. The local processing and lightweight design of each module reduce the amount of data transmission. Combined with a low-latency feedback adjustment mechanism, it can meet the real-time requirements of channel perception for high-speed mobile devices such as autonomous driving.
[0023] 3. This invention constructs a collaborative mechanism for three-dimensional clustering and fusion learning. The three-dimensional clustering module extracts three core features: channel characteristics, device capabilities, and network load. By dynamically allocating weights, it achieves accurate device clustering, adapting to different scenarios such as densely populated urban areas and high-speed mobile environments. The feature learning module adopts a CNN-LSTM fusion model, which can extract spatial domain interference distribution features and capture the dynamic laws of channel fading in the time domain, while incorporating short-term prediction results to enhance predictive capabilities. During federated training, L2 regularization and dynamic early stopping mechanisms are introduced to ensure the compatibility of local models and global models within clusters. This enables the model to accurately adapt to heterogeneous data characteristics, effectively improving the perception accuracy of key parameters such as channel capacity and interference intensity, and enhancing the system's adaptability to complex dynamic channel environments. Attached Figure Description
[0024] Figure 1 This is a flowchart of the overall system of the present invention; Figure 2 This is a flowchart of the data processing and feature learning process of this invention; Figure 3 This is a flowchart of the encryption aggregation and resource optimization process of this invention. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] like Figures 1 to 3 As shown, this embodiment of the invention provides a distributed device group collaborative wireless channel sensing system integrating artificial intelligence and federated learning, including a sensing prediction module, a data preprocessing module, a three-dimensional clustering module, a feature learning module, an encryption aggregation module, a resource optimization module, and a security protection module. The specific implementation of each module is as follows: The perception and prediction module is specifically as follows: Each distributed device has a built-in integrated sensing and prediction unit, which integrates a MIMO multi-antenna receiving module, a high-precision signal analysis unit, and a lightweight channel prediction submodule. The channel data collected covers multi-dimensional core features such as channel state information (CSI), received signal strength (RSSI), delay spread, and Doppler shift. It also collects dynamic operating parameters such as the bandwidth utilization rate of the device's access link, the current actual packet loss rate, and the base station load. The device's capabilities include inherent attribute parameters such as computing power level, maximum supported bandwidth of the communication link, data acquisition hardware accuracy, and upper limit of packet loss rate design. The integrated perception and prediction unit is equipped with a lightweight AI algorithm, which uses a two-layer shallow neural network combined with an attention mechanism to quickly complete the initial local channel perception, including channel type identification, interference signal judgment, and signal-to-noise ratio estimation. The prediction submodule uses an LSTM network to make short-term predictions of channel gain and interference intensity in the next 5 to 20 ms based on historical channel data and device movement trajectory, such as GPS information of vehicle-mounted devices. It generates a binary data packet of perception results and prediction trends, which is then transmitted to the data preprocessing module via the internal bus after being marked with device ID and timestamp. The system also integrates a load adaptive mechanism to detect the device's computing load and wireless link quality in real time, dynamically adjust the data acquisition frequency (adjustable from 50 to 200 Hz) and feature acquisition dimensions. When the link bandwidth occupancy rate exceeds 70%, the system automatically starts the core feature priority acquisition mode, retains key features such as CSI and RSSI, and eliminates redundant dimensions to ensure perception continuity in resource-constrained scenarios.
[0027] Computing load detection is determined by the device's CPU utilization (threshold set to 70%) and memory utilization (threshold set to 80%), while link quality is comprehensively evaluated by real-time bandwidth, packet loss rate, and latency (threshold set to 50ms). The core features are defined as three categories: Channel State Information (CSI), Received Signal Strength (RSSI), and Doppler frequency shift. Redundant features include the number of secondary interference sources in the channel and the repeated marking of historical data. The core feature collection dimension is fixed at 6 dimensions, while the redundant feature dimension can be dynamically reduced by 1-3 dimensions according to the link status.
[0028] In the lightweight AI algorithm's two-layer shallow neural network, the input layer has 32 neurons, matching the concatenation dimension of 6-dimensional core channel features + 4-dimensional network state parameters; the hidden layer has 16 neurons, and the activation function is ReLU; the attention mechanism adopts single-head self-attention, with an attention head dimension of 16, used to strengthen the weights of signal-to-noise ratio related features.
[0029] The LSTM prediction submodule has an input dimension of (time step = 20, feature dimension = 10) and an output dimension of 2, corresponding to the predicted channel gain and interference intensity, respectively. The optimizer is Adam, with an initial learning rate of 0.001, a training batch size of 32, and an upper limit of 100 iterations. The activation functions for the forget gate, input gate, and output gate are all sigmoid, and the cell state activation function is tanh.
[0030] The data preprocessing module is specifically as follows: Each distributed device deploys a dedicated data processing unit locally. In accordance with the spectrum sensing requirements, it performs processing according to the process of spectrum status judgment, data cleaning, feature alignment, enhancement, and compression, so as to achieve deep collaboration between preprocessing and wireless spectrum resource optimization. A spectrum occupancy detection subunit is set up to quickly determine the current spectrum busyness based on the sub-Nyquist sampling principle; data cleaning adopts the 3σ criterion combined with the gradient anomaly detection algorithm to accurately remove abnormal data caused by noise and electromagnetic interference; for channel mutation data in high-speed mobile scenarios, an adaptive adjustment mechanism for mutation threshold is set up; feature alignment is based on the network time synchronization protocol to calibrate the timestamps of each device, and a standard feature lookup table is used to achieve unified matching of channel feature dimensions between different devices to ensure compatibility with wireless communication network parameters; Data augmentation employs WGAN-GP generative adversarial networks, focusing on expanding data for typical wireless communication scenarios such as high-speed mobility, strong interference, and scarce spectrum, thereby improving the model's adaptability to dynamic channel environments. Adaptive compression dynamically adjusts the compression strategy based on spectrum status judgment results. When the spectrum is busy, an efficient quantization coding algorithm is used to control data accuracy loss to within 5%, increasing the compression ratio to over 40%. When the spectrum is idle, the compression ratio is appropriately reduced to ensure data integrity, and the coding format is adapted to the federated learning parameter transmission format, reducing transmission adaptation overhead. In this invention, the compression ratio refers to the ratio of the compressed data volume to the original data volume, and the compression ratio refers to the ratio of the original data volume to the compressed data volume.
[0031] The threshold for determining spectrum activity level is as follows: spectrum occupancy rate ≥60% is busy, ≤30% is idle, and 30%-60% is moderate. The efficient quantization coding algorithm adopts adaptive uniform quantization coding, and the number of quantization bits is adjusted according to the spectrum status: 8 bits when busy and 16 bits when idle. The encoded data format is a binary stream, which is seamlessly compatible with the protobuf format for federated learning parameter transmission and requires no additional format conversion.
[0032] The generator of WGAN-GP uses a 3-layer fully connected network. The input is a 10-dimensional random noise vector, and the number of hidden layer neurons is 64 and 128 respectively. The output layer dimension is consistent with the dimension of the preprocessed data, which is 12-dimensional. The activation function is LeakyReLU with a negative slope of 0.2. The discriminator is also a 3-layer fully connected network with an input dimension of 12, 128 and 64 hidden neurons respectively, and a 1-dimensional discriminant value as the output. The activation function is LeakyReLU. The number of training epochs is 50, the batch size is 64, the gradient penalty coefficient is 10, the learning rate of both the generator and the discriminator is 0.0001, and the optimizer uses RMSProp.
[0033] The three-dimensional clustering module is specifically as follows: A decentralized clustering strategy is adopted, and network load status parameters are introduced to construct a three-dimensional clustering system of channel characteristics, device capabilities, and network load, so as to achieve accurate clustering and adaptive updating of dynamic device clusters. Three-dimensional feature extraction covers channel features and network status parameters collected by the perception and prediction module, as well as the device's own capability parameters; Channel characteristics include fading type and interference level; equipment capabilities include computing power level, communication link bandwidth and packet loss rate, and data acquisition quality; network load includes base station load and spectrum utilization in the area. Based on resource management requirements, clustering weights for 3D features are dynamically allocated through a Voter-Model election mechanism. The weight allocation logic is as follows: When network load is too high: device capacity weight 40%, channel characteristic weight 30%, network load weight 30%; When the channel environment is complex: channel characteristics weight 45%, equipment capability weight 35%, network load weight 20%.
[0034] In addition to triggering re-clustering when the device location offset exceeds 50 meters or the channel feature similarity is below the threshold, the dynamic update of the cluster structure also includes network load fluctuations exceeding 30% as a re-clustering trigger condition. The cluster head node integrates link quality monitoring functions to coordinate the allocation of communication resources within the cluster in real time and avoid communication conflicts within the cluster.
[0035] The decentralized clustering algorithm uses an improved K-means algorithm. The initial number of clusters is dynamically set according to "20-30 devices per cluster", and the maximum number of clusters does not exceed 10. During the clustering process, each device completes the cluster affiliation judgment by calculating the three-dimensional feature similarity (cosine similarity) with the neighboring devices in the cluster locally. No central node coordination is required. After the cluster affiliation is confirmed, it is synchronized to all devices in the cluster through a broadcast mechanism.
[0036] The Voter-Model election mechanism is implemented as follows: each cluster has no fewer than 5 participating voting devices. The devices vote based on their own 3D feature acquisition accuracy and link stability. The weight allocation scheme with a vote rate of more than 60% takes effect. The channel feature similarity threshold is set to 0.7, and cosine similarity is used for calculation. If the value is lower than this, re-clustering is triggered.
[0037] The cluster head node election criteria are as follows: device computing power level ≥ 8 (levels 1-10, with level 10 being the highest), link transmission success rate ≥ 95% in the past 10 seconds, and packet loss rate ≤ 3%. Among the devices that meet the criteria, the one with the highest integrity of channel feature collection is selected as the cluster head.
[0038] The computing power level of the equipment is divided into 1-10 levels (level 10 is the highest) based on the processor speed and memory bandwidth. The standard for electing cluster head nodes is a computing power level ≥ 8.
[0039] The feature learning module is specifically as follows: Based on the standardized data output by the data preprocessing module, and combined with the device clusters divided by the three-dimensional clustering module, a feature learning module is constructed to achieve deep extraction of channel features and model training. Specifically, a prediction feedback deep model that integrates CNN and LSTM is adopted to enhance the extraction and forward-looking perception of the spatiotemporal dynamic features of the wireless channel and improve the perception accuracy of the model in dynamic wireless environments. Spatial domain feature extraction is achieved through a CNN module, employing a 3-layer convolutional layer combined with a 2-layer max pooling layer structure and introducing an attention mechanism. It focuses on extracting spatial distribution differences and spatial interference distribution features of multi-antenna received signals, while also integrating scene features such as base station coverage and terrain features to improve the targeting of spatial feature extraction. Temporal domain feature and prediction fusion is accomplished by an LSTM module, which adopts a 2-layer hidden layer design with 128 hidden units. It not only extracts the channel fading time series variation law and Doppler frequency shift state evolution trend, but also integrates the short-term channel prediction results of the local prediction submodule. Through the attention gating mechanism, it achieves deep fusion of historical sensing data and future prediction data, improving the model's ability to predict dynamic channel changes. Federated training optimization is performed iteratively on the preprocessed locally labeled dataset. L2 regularization constraints are introduced into federated learning, with a regularization coefficient of 0.001 to ensure compatibility between the local model and the global model within the cluster. A dynamic early stopping mechanism based on channel prediction error is employed. By dynamically adjusting the stopping threshold, training stops when the validation set loss does not decrease for three consecutive rounds and the prediction error exceeds 2dB, improving training efficiency and model usability. After data preprocessing, annotations are automatically added to the preprocessed data based on the scene label library stored locally on the device (such as high-speed moving / stationary scene labels, interference type labels, etc.) or real-time scene recognition results, forming a locally labeled dataset for subsequent federated training.
[0040] The maximum number of iterations for local training is 200 rounds. After each iteration, the validation set loss is calculated. In federated learning, the aggregation frequency of the global model within the cluster is once every 10 rounds of local iteration. Weighted average is used during aggregation, and the weight allocation is consistent with the encrypted aggregation module. The initial parameters of the local model are initialized using Xavier, and the parameters of the global model are updated using momentum gradient descent with a momentum coefficient of 0.9.
[0041] In the CNN module, the kernel sizes of the three convolutional layers are 3×3, 3×3, and 2×2, respectively, with a stride of 1 and a padding method of "same". The kernel size of the two max pooling layers is 2×2, with a stride of 2. The attention mechanism adopts spatial attention, which generates an attention weight map based on the average pooling of the channel dimension of the feature map.
[0042] Scene features (base station coverage quantized into 3 categories, terrain features quantized into 4 categories) are converted into 7-dimensional vectors through one-hot encoding, and then concatenated with the 64-dimensional spatial features extracted by CNN as the input to LSTM. The LSTM module has an input dimension of 71 (64-dimensional spatial features + 7-dimensional scene features) and an output dimension of 32 (temporal domain features). The fusion of CNN and LSTM involves feature concatenation, resulting in a 64-dimensional + 32-dimensional = 96-dimensional feature vector, which is then output through a single fully connected layer. The batch size for local training is 64, the loss function is mean squared error, and the initial dynamic early stopping threshold is 1.5dB.
[0043] The encryption aggregation module is specifically as follows: A collaborative strategy of layered adaptive encryption, link adaptive dynamic weighted aggregation, and asynchronous update is constructed to achieve a dynamic balance between security and efficiency. Layered adaptive encryption refers to the dynamic selection of encryption schemes based on link quality for model parameters trained locally on each device. When the link bandwidth is greater than 10Mbps and the packet loss rate is less than 5%, the Paillier homomorphic encryption algorithm is used for full-dimensional encryption. When link resources are scarce, a lightweight homomorphic encryption algorithm is switched to, which performs high-strength encryption only on core parameters and uses symmetric encryption for non-core parameters. The encryption key is generated locally on the device and is shared only with the cluster head node currently elected by the 3D clustering module. When the cluster head node is updated, the key is synchronized to the new cluster head node through a secure key handover mechanism. Dynamic weighted aggregation is optimized by the cluster head node based on communication scenario requirements. The aggregation weight is jointly determined by data quality score (accounting for 50%), model local validation accuracy (accounting for 30%), and link contribution (accounting for 20%, i.e. the link stability of device uploaded parameters), generating global model parameters within the cluster. Cross-cluster parameter fusion is achieved through a secure multi-party computation protocol between clusters to build a system-level global channel awareness model. The asynchronous update mechanism supports parameter retransmission after offline devices reconnect. It is designed with parameter caching and incremental update mechanisms. During device offline, the cluster head node caches global model incremental parameters. After the device reconnects, only the difference between the local model and the cached incremental parameters needs to be transmitted, reducing the communication overhead during reconnection and adapting to the dynamic access requirements of mobile devices.
[0044] The secure key handover mechanism is as follows: After the new cluster head completes its identity verification through the consortium blockchain, it sends a key request to the original cluster head. The original cluster head verifies the identity of the new cluster head using SHA-256 hash. After successful verification, it uses AES-256 encryption to encrypt the key and transmits it through a dedicated secure channel. The lightweight homomorphic encryption algorithm selected is the BFV algorithm. The core parameters are defined as key parameters that affect the prediction accuracy, such as model weights and bias terms, accounting for about 30%. Non-core parameters include auxiliary information such as the number of training iterations and local loss values.
[0045] Data quality score is calculated by weighting feature completeness, percentage of outlier data, and timestamp consistency, with a maximum score of 100. Data scoring 80 or above is considered high-quality. The accuracy of local model validation is quantified by the inverse of the mean square error between the predicted value and the actual channel parameters. The link contribution is calculated by the average delay and transmission success rate of the last 5 parameter transmissions.
[0046] Cross-cluster integration adopts the SPDZ secure multi-party computation protocol, with each cluster head node participating. The communication process uses lightweight symmetric encryption (AES-128) to ensure transmission security.
[0047] The resource optimization module is specifically as follows: Channel sensing and prediction results are directly used to drive radio resource scheduling, forming a closed-loop optimization across the entire link. The refined channel assessment receives the channel state assessment results output by the global model, including channel capacity, interference type and intensity, link stability level, and transmission delay budget. It also integrates short-term channel prediction results from the perception and prediction module to generate a complete channel assessment report showing the current state and future trends. Dynamic resource scheduling, based on the assessment report and in accordance with wireless resource management specifications, uses a genetic algorithm to optimize spectrum resource allocation, achieving efficient reuse of spectrum resources for devices in the same cluster divided by the three-dimensional clustering module. Simultaneously, based on the Doppler frequency shift trend predicted by the channel, it adjusts the transmission power (adjustable from 10 to 50 dBm) and modulation / coding scheme in advance. When high-speed mobile scenarios are predicted, LDPC coding is prioritized to ensure low latency; when strong interference scenarios are predicted, Turbo coding is switched to improve reliability. The feedback adjustment mechanism feeds back the scheduling results to the perception prediction module and the 3D clustering module through a low-latency control channel, dynamically adjusting the data acquisition frequency, clustering weights, and model training parameters.
[0048] The security protection module is specifically as follows: By integrating homomorphic encryption, differential privacy, blockchain, and wireless anti-interference technologies, a full-process security protection system adapted to wireless communication scenarios is formed. Identity authentication and behavior tracing adopt a consortium blockchain architecture, implementing identity authentication and operation behavior tracing for all participating nodes, including the perception and prediction module and the data preprocessing module. When a node joins, it must complete key authentication and permission registration. All interactive operations generate tamper-proof blockchain transaction records. At the same time, the blockchain data adopts a lightweight storage solution adapted to edge devices. Anti-interference and data verification are achieved through real-time power spectral density analysis combined with a deep learning interference identification model. The interference detection frequency is 50 to 200 Hz, consistent with the data acquisition frequency, accurately identifying malicious interference and data poisoning attacks, triggering channel switching or power enhancement strategies. At the same time, the perception data output by the perception and prediction module and the model parameters transmitted by the encryption aggregation module are verified for integrity using the SHA-256 hash verification algorithm. If the verification fails, data access or parameter aggregation is rejected. The privacy protection mechanism introduces an adaptive ε-differential privacy mechanism during the parameter aggregation process of the encrypted aggregation module. The ε value is dynamically adjusted according to the wireless link quality. When the link is secure, ε=0.1 and when the link is risky, ε=0.05. Controllable Laplace noise is injected, and combined with the local training characteristics of federated learning, the dual goals of data "usable but not visible" and privacy protection are achieved.
[0049] The deep learning interference identification model adopts a lightweight CNN architecture. The input is a 1×64-dimensional power spectral density sampling vector with 64 sampling points. The network consists of 2 convolutional layers (3×1 kernels, stride 1, padding with "same"), 1 max pooling layer (2×1 pooling kernels, stride 2), and 1 fully connected layer (32 neurons). The output is 3 types of interference labels: no interference, malicious interference, and electromagnetic interference. The activation function is Softmax.
[0050] The training data came from 1000 labeled power spectral density samples (300 samples per class of interference + 100 samples without interference). The optimizer was Adam (learning rate = 0.0005), the batch size was 32, the number of training epochs was 80, and the loss function was cross-entropy loss.
[0051] This invention also provides a distributed device group cooperative wireless channel sensing method integrating artificial intelligence and federated learning. Based on the above system, the specific steps are as follows: Perception preprocessing stage: Each device collects core channel features and network status parameters through its built-in sensing unit, completes local preliminary perception using a lightweight AI algorithm, and achieves short-term channel prediction by combining an LSTM model; the generated perception results and predicted trend binary data packets are standardized according to the process of spectrum judgment, cleaning, alignment, enhancement, and compression. Three-dimensional clustering stage: A decentralized strategy is adopted to extract three-dimensional features of channel features, device capabilities, and network load, dynamically allocate clustering weights to achieve accurate device clustering, complete the adaptive update of cluster structure based on multiple triggering conditions, and coordinate intra-cluster communication resources by the cluster head node; Fusion learning phase: Based on preprocessed data and device cluster division, channel spatiotemporal dynamic features are extracted through CNN-LSTM fusion model, local model training is completed by combining federated learning, and regularization constraints and dynamic early stopping mechanism are adopted to ensure model compatibility and training efficiency. In the aggregation and optimization phase: encryption schemes are dynamically selected based on link quality, cluster head nodes are weighted and aggregated to generate a global model, and cross-cluster fusion is used to build a system-level model; spectrum resource scheduling is driven based on global perception results, and collection, clustering and training parameters are dynamically optimized through feedback adjustment, while a full-process security protection mechanism is integrated to ensure data and collaboration security.
[0052] Cross-cluster fusion process: Each cluster head node first encrypts and uploads the global model parameters within the cluster to the consortium blockchain consensus node. The consensus node then performs weighted fusion according to the cluster size weight (60%) and the model accuracy weight (40%) to generate a system-level global model. The global model is updated once every 3 rounds of intra-cluster aggregation, triggering cross-cluster fusion. The updated global model is distributed to all devices within the cluster through the cluster head node, and the devices complete local model calibration within 200ms after receiving the model.
[0053] Feedback adjustment quantification rules: ① Frequency adjustment: Set to 200Hz when the channel assessment is "high stability" (link stability level 1-2), 150Hz when it is "medium stability" (level 3), and 50-100Hz when it is "low stability" (level 4-5); ② Clustering weight adjustment: When spectrum resources are scarce, the weight of equipment capability is increased by 10%, and the weight of network load is decreased by 10%; ③ Model training parameter adjustment: When the channel prediction error is ≤1dB, the regularization coefficient is maintained at 0.001; when the error is 1-2dB, it is adjusted to 0.002; and when the error is >2dB, it is adjusted to 0.003.
[0054] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0055] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A distributed device-group cooperative wireless channel sensing system integrating artificial intelligence and federated learning, characterized in that, It includes a perception and prediction module, a data preprocessing module, a 3D clustering module, a feature learning module, an encryption aggregation module, a resource optimization module, and a security protection module; Perception and Prediction Module: The distributed device has a built-in integrated perception and prediction unit that collects core channel features and network status parameters, completes local channel perception through AI algorithms, uses a prediction model to achieve short-term channel status prediction, and generates binary data packets to be transmitted to the data preprocessing module. Data preprocessing module: Receives data packets output by the perception and prediction module, completes local processing according to a standardized process, ensures compatibility through unified feature matching, and adopts data augmentation and adaptive compression strategies to achieve synergy between preprocessing and spectrum resource optimization; 3D clustering module: It adopts a decentralized strategy to construct a 3D clustering system, extracts three core features: channel features, device capabilities, and network load, dynamically allocates clustering weights to adapt to different scenario requirements, realizes adaptive updates of cluster structure based on multiple triggering conditions, and coordinates intra-cluster communication resources by cluster head nodes. Feature learning module: Based on the preprocessed data and device cluster partitioning results, a fusion-type deep model is used to extract the spatiotemporal dynamic features of the wireless channel. The fusion of scenario features improves the extraction targeting, and the integration of short-term prediction results enhances the ability to predict channel changes. Federated training optimizes and ensures the compatibility of local and global models. Encryption Aggregation Module: Constructs a hierarchical adaptive encryption, link adaptive dynamic weighted aggregation and asynchronous update collaborative strategy, dynamically selects encryption schemes based on link quality, optimizes aggregation weights at the cluster head node to generate global parameters within the cluster, and constructs a system-level perception model through cross-cluster fusion to adapt to the dynamic access requirements of mobile devices; Resource optimization module: Based on the global channel awareness model, it generates a channel assessment report by fusing awareness prediction results, uses optimization algorithms to achieve efficient reuse of spectrum resources, adjusts transmission parameters in combination with channel prediction trends, and feeds back the scheduling results to the aforementioned modules through a feedback adjustment mechanism to dynamically adjust relevant parameters; Security Protection Module: Constructs a full-process protection system to achieve node identity authentication and behavior traceability. Through interference detection, data verification, and privacy protection mechanisms, it ensures the safe and reliable collaboration of all modules.
2. The distributed device group cooperative wireless channel sensing system integrating artificial intelligence and federated learning as described in claim 1, characterized in that, In the perception and prediction module, the integrated perception and prediction unit integrates a multi-antenna receiving module, a signal analysis unit, and a lightweight channel prediction submodule to collect multi-dimensional channel core features and network status parameters; it uses a lightweight AI algorithm to complete local preliminary channel perception; the prediction submodule uses a time-series prediction network to achieve short-term channel status prediction and generates binary data packets with device IDs and timestamps, which are then transmitted to the data preprocessing module; the system integrates a load adaptive mechanism to detect device computing power and link quality in real time and dynamically adjust the data acquisition frequency and feature dimensions.
3. The distributed device group cooperative wireless channel sensing system integrating artificial intelligence and federated learning as described in claim 2, characterized in that, In the data preprocessing module, each distributed device locally deploys a dedicated data processing unit to perform processing according to the process of spectrum status judgment, data cleaning, feature alignment, enhancement, and compression. The spectrum occupancy detection subunit determines the spectrum busyness, anomaly detection algorithm removes interference and abnormal data, and time stamp calibration is based on time synchronization protocol to achieve unified feature matching. WGAN-GP generative adversarial network is used for data augmentation, and adaptive compression strategy is dynamically adjusted based on spectrum status. The encoding format is adapted to the transmission requirements of federated learning parameters.
4. The distributed device group cooperative wireless channel sensing system integrating artificial intelligence and federated learning as described in claim 3, characterized in that, In the three-dimensional clustering module, the three-dimensional feature extraction covers three core features: channel features, device capabilities, and network load. Based on resource management requirements, the clustering weights of the three features are dynamically allocated through an election mechanism to adapt to different scenario requirements. The dynamic update triggering conditions for the cluster structure include device location offset, changes in channel feature similarity, and network load mutations exceeding 30%. The cluster head node coordinates the allocation of communication resources within the cluster in real time.
5. A distributed device group cooperative wireless channel sensing system integrating artificial intelligence and federated learning as described in claim 4, characterized in that, The feature learning module adopts a prediction feedback deep model that integrates convolutional neural networks and temporal prediction networks. It extracts channel spatial domain features and integrates scene features through convolutional neural networks, and extracts channel temporal domain features and integrates short-term prediction results through temporal prediction networks, thereby achieving deep integration of historical and prediction data. Federated training optimization introduces regularization constraints to ensure model compatibility, and adopts a dynamic early stopping mechanism based on channel prediction error to improve training efficiency.
6. A distributed device group cooperative wireless channel sensing system integrating artificial intelligence and federated learning as described in claim 5, characterized in that, In the encryption aggregation module, the hierarchical adaptive encryption dynamically selects the encryption scheme based on the link quality. The encryption key is shared only with the current cluster head node, and the key handover is completed through a secure mechanism when the cluster head is updated. The cluster head node optimizes the aggregation weights based on data quality, model accuracy, and link contribution to generate global parameters within the cluster. Cross-cluster fusion is achieved through secure multi-party computation between clusters. The asynchronous update mechanism supports offline devices to reconnect and retransmit parameters, and reduces communication overhead through parameter caching and incremental updates.
7. A distributed device group cooperative wireless channel sensing system integrating artificial intelligence and federated learning as described in claim 6, characterized in that, In the resource optimization module, the refined channel assessment integrates the global model assessment results and short-term prediction results to generate a complete channel assessment report; the dynamic resource scheduling uses optimization algorithms to allocate spectrum resources and adjusts the transmission power and modulation and coding schemes in advance based on the channel prediction trend. The feedback adjustment mechanism feeds back the scheduling results to the perception prediction module and the 3D clustering module through a low-latency control channel, dynamically adjusting the acquisition, clustering, and training parameters.
8. A distributed device group cooperative wireless channel sensing system integrating artificial intelligence and federated learning according to claim 7, characterized in that, The security protection module integrates homomorphic encryption, differential privacy, blockchain and wireless anti-interference technologies to build a full-process protection system; it adopts a consortium blockchain architecture to realize node identity authentication and behavior traceability, detects malicious interference through a deep learning interference identification model, and uses the SHA-256 hash verification algorithm to ensure data and parameter integrity. An adaptive differential privacy mechanism is introduced during parameter aggregation to achieve the dual goals of data usability without visibility and privacy protection.
9. A distributed device group cooperative wireless channel sensing method integrating artificial intelligence and federated learning, based on the system described in claim 8, characterized in that, The specific steps are as follows: Perception preprocessing stage: Each device collects core channel features and network status parameters through its built-in sensing unit, completes local preliminary perception using a lightweight AI algorithm, and achieves short-term channel prediction by combining a time-series prediction model; the generated binary data packets are processed according to a standardized process. Three-dimensional clustering stage: A decentralized strategy is adopted to extract three types of core features, and clustering weights are dynamically allocated to achieve accurate clustering of devices. The cluster structure is adaptively updated based on multiple triggering conditions, and the cluster head node coordinates intra-cluster communication resources. Fusion learning phase: Based on preprocessed data and device cluster partitioning, channel spatiotemporal dynamic features are extracted through fusion deep model, local model training is completed by combining federated learning, and regularization constraints and dynamic early stopping mechanism based on channel prediction error are adopted to ensure model compatibility and training efficiency. In the aggregation optimization phase: encryption schemes are dynamically selected based on link quality, and cluster head nodes are weighted and aggregated to generate a global model and complete cross-cluster fusion; Driven by global perception results, spectrum resource scheduling is implemented, and relevant parameters are optimized through feedback adjustment. Simultaneously, a full-process security protection mechanism is integrated to ensure collaborative security.