A Cognition-Oriented Intelligent Control Method for Wireless Networks

Abstract

This invention discloses a cognitively guided intelligent control method for wireless networks. First, a cognitively guided intelligent control system for wireless networks is constructed. Multi-dimensional scene cognitive features of the wireless network are collected, and these features are analyzed and encoded along four dimensions: topological connectivity, interference assessment, traffic pattern recognition, and channel quality assessment. Each dimension generates a feature vector. A sequence of feature vectors is constructed using classification symbols and input into a cognitively guided analysis model based on a multi-head attention mechanism. The model outputs the core scene category of the current network. Finally, based on the cognitively guided results, corresponding network parameter adjustment strategy chains are triggered sequentially from macro to micro levels, achieving cross-layer intelligent control. This invention overcomes the limitations of existing methods that rely on single-index judgment and isolated hierarchical optimization, achieving a unified multi-dimensional scene intelligent cognition and global cross-layer collaborative control, significantly improving network stability and communication performance under various harsh scenarios.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of wireless communication, unmanned aerial vehicle networks, and artificial intelligence, and specifically relates to a cognitively guided intelligent control method for wireless networks. Background Technology

[0002] In recent years, Flying Ad-Hoc Networks (FANETs) have been widely used in emergency communication and rescue, large-scale environmental monitoring, and other fields due to their flexible deployment, strong survivability, and rapid expansion capabilities. However, due to the high-speed three-dimensional mobility of UAV nodes, limited carrying capacity and computing resources, as well as the openness and time-varying nature of the wireless communication environment, FANETs face severe challenges such as frequent topology changes (e.g., topology failure caused by node crashes or power depletion), sudden aggregation of local traffic (e.g., multiple nodes converging data to the cluster head), and complex and ever-changing electromagnetic interference.

[0003] Existing wireless ad hoc network control methods typically follow the principle of independent design of layered protocol stacks, meaning they optimize based on fixed rules for a single network layer (such as routing protocol optimization at the routing layer, channel access mechanism optimization at the MAC layer, or power control at the physical layer) or a single specific scenario (such as simple anti-interference or simple topology repair). This localized and rigid control strategy has significant limitations when facing highly dynamic FANETs: First, network nodes lack global, multi-dimensional scenario awareness and often rely on single, one-sided indicators (such as signal strength or packet loss rate) for passive responses, failing to accurately analyze the core factors causing current network performance fluctuations. Second, single-layer parameter adjustments are insufficient to effectively address complex system-level network state deterioration. For example, when encountering strong interference, adjusting modulation and coding schemes only at the physical layer without changing network routing or node formation is often only a temporary solution; the lack of cross-layer collaborative mechanisms results in weak overall robustness and adaptability of the network in changing environments. Furthermore, the parameter optimization objectives between different layers may be mutually restrictive or even contradictory. For example, reducing the modulation order to improve the physical layer's anti-interference performance can lead to a decrease in transmission rate, causing a shortage of MAC layer time slot resources and a decrease in end-to-end throughput at the routing layer. Similarly, when the routing layer switches to a backup path to maintain connectivity after a topology failure, if the MAC layer continues to operate according to the old topology's time slot allocation scheme, relay nodes on the new path will lack sufficient channel access time slots in the short term, creating new transmission bottlenecks. Due to the lack of a unified cross-layer coordination mechanism, independent optimization at each layer can easily trigger a chain reaction of deterioration, like "pressing down one gourd and another floats up."

[0004] Therefore, there is an urgent need in this field for an intelligent method that can intelligently perceive and integrate multi-dimensional complex network features, accurately judge the current core scenario situation, and perform global cross-layer collaborative parameter control accordingly. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a cognitively-oriented intelligent control method for wireless networks, which solves the issues of insufficient cognitive ability and poor cross-layer collaboration in unmanned aerial vehicle (UAV) self-organizing networks when dealing with complex dynamic scenarios.

[0006] The technical solution adopted in this invention is: a cognitively-oriented intelligent control method for wireless networks, the specific steps of which are as follows:

[0007] S1. Construct a cognitively-oriented intelligent control system for wireless networks;

[0008] The control system includes: a scene feature acquisition module, a feature sequence optimization module, a cognitive guidance analysis module, and a cross-layer parameter adjustment module.

[0009] The cross-layer parameter adjustment module includes: a recruitment strategy engine, a formation control engine, a route reconstruction engine, a MAC layer scheduling engine, and a physical layer AMC engine.

[0010] The control system is applied to the UAV self-organizing network and is deployed at the central control node of the UAV self-organizing network for centralized execution. The central control node is a ground control station or a UAV cluster leader node. The cross-layer parameter adjustment module generates configuration instructions and sends them to the controlled UAV nodes in the network to execute the scene feature acquisition module, which is used to collect multi-dimensional scene cognitive features of the current wireless network.

[0011] The multi-dimensional scene cognition features include: network topology status, link information, service requirement fulfillment rate, interference type, interference parameters, and frequency band occupancy probability.

[0012] S2. Based on step S1, collect multi-dimensional scene cognitive features of the current wireless network through the scene feature acquisition module;

[0013] The central control node collects multi-dimensional scene cognitive features of the current network through periodic status reporting and network monitoring.

[0014] S3. Based on step S2, the collected multi-dimensional scene cognitive features are input into the feature sequence optimization module, and classified and coded according to the preset analysis dimensions. Each analysis dimension generates a feature vector, and the obtained feature vectors are concatenated with a classification symbol CLS token to construct a feature vector sequence.

[0015] The central control node first preprocesses and standardizes the multi-dimensional scene cognitive feature data obtained in step S2, and then classifies and extracts it according to four analysis dimensions: topology connectivity analysis, interference assessment, traffic pattern recognition, and channel quality assessment. It then performs analysis and encoding through linear mapping layers of each dimension, and outputs a fixed-dimensional feature vector Token for each analysis dimension.

[0016] Simultaneously, the control system initializes a learnable classification symbol, namely the CLS token. The four feature vectors obtained are then concatenated sequentially with the CLS token to construct a feature vector sequence of length 5. The expression is as follows:

[0017] ;

[0018] in, This represents the sequence of feature vectors input to the cognitive guidance analysis module; This represents the input vector corresponding to the CLS token; to These represent the feature vectors corresponding to the four analysis dimensions, and each vector has the same feature dimension.

[0019] S4. Input the feature vector sequence obtained in step S3 into the preset cognitive guidance analysis module for processing. Extract the association through the multi-head self-attention mechanism, output the cognitive guidance result, and indicate the core scene category of the current wireless network.

[0020] The core scenario categories include: topology failure scenarios, convergence service scenarios, and interference scenarios. The cognitive-guided analysis module includes at least one layer of network architecture based on a multi-head attention mechanism.

[0021] The feature vector sequence obtained in step S3 The cognitive guidance analysis module first obtains the updated sequence representation that integrates contextual information through a multi-head attention mechanism, and then classifies the updated sequence representation corresponding to the CLS token: the corresponding updated sequence representation vector is input into a linear classification layer to calculate the probability distribution of each candidate core scene category; the one with the largest value in the probability distribution is selected as the cognitive guidance result.

[0022] S5. Based on step S4, the cross-layer parameter adjustment module, according to the cognitive guidance result, sequentially triggers the network parameter adjustment strategy chain corresponding to the core scenario category in a hierarchical order from macro to micro, and performs cross-layer intelligent control of the network.

[0023] The hierarchical order includes the sequential evaluation and adjustment of network node recruitment, node grouping, network routing, Media Access Control (MAC) protocol, and physical layer transmission parameters.

[0024] Furthermore, in step S1, the control system executes a triggering mechanism that combines periodic triggering with event triggering.

[0025] In the absence of any emergencies, the control system executes a process once every fixed period. When a drastic change is detected in the key characteristics of the underlying sensors, an event is triggered, the current period is immediately interrupted, and an emergency cognitive analysis and parameter adjustment are initiated.

[0026] The event triggering conditions include: 1) detecting a mutation in any feature of the multidimensional scene cognitive features that exceeds a preset security threshold, and 2) the network packet loss rate exceeding a security threshold.

[0027] Furthermore, step S4 is specifically as follows:

[0028] The cognitive-guided analysis module processes the feature vector sequence obtained in step S3 using a model based on the Transformer Encoder architecture. After the feature vector sequence is input into the model, it is processed through a multi-head self-attention mechanism, and scaling dot product attention is calculated between each feature vector. The expression is as follows:

[0029] ;

[0030] in, , , These represent the query matrix, key matrix, and value matrix obtained by linear transformation of the feature sequence, respectively. This represents the dimension of the key vector, used to scale the dot product result. This indicates the transpose operation.

[0031] Then, the multi-head self-attention mechanism computes attention in parallel across multiple representation subspaces, and concatenates the outputs of each attention head to map them into an updated sequence representation. The expression is as follows:

[0032] ;

[0033] in, , Indicates the number of heads of attention. ; , , They represent the first A linear mapping matrix of queries, keys, and values ​​corresponding to each attention head; This represents the output mapping matrix after multiple heads are concatenated. This indicates a splicing operation.

[0034] The output of the multi-head self-attention layer is processed by residual connections, normalization, and a feedforward neural network to obtain an update sequence representation of the same length as the input sequence. The expression is as follows:

[0035] ;

[0036] in, This represents the update sequence output by the cognitive-guided analysis module; This indicates the output representation vector corresponding to the CLS token, which has already incorporated global context information from four analysis dimensions; to These represent the update representation vectors corresponding to the four analysis dimensions.

[0037] Then, Input a linear classification layer and calculate the probability distribution of each candidate core scene category using the Softmax function, as shown in the following expression:

[0038] ;

[0039] in, Indicates the first One candidate core scenario category; This represents a conditional probability function. Indicates the input sequence Given the current network type, it belongs to the category... The probability of; and These represent the weight matrix and bias parameters of the classification layer, respectively; subscripts This indicates taking the values ​​from the Softmax output vector that correspond to the class. The corresponding components. Finally, the category with the highest probability is selected as the cognitive guidance result of the current network, as shown in the following expression:

[0040] ;

[0041] in, This represents the final determined core scenario category, which is the current cognitive guidance result of the network.

[0042] Furthermore, in step S5, the central control node triggers different parameter adjustment strategy chains for different cognitive guidance results, as follows:

[0043] 1) When the cognitive guidance result is a topology failure scenario, the network parameter adjustment strategy chain is executed sequentially:

[0044] The macro layer first performs a network node recruitment operation, introducing new nodes into the network. That is, the central control node calls a preset recruitment algorithm to find suitable standby drones or redundant nodes in the surrounding environment to join the network and fill the gaps. After the recruitment is completed, a network topology reconstruction operation is performed to update the logical topology connection relationship of all network nodes. Finally, a path reconstruction operation is performed at the network layer, recalculating and issuing routing paths based on the updated logical topology, that is, replanning the optimal path for business flows based on the new topology.

[0045] 2) When the cognitive guidance result is a convergence business scenario, the network parameter adjustment strategy chain is executed sequentially:

[0046] The macro layer performs topology optimization, multipath routing, and load balancing operations at the formation and routing layers; then it performs time slot allocation optimization and congestion control operations at the MAC layer; and finally, it performs adaptive modulation and coding operations at the physical layer.

[0047] 3) When the cognitive guidance result is an interference scenario, the network parameter adjustment strategy chain is executed sequentially:

[0048] At the formation layer, a UAV formation shrinking operation is performed to control the nodes in the network to reduce the physical distance between them. Then, at the routing layer, a route switching operation to avoid interference areas is performed, selecting safe nodes far away from interference areas as relays. Next, at the MAC layer, a dynamic frequency switching operation is performed to switch the communication operating frequency band and control the interference-affected nodes to jump to an unoccupied safe frequency band. Finally, at the physical layer, an adaptive modulation and coding operation is performed to reduce the modulation order and increase the redundancy of error correction coding.

[0049] The beneficial effects of this invention are as follows: The method of this invention first constructs a cognitively guided intelligent control system for wireless networks. It collects multi-dimensional scene cognitive features of the wireless network and analyzes and encodes these features according to four analytical dimensions: topological connectivity, interference assessment, traffic pattern recognition, and channel quality assessment. Each dimension generates a feature vector. This feature vector sequence, combined with classification symbols, is input into a cognitively guided analysis model based on a multi-head attention mechanism. The model outputs the core scene category of the current network. Finally, based on the cognitively guided results, the corresponding network parameter adjustment strategy chain is triggered sequentially from macro to micro levels, achieving cross-layer intelligent control. This invention overcomes the limitations of existing methods that rely on single-index judgment and isolated hierarchical optimization. It achieves the unification of multi-dimensional scene intelligent cognition and global cross-layer collaborative control, significantly improving network stability and communication performance under various harsh scenarios. Attached Figure Description

[0050] Figure 1 This is a flowchart of a cognitively guided intelligent control method for wireless networks according to the present invention.

[0051] Figure 2 This is a block diagram of the cognitively guided wireless network intelligent control system in an embodiment of the present invention.

[0052] Figure 3 This is a cognitive-guided analysis architecture diagram of the multi-head attention mechanism in an embodiment of the present invention.

[0053] Figure 4 This is a flowchart of the network parameter adjustment based on cognition guidance in an embodiment of the present invention. Detailed Implementation

[0054] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0055] like Figure 1 The flowchart of a cognitively-oriented intelligent control method for wireless networks of the present invention is shown below. The specific steps are as follows:

[0056] S1. Construct a cognitively-oriented intelligent control system for wireless networks;

[0057] like Figure 2 As shown, the control system includes: a scene feature acquisition module, a feature sequence optimization module, a cognitive guidance analysis module, and a cross-layer parameter adjustment module.

[0058] The cross-layer parameter adjustment module is an action execution engine, including: a recruitment strategy engine, a formation control engine, a route reconstruction engine, a MAC layer scheduling engine, and a physical layer AMC engine.

[0059] The division of the control system and its functional modules is for descriptive convenience only. In actual applications, they can be combined into a single physical module or split into more sub-modules. These functional modules can be implemented by chips with computing capabilities (such as CPU, GPU, FPGA, or dedicated NPU) in conjunction with computer program logic stored in memory.

[0060] In this embodiment, the control system is applied to a self-organizing unmanned aerial vehicle (FANET) network constructed by multiple unmanned aerial vehicles (UAVs). Due to the highly dynamic nature of FANET, the control system is deployed at the central control node of the UAV FANET for centralized execution, in order to grasp global information and coordinate the issuance of decision commands. The central control node is a ground control station or a UAV cluster leader node (a node with strong computing power in the network that has been elected as the cluster leader). It generates configuration commands through the cross-layer parameter adjustment module and issues them to the controlled UAV nodes in the network to execute the scene feature acquisition module, which is used to collect multi-dimensional scene cognitive features of the current wireless network.

[0061] The multi-dimensional scene recognition features include: network topology status, link information, service requirement fulfillment rate, interference type, interference parameters, and frequency band occupancy probability. If the control system operates at a ground control station, it receives feature reports from each UAV via a telemetry link; if it operates at the UAV cluster head node, it aggregates cluster member information via an air-to-air link.

[0062] S2. Based on step S1, collect multi-dimensional scene cognitive features of the current wireless network through the scene feature acquisition module;

[0063] The central control node collects multi-dimensional scene cognitive features of the current network through periodic status reporting and network monitoring.

[0064] This embodiment selects six key dimensions: network topology status (such as node neighbor table, connectivity), link information (such as RSSI, signal-to-noise ratio SNR, link delay), service requirement fulfillment (such as whether the latency and throughput of the current service flow meet QoS requirements), interference type (such as broadband congestion interference, narrowband targeting interference, frequency sweeping interference, etc.), interference parameters (such as the location and intensity of the interference source), and the occupancy probability of each frequency band. These features cover both the internal state of the network and external environmental threats, and are the foundation for accurate network diagnosis.

[0065] S3. Based on step S2, the collected multi-dimensional scene cognitive features are input into the feature sequence optimization module, and classified and coded according to the preset analysis dimensions. Each analysis dimension generates a feature vector, and the obtained feature vectors are concatenated with a classification symbol CLS token to construct a feature vector sequence.

[0066] Since the data types and physical dimensions of the six features described in step S2 are different, the central control node first preprocesses and standardizes the multi-dimensional scene cognitive feature data obtained in step S2, and then classifies and extracts them according to four analysis dimensions: topology connectivity analysis, interference assessment, traffic pattern recognition and channel quality assessment. The data is then analyzed and encoded through the linear projection layer of each dimension, and each analysis dimension outputs a feature vector (Token) of a fixed dimension.

[0067] Simultaneously, the control system initializes a learnable classification symbol, namely the CLS token. The four obtained feature vectors are then concatenated sequentially with the CLS token to construct a feature vector sequence of length 5. The expression is as follows:

[0068] ;

[0069] in, This represents the sequence of feature vectors input to the cognitive guidance analysis module; This represents the input vector corresponding to the CLS token; to These represent the feature vectors corresponding to the four analysis dimensions, and each vector has the same feature dimension. This approach borrows from mechanisms in the field of natural language processing, enabling various heterogeneous network features to undergo subsequent computational interactions within a unified vector space.

[0070] S4. Input the feature vector sequence obtained in step S3 into the preset cognitive guidance analysis module for processing. Extract the association through the multi-head self-attention mechanism, output the cognitive guidance result, and indicate the core scene category of the current wireless network.

[0071] The core scenario categories include: topology failure scenarios, convergence service scenarios, and interference scenarios. The cognitive-guided analysis module includes at least one layer of network architecture based on a multi-head attention mechanism. This architecture remains lightweight (e.g., deploying only two encoder layers) to accommodate the limited computing resources of the UAV control node while ensuring real-time inference.

[0072] like Figure 3 As shown, the feature vector sequence obtained in step S3 The cognitive guidance analysis module first obtains the updated sequence representation that integrates contextual information through a multi-head attention mechanism, and then classifies the updated sequence representation corresponding to the CLS token: the corresponding updated sequence representation vector is input into a linear classification layer to calculate the probability distribution of each candidate core scene category; the one with the largest value in the probability distribution is selected as the cognitive guidance result.

[0073] S5. Based on step S4, the cross-layer parameter adjustment module, according to the cognitive guidance result, sequentially triggers the network parameter adjustment strategy chain corresponding to the core scenario category in a hierarchical order from macro to micro, and performs cross-layer intelligent control of the network.

[0074] Because communication networks contain multiple layers of protocols from top to bottom, and the parameters are mutually constrained, the adjustment operation follows the order of "execution in sequence", that is, starting from the macroscopic physical topology and formation structure, then the routing of the network layer, then the MAC protocol of the data link layer, and finally the microscopic physical layer parameters.

[0075] The hierarchical order includes the sequential evaluation and adjustment of network node recruitment, node grouping, network routing, media access control (MAC) protocol, and physical layer transmission parameters.

[0076] In this embodiment, step S4 is specifically as follows:

[0077] like Figure 3 As shown, the cognitive-guided analysis module uses a model based on the Transformer Encoder architecture to process the feature vector sequence obtained in step S3. After the feature vector sequence is input into the model, it is processed through a multi-head attention mechanism. Scaled dot-product attention is calculated between each feature vector (including the CLS token). The expression is as follows:

[0078] ;

[0079] in, , , These represent the query matrix, key matrix, and value matrix obtained by linear transformation of the feature sequence, respectively. This represents the dimension of the key vector, used to scale the dot product result. This indicates the transpose operation.

[0080] Then, the multi-head self-attention mechanism computes attention in parallel across multiple representation subspaces, and concatenates the outputs of each attention head to map them into an updated sequence representation. The expression is as follows:

[0081] ;

[0082] in, , Indicates the number of heads of attention. ; , , They represent the first A linear mapping matrix of queries, keys, and values ​​corresponding to each attention head; This represents the output mapping matrix after multiple heads are concatenated. This represents the splicing operation. Through this mechanism, the model can learn the correlation weights between topological connectivity, interference assessment, traffic pattern recognition, and channel quality assessment. For example, there may be a strong correlation between interference assessment results and channel quality assessment, and the model can assign higher importance to this interaction through attention weights.

[0083] The output of the multi-head self-attention layer is processed by residual connections, normalization, and a feedforward neural network (Linear mapping) to obtain an updated sequence representation of the same length as the input sequence. The expression is as follows:

[0084] ;

[0085] in, This represents the update sequence output by the cognitive-guided analysis module; This indicates the output representation vector corresponding to the CLS token, which has already incorporated global context information from four analysis dimensions; to These represent the update representation vectors corresponding to the four analysis dimensions.

[0086] Then, Input the linear classification layer and calculate the probability distribution of each candidate core scene category using the Softmax function, as shown in the following expression:

[0087] ;

[0088] in, Indicates the first One candidate core scenario category; This represents a conditional probability function. Indicates the input sequence Given the current network type, it belongs to the category... The probability of; and These represent the weight matrix and bias parameters of the classification layer, respectively; subscripts This indicates taking the values ​​from the Softmax output vector that correspond to the class. The corresponding components. Finally, the category with the highest probability is selected as the cognitive guidance result of the current network, as shown in the following expression:

[0089] ;

[0090] in, This represents the final determined core scenario category, which is the current cognitive guidance result of the network.

[0091] like Figure 4 As shown, in this embodiment, in step S5, the central control node triggers different parameter adjustment strategy chains for different cognitive guidance results, as follows:

[0092] 1) When the cognitive guidance result is a topology failure scenario (e.g., a critical relay node crashes, causing network fragmentation), the network parameter adjustment strategy chain is executed sequentially:

[0093] The macro layer first performs a network node recruitment operation, introducing new nodes into the network. That is, the central control node calls a preset recruitment algorithm (such as by broadcasting recruitment requests) to find suitable standby drones or redundant nodes in the surrounding environment to join the network and fill the gaps. After the recruitment is completed, a network topology reconstruction operation is performed to update the logical topology connection relationship of all network nodes. Finally, a path reconstruction operation is performed at the network layer to recalculate and distribute routing paths based on the updated logical topology, that is, to replan the optimal path for business flows based on the new topology.

[0094] 2) When the cognitive guidance result is a convergence business scenario (e.g., multiple reconnaissance drones simultaneously transmitting high-definition video data to the cluster head, causing local congestion), the network parameter adjustment strategy chain is executed sequentially:

[0095] The macro layer performs topology optimization (fine-tuning the positions of local nodes to improve spatial distribution) and multipath routing and load balancing operations (splitting a concentrated single service flow into multiple paths for transmission) at the formation and routing layers; then, it performs time slot allocation optimization and congestion control operations at the MAC layer to allocate more transmission time slot resources to high-volume nodes; finally, it performs adaptive modulation and coding (AMC) operations at the physical layer to increase the higher-order modulation ratio and improve the transmission rate when the channel allows.

[0096] 3) When the cognitive guidance result is an interference scenario (e.g., subjected to electromagnetic suppression by an external enemy), the network parameter adjustment strategy chain is executed sequentially:

[0097] At the formation layer, a UAV formation shrinking operation is performed, controlling the nodes within the network to reduce the physical distance between them. This involves commanding UAVs to move closer together to shorten the communication distance. Utilizing the free space propagation loss formula, this significantly improves the signal-to-interference ratio (SINR) at the receiver. Then, at the routing layer, a route switching operation is performed to avoid interference areas. This deliberately avoids nodes geographically close to the interference source and selects safe nodes far from the interference area as relays. Next, at the MAC layer, a dynamic frequency switching operation is performed to switch the communication operating frequency band and control the entire group of interfered nodes to jump to an unoccupied safe frequency band. Finally, at the physical layer, an adaptive modulation and coding operation is performed to reduce the modulation order (e.g., back to BPSK) and increase the redundancy of error correction coding, sacrificing data rate for anti-interference reliability.

[0098] This top-down, logically rigorous adjustment chain avoids the piecemeal problem of single-layer adjustments. For example, in interference scenarios, if only the coding rate is reduced at the physical layer, but the nodes are too far apart (the formation is not contracted) and the frequency band is completely suppressed (no frequency switching), communication will still be interrupted; only multi-layer coordination can fundamentally restore network connectivity.

[0099] To ensure the efficient operation of the system, in this embodiment, the control system execution triggering mechanism in step S1 adopts a combination of periodic triggering and event triggering.

[0100] In the absence of any emergencies, the control system executes a process once every fixed period (e.g., 1 second). When a drastic change is detected in the key characteristics of the underlying sensors (e.g., a sudden increase in interference power exceeding the safety threshold, or a series of large-scale packet loss events), an event is triggered, immediately interrupting the current period and triggering an emergency cognitive analysis and parameter adjustment.

[0101] In summary, the method of this invention breaks through the limitations of single-indicator judgment, integrating multi-dimensional features such as topology, links, services, and interference. It utilizes a lightweight cognitive-guided analysis model based on the Transformer mechanism to capture deep correlations between features (through a multi-head attention mechanism), enabling accurate identification of the root causes of network performance degradation and achieving a leap from "passive response" to "proactive cognition." This method abandons isolated optimization of a single protocol layer, customizing a progressive strategy chain covering recruitment, grouping, routing, MAC, and the physical layer (macro to micro) according to different scenario motivations. By strictly executing these strategies sequentially, it ensures the stability of the macro-level network architecture while achieving optimal matching of micro-level transmission parameters, effectively improving the network's survivability and data transmission efficiency in environments with topology damage, service overload, and strong electromagnetic interference. The method of this invention, through a mechanism combining periodicity and event triggering, ensures that the control system can respond to network mutations in real time. Simultaneously, the network architecture employing feature encoding + attention mechanisms has good scalability, allowing for easy integration of more dimensional feature indicators and expansion of more core scenario categories in the future, adapting to the needs of different complex environments.

[0102] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A cognitively-oriented intelligent control method for wireless networks, comprising the following steps: S1. Construct a cognitively-oriented intelligent control system for wireless networks; The control system includes: Scene feature acquisition module, feature sequence optimization module, cognitive guidance analysis module, cross-layer parameter adjustment module; The cross-layer parameter adjustment module includes: a recruitment strategy engine, a formation control engine, a route reconstruction engine, a MAC layer scheduling engine, and a physical layer AMC engine. The control system is applied to the UAV self-organizing network and is deployed at the central control node of the UAV self-organizing network for centralized execution. The central control node is a ground control station or a UAV cluster leader node. The cross-layer parameter adjustment module generates configuration instructions and sends them to the controlled UAV nodes in the network to execute the scene feature acquisition module, which is used to collect multi-dimensional scene cognitive features of the current wireless network. The multi-dimensional scene cognition features include: network topology status, link information, service requirement fulfillment, interference type, interference parameters, and frequency band occupancy probability. S2. Based on step S1, collect multi-dimensional scene cognitive features of the current wireless network through the scene feature acquisition module; Among them, the central control node collects multi-dimensional scene cognitive features of the current network through periodic status reporting and network monitoring; S3. Based on step S2, the collected multi-dimensional scene cognitive features are input into the feature sequence optimization module, and classified and coded according to the preset analysis dimensions. Each analysis dimension generates a feature vector, and the obtained feature vectors are concatenated with a classification symbol CLS token to construct a feature vector sequence. The central control node first preprocesses and standardizes the multi-dimensional scene cognitive feature data obtained in step S2, and then classifies and extracts it according to four analysis dimensions: topology connectivity analysis, interference assessment, traffic pattern recognition and channel quality assessment. It then performs analysis and encoding through the linear mapping layer of each dimension, and outputs a fixed-dimensional feature vector Token for each analysis dimension. Simultaneously, the control system initializes a learnable classification symbol, namely the CLS token; the obtained four feature vectors are then concatenated with the CLS token in sequence to construct a feature vector sequence of length 5. The expression is as follows: ； in, This represents the sequence of feature vectors input to the cognitive guidance analysis module; This represents the input vector corresponding to the CLS token; to These represent the feature vectors corresponding to the four analysis dimensions, and each vector has the same feature dimension; S4. Input the feature vector sequence obtained in step S3 into the preset cognitive guidance analysis module for processing. Extract the association through the multi-head self-attention mechanism, output the cognitive guidance result, and indicate the core scene category of the current wireless network. The core scenario categories include: topology failure scenarios, convergence service scenarios, and interference scenarios; the cognitive-guided analysis module includes at least one layer of network architecture based on a multi-head attention mechanism. The feature vector sequence obtained in step S3 The cognitive guidance analysis module first obtains the updated sequence representation fused with contextual information through a multi-head attention mechanism, and then classifies the updated sequence representation corresponding to the CLS token: the corresponding updated sequence representation vector is input into a linear classification layer to calculate the probability distribution of each candidate core scene category; the one with the largest value in the probability distribution is selected as the cognitive guidance result; S5. Based on step S4, the cross-layer parameter adjustment module, according to the cognitive guidance result, sequentially triggers the network parameter adjustment strategy chain corresponding to the core scenario category in a hierarchical order from macro to micro, and performs cross-layer intelligent control of the network. The hierarchical order includes the sequential evaluation and adjustment of network node recruitment, node grouping, network routing, Media Access Control (MAC) protocol, and physical layer transmission parameters.

2. The cognitively-guided intelligent control method for wireless networks according to claim 1, characterized in that, In step S1, the control system executes a triggering mechanism that combines periodic triggering with event triggering, as detailed below: In the absence of any emergencies, the control system executes the process once every fixed period. When a drastic change is detected in the key features of the underlying sensors, an event is triggered, the current period is immediately interrupted, and an emergency cognitive analysis and parameter adjustment are initiated. The event triggering conditions include: 1) detecting a mutation in any feature of the multidimensional scene cognitive features that exceeds a preset security threshold, and 2) the network packet loss rate exceeding a security threshold.

3. The cognitively-guided intelligent control method for wireless networks according to claim 1, characterized in that, Step S4 is as follows: The cognitive-guided analysis module processes the feature vector sequence obtained in step S3 using a model based on the Transformer Encoder architecture. After the feature vector sequence is input into the model, it is processed through a multi-head self-attention mechanism, and scaled dot product attention is calculated between each feature vector. The expression is as follows: ； in, , , These represent the query matrix, key matrix, and value matrix obtained by linear transformation of the feature sequence, respectively. This represents the dimension of the key vector, used to scale the dot product result. Indicates the transpose operation; Then, the multi-head self-attention mechanism computes attention in parallel across multiple representation subspaces, and concatenates the outputs of each attention head to map them into an updated sequence representation. The expression is as follows: ； in, , Indicates the number of heads of attention. ; , , They represent the first A linear mapping matrix of queries, keys, and values ​​corresponding to each attention head; This represents the output mapping matrix after multi-head concatenation; Indicates a splicing operation; The output of the multi-head self-attention layer is processed by residual connections, normalization, and a feedforward neural network to obtain an update sequence representation of the same length as the input sequence. The expression is as follows: ； in, This represents the update sequence output by the cognitive-guided analysis module; This indicates the output representation vector corresponding to the CLS token, which has already incorporated global context information from four analysis dimensions; to These represent the update representation vectors corresponding to the four analysis dimensions; Then, Input a linear classification layer and calculate the probability distribution of each candidate core scene category using the Softmax function, as shown in the following expression: ； in, Indicates the first One candidate core scenario category; This represents a conditional probability function. Indicates the input sequence Given the current network type, it belongs to the category... The probability of; and These represent the weight matrix and bias parameters of the classification layer, respectively; subscripts This indicates taking the values ​​from the Softmax output vector that correspond to the class. The corresponding components; finally, the category with the highest probability is selected as the cognitive guidance result of the current network, as shown in the following expression: ； in, This represents the final determined core scenario category, which is the current cognitive guidance result of the network.

4. The cognitively-guided intelligent control method for wireless networks according to claim 1, characterized in that, In step S5, the central control node triggers different parameter adjustment strategy chains for different cognitive guidance results, as follows: 1) When the cognitive guidance result is a topology failure scenario, the network parameter adjustment strategy chain is executed sequentially: The macro layer first performs a network node recruitment operation, introducing new nodes into the network. That is, the central control node calls a preset recruitment algorithm to find suitable standby drones or redundant nodes in the surrounding environment to join the network and fill the gaps. After the recruitment is completed, a network topology reconstruction operation is performed to update the logical topology connection relationship of all network nodes. Finally, a path reconstruction operation is performed at the network layer to recalculate and distribute routing paths based on the updated logical topology. That is, the optimal path is replanned for the business flow based on the new topology. 2) When the cognitive guidance result is a convergence business scenario, the network parameter adjustment strategy chain is executed sequentially: The macro layer performs topology optimization, multipath routing, and load balancing operations at the formation and routing layers; then it performs time slot allocation optimization and congestion control operations at the MAC layer; and finally, it performs adaptive modulation and coding operations at the physical layer. 3) When the cognitive guidance result is an interference scenario, the network parameter adjustment strategy chain is executed sequentially: At the formation layer, a UAV formation shrinking operation is performed to control the nodes in the network to reduce the physical distance between them. Then, at the routing layer, a route switching operation to avoid interference areas is performed, selecting safe nodes far away from interference areas as relays. Next, at the MAC layer, a dynamic frequency switching operation is performed to switch the communication operating frequency band and control the interference-affected nodes to jump to an unoccupied safe frequency band. Finally, at the physical layer, an adaptive modulation and coding operation is performed to reduce the modulation order and increase the redundancy of error correction coding.

Images