Unmanned aerial vehicle-to-unmanned aerial vehicle semantic communication control method based on large language model

Abstract

This invention provides a semantic communication control method for UAVs based on a large language model. An encoder and decoder are added to the communication network framework to form a semantic communication system for semantic-level command transmission. During this transmission, only key semantic content within the command is transmitted, effectively reducing the amount of data transmitted and avoiding redundant transmission of invalid commands. Semantic comparison is also performed on the commands of the sender and receiver to assess their semantic consistency and prevent errors in command communication. Encoding, decoding, and error correction are performed only on key semantic content within the command, reducing computational load and resource consumption. Transmitting key semantic content within the command between UAVs, rather than transmitting the entire data, and relying on context and semantic reasoning to recover non-key semantic content, enables a higher fault tolerance rate and improves the reliability and stability of communication between UAVs.

Description

Technical Field

[0001] This invention relates to the field of information communication, and more particularly to a semantic communication control method for unmanned aerial vehicles based on a large language model. Background Technology

[0002] Currently, most UAV communication network frameworks are based on ns-3 networks. This requires establishing communication connections between the UAV swarm and ns-3 network nodes, configuring routing and message forwarding rules within the ns-3 network, and designing and constructing the ns-3 network topology and routing table to ensure accurate communication between UAVs. However, these UAV communication methods have the following shortcomings: First, information is transmitted as a bitstream, resulting in low transmission efficiency. High precision is required for the transmitted content, and the receiver needs to parse and reconstruct the bitstream, potentially leading to significant redundant transmission. Second, the communication network framework cannot be evaluated, making it impossible to accurately determine the consistency between sent and received messages. Third, it requires substantial computing resources and places high demands on computational performance. Fourth, it has low fault tolerance, making communication failures due to data loss or errors more likely. Therefore, existing UAV communication networks cannot achieve efficient and accurate point-to-point communication between different UAVs, and they fail to reduce computational resource consumption and improve communication fault tolerance. Summary of the Invention

[0003] The purpose of this invention is to provide a semantic communication control method for unmanned aerial vehicles (UAVs) based on a large language model. This method adds an encoder and decoder to the communication network framework to form a semantic communication system for semantic-level command transmission. During this transmission, only key semantic content within the command is transmitted, effectively reducing the amount of data transmitted and avoiding redundant transmission of invalid commands, thus improving communication efficiency between UAVs. Furthermore, the invention performs semantic comparison between the commands of the sender and receiver to assess their semantic consistency and prevent errors in command transmission. Encoding, decoding, and error correction are performed only on key semantic content within the command, reducing computational load and avoiding the processing of large amounts of redundant data, thereby reducing computational resource consumption. Transmitting key semantic content within the command between UAVs, rather than transmitting the entire data, and relying on context and semantic reasoning to recover non-key semantic content, enables a higher fault tolerance rate in the communication process, improving the reliability and stability of communication between UAVs.

[0004] This invention is achieved through the following technical solution:

[0005] A semantic communication control method for UAVs based on a large language model includes:

[0006] The encoder of the client is instructed to convert the original instructions issued by the client into semantic communication messages and send the semantic communication messages to the main UAV;

[0007] The decoder of the main UAV is instructed to convert the semantic communication message into a restoration instruction, and to verify the restoration instruction and the original instruction to determine whether the restoration instruction and the original instruction meet the preset semantic consistency conditions.

[0008] When the preset semantic consistency condition is met, large language model matching is set for the encoder of the master UAV and the decoder of the slave UAV swarm; the encoder of the master UAV is instructed to convert the control command issued by the client into a semantic control message and send the semantic control message to the slave UAV swarm.

[0009] The decoder of the drone swarm is instructed to convert the semantic control message into a restoration control command, thereby adjusting the working state of the corresponding drone within the drone swarm according to the restoration control command.

[0010] Optionally, the encoder of the client is instructed to convert the original instructions issued by the client into semantic communication messages, including:

[0011] The encoder of the client is instructed to perform semantic encoding on the original instructions issued by the client to obtain the semantic information of the original instructions;

[0012] The semantic information is subjected to text recognition to extract key semantic content, which is then converted into semantic communication messages.

[0013] Optionally, before sending the semantic communication message to the main drone, the process includes:

[0014] Search for all drones in the vicinity of the client to obtain the identity information of all drones; based on the identity information, identify the master drone and establish a communication channel between the client and the master drone.

[0015] Optionally, sending the semantic communication message to the main UAV includes:

[0016] Based on the available bandwidth information of the communication channel between the client and the main drone, the semantic communication message is packaged and sent to the main drone.

[0017] Optionally, the decoder of the main UAV is instructed to convert the semantic communication message into a reconstruction instruction, including:

[0018] The decoder of the main UAV is instructed to perform contextual and semantic reasoning on all key semantic content within the semantic communication message, and convert the semantic communication message into a restoration instruction.

[0019] Optionally, the preset similarity threshold is dynamically adjusted in real time based on the signal-to-noise ratio of the communication channel, including:

[0020] The signal-to-noise ratio (SNR) of the communication link is monitored in real time. If the absolute value of the difference between the current SNR and the previous SNR exceeds a preset fluctuation threshold, the semantic information entropy of the current instruction to be verified is determined. The semantic information entropy of the current instruction to be verified is obtained using the following formula:

[0021]

[0022] Among them, H cmd The semantic information entropy of the instruction to be verified is represented by n; the total number of words appearing in the instruction is represented by p. i The probability of the i-th word in the instruction appearing in the UAV communication instruction library can be obtained through a large language model or statistical word frequency analysis.

[0023] The semantic information entropy of the current instruction to be verified is compared with the semantic information entropy of the previous instruction to obtain the absolute difference between the semantic information entropy of the current instruction to be verified and the semantic information entropy of the previous instruction.

[0024] The absolute difference between the semantic information entropy of the current instruction to be verified and the semantic information entropy of the previous instruction is compared with a preset entropy change threshold.

[0025] When the absolute difference exceeds the preset entropy change threshold, the semantic information entropy of the current instruction to be verified and the signal-to-noise ratio of the current communication channel are retrieved.

[0026] The information entropy of the current communication channel is obtained using the signal-to-noise ratio (SNR); wherein the information entropy of the current communication channel is obtained using the following formula:

[0027]

[0028] Among them, H ch The SNR represents the information entropy of the current communication channel;

[0029] The current similarity threshold is adjusted using the semantic information entropy of the current instruction to be verified and the signal-to-noise ratio of the current communication channel; wherein the adjusted similarity threshold is obtained by the following formula:

[0030]

[0031] Among them, P mλ represents the adjusted similarity threshold; λ and x represent the first adjustment parameter and the second adjustment coefficient, respectively, with values ​​ranging from [0.25, 1.0] to [0.1, 0.5], used to control the steepness and offset of the sigmoid curve; SNR0 represents the preset reference signal-to-noise ratio.

[0032] Optionally, the restoration instruction and the original instruction are verified to determine whether they satisfy a preset semantic consistency condition, including:

[0033] Based on the BLEU metric, semantic content verification is performed on the restored instruction and the original instruction to obtain the semantic content similarity between the restored instruction and the original instruction; the semantic content similarity is compared with a preset similarity threshold. If the semantic content similarity is greater than or equal to the preset similarity threshold, it is determined that the restored instruction and the original instruction meet the preset semantic consistency condition; otherwise, it is determined that the restored instruction and the original instruction do not meet the preset semantic consistency condition.

[0034] Optionally, when a preset semantic consistency condition is met, large language model matching settings are performed on the encoder of the master UAV and the decoder of the slave UAV swarm, including:

[0035] When the preset semantic consistency condition is met, the large language model parameters of the encoder of the client and the decoder of the master drone are matched and set based on their respective large language model parameters.

[0036] Optionally, instructing the encoder of the master UAV to convert the control commands issued by the client into semantic control messages, and sending the semantic control messages to the swarm of slave UAVs, includes:

[0037] The encoder of the main UAV is instructed to perform semantic encoding on the control commands issued by the client to obtain the semantic information of the control commands;

[0038] The semantic information is subjected to text recognition to extract key semantic content, which is then converted into a semantic control message and sent to the corresponding slave drone in the drone swarm.

[0039] Optionally, instructing the decoder from the drone swarm to convert the semantic control message into a restore control command includes:

[0040] The decoder of the corresponding slave drone in the drone swarm is instructed to perform contextual and semantic reasoning on all key semantic content in the semantic control message, and convert the semantic control message into a restored control command;

[0041] According to the restore control command, adjust the working status of the corresponding slave drones in the drone swarm, including:

[0042] The restore control command is parsed and processed, and corresponding task content is added to the task queue of the slave drone, thereby adjusting the working status of the slave drone.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The semantic communication control method for UAVs based on a large language model provided in this application adds an encoder and decoder to the communication network framework to form a semantic communication system for semantic-level command transmission. During semantic-level command transmission, only key semantic content within the command is transmitted, effectively reducing the amount of data transmitted and avoiding redundant transmission of invalid command content, thus improving the communication efficiency between UAVs. Furthermore, semantic comparison is performed on the commands corresponding to the sender and receiver to assess their semantic consistency and avoid errors in command communication. Encoding, decoding, and error correction are performed only on key semantic content within the command, reducing computational load and avoiding the processing of large amounts of redundant data, thereby reducing computational resource consumption. Transmitting key semantic content within the command between UAVs, rather than transmitting the entire data, and relying on context and semantic reasoning to recover non-key semantic content, enables a higher fault tolerance rate in the communication process, improving the reliability and stability of communication between UAVs. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0046] Figure 1 This is a flowchart illustrating the semantic communication control method for unmanned aerial vehicles based on a large language model provided by the present invention. Detailed Implementation

[0047] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, it should be noted that, for ease of description, only the parts relevant to this application are shown in the accompanying drawings, not the entire structure. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0048] The terms “comprising” and “having”, and any variations thereof, used in this application are intended to cover non-exclusive inclusion. For example, a process, method, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0049] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0050] Please see Figure 1 As shown, an embodiment of this application provides a semantic communication control method for inter-UAVs based on a large language model. This semantic communication control method for inter-UAVs based on a large language model includes:

[0051] The encoder of the client is instructed to convert the original command issued by the client into a semantic communication message and send the semantic communication message to the main UAV;

[0052] The decoder of the main UAV is instructed to convert the semantic communication message into a restoration command, verify the restoration command and the original command, and determine whether the restoration command and the original command meet the preset semantic consistency conditions.

[0053] When the preset semantic consistency condition is met, the encoder of the master UAV and the decoder of the slave UAV swarm are configured with large language model matching; the encoder of the master UAV is instructed to convert the control command issued by the client into a semantic control message and send the semantic control message to the slave UAV swarm.

[0054] The decoder of the drone swarm is instructed to convert the semantic control message into a restoration control command, thereby adjusting the working state of the corresponding drone within the drone swarm according to the restoration control command.

[0055] The beneficial effects of the above embodiments are as follows: This UAV semantic communication control method based on a large language model adds an encoder and decoder to the communication network framework to form a semantic communication system for semantic-level instruction transmission. During semantic-level instruction transmission, only key semantic content within the instruction is transmitted, effectively reducing the amount of instruction data transmitted and avoiding redundant transmission of invalid instruction content, thus improving the communication efficiency between UAVs. Furthermore, semantic comparison is performed on the instructions corresponding to the sender and receiver to assess their semantic consistency and avoid errors in instruction communication. Encoding, decoding, and error correction are performed only on key semantic content within the instruction, reducing computational load, avoiding the processing of large amounts of redundant data, and reducing computational resource consumption. Transmitting key semantic content within the instruction between UAVs, rather than transmitting the entire data, and relying on context and semantic reasoning to recover non-key semantic content, enables a higher fault tolerance rate in the communication process, improving the reliability and stability of communication between UAVs.

[0056] In another embodiment, the encoder of the client is instructed to convert the raw instructions issued by the client into semantic communication messages, including:

[0057] The encoder of the client is instructed to perform semantic encoding on the original command issued by the client to obtain the semantic information of the original command;

[0058] The semantic information is subjected to text recognition to extract key semantic content, which is then converted into a semantic communication message.

[0059] The beneficial effects of the above embodiments, in actual UAV swarm control scenarios, involve first selecting one UAV as the master UAV, with all other UAVs forming a slave UAV swarm. Both the master UAV and all slave UAVs within the swarm can operate independently and execute corresponding tasks. The master UAV acts as a communication intermediary between the client and the slave UAV swarm. When the client needs to send a command to a slave UAV, it first sends the command to the master UAV, which then forwards the command to the corresponding slave UAV. This avoids the client directly sending commands to the slave UAV, increasing the client's command sending workload and causing command reception delays for the slave UAVs. Furthermore, to effectively reduce information transmission redundancy between the client and the master UAV, and between the master UAV and slave UAVs, an encoder is added to the client, an encoder and decoder are added to the master UAV, and a decoder is added to the slave UAVs, based on the existing network communication architecture of the client, master UAV, and slave UAV swarm. The encoder and decoder enable semantic-level communication transmission between the client and the master UAV, and between the master UAV and slave UAVs. In addition, the encoder on the client corresponds to the decoder on the master drone, and the encoder on the master drone corresponds to the decoder on the slave drone, realizing instruction semantic communication between the encoder and decoder. Furthermore, the encoder on the client and the encoder on the master drone, as well as the decoder on the master drone and the decoder on the slave drone, all run the same large language model. After debugging the large language model running parameters between the encoder on the client and the decoder on the master drone, the large language model running parameters between the encoder on the master drone and the decoder on the slave drone are adjusted accordingly. That is, the encoder on the client and the decoder on the master drone are used as adjustment references for the encoder on the master drone and the decoder on the slave drone, respectively, thereby ensuring that the encoder on the master drone and the decoder on the slave drone can achieve accurate and efficient semantic content communication and transmission.

[0060] To ensure the accuracy of semantic content communication between the master and slave UAVs, it is necessary to first verify the semantic communication of instructions between the encoder on the client and the decoder on the master UAV. Specifically, during the verification process of the client's encoder and the master UAV's decoder, when the client needs to send raw instructions to the master UAV, the client's encoder is first instructed to semantically encode the raw instructions sent by the client to obtain the semantic information of the raw instructions; then, text recognition is performed on the semantic information to determine the key and non-key semantic content in the global text content corresponding to the semantic information, and only all the key semantic content is integrated and converted into semantic communication messages. This can effectively compress the data volume of semantic communication messages while fully representing the true content of the raw instructions. The client's encoder can perform semantic encoding and text recognition operations through, but is not limited to, the DeepSC model, and can be continuously optimized and trained during the communication and interaction between the encoder on the client and the encoder on the master UAV to achieve the corresponding functions. The semantic encoding and text recognition of raw instructions by the DeepSC model are conventional techniques in this field, and the above semantic encoding and text recognition processes will not be described in detail here.

[0061] In another embodiment, before sending the semantic communication message to the main drone, the following steps are included:

[0062] Search for all drones in the vicinity of the client to obtain the identity information of all drones; based on the identity information, identify the master drone and establish a communication channel between the client and the master drone.

[0063] The beneficial effect of the above embodiments is that the master drone in a drone swarm is not fixed. Typically, the drone with the most available computing and communication resources is selected as the master drone. However, the available computing and communication resources of each drone in the swarm change depending on its task execution status. For example, one drone may have the most available computing and communication resources at one time, while at another time these resources may decrease. To ensure a consistently good and stable communication between the client and the drone swarm, the client needs to periodically search for all drones in the vicinity to obtain their individual identity information. Based on this identity information, the client determines the amount of available computing and communication resources for each drone, selects a suitable drone as the master drone, and establishes a communication channel between the client and the master drone. Alternatively, a drone in the drone swarm can be designated as the master drone in advance. The designated master drone has a large amount of computing and communication resources, and it is only responsible for relaying communication between the client and the drone swarm, without performing other tasks. This can maintain high computing and communication performance at all times. The client only needs to compare the identity information of all the drones it has searched with the identity information of the designated drone to identify the master drone in the vicinity and establish a communication channel between the client and the master drone.

[0064] In another embodiment, sending the semantic communication message to the master drone includes:

[0065] Based on the available bandwidth information of the communication channel between the client and the main drone, the semantic communication message is packaged and sent to the main drone.

[0066] The beneficial effect of the above embodiments is that when the encoder of the client generates a semantic communication message, in order to ensure that the semantic communication message is completely sent to the main drone, the available bandwidth of the communication channel between the client and the main drone is first determined based on the available bandwidth information between the client and the main drone. This determines whether the semantic communication message can be completely sent to the main drone in one transmission. If so, the semantic communication message is packaged and compressed as a whole and sent directly to the main drone. If not, the semantic communication message is divided into several message packets and then sent to the main drone in sequence, thereby avoiding omission of semantic communication messages during the transmission process.

[0067] In another embodiment, the decoder of the master drone is instructed to convert the semantic communication message into a reconstruction instruction, including:

[0068] The decoder of the main UAV is instructed to perform contextual and semantic reasoning on all key semantic content within the semantic communication message, and convert the semantic communication message into a restoration instruction.

[0069] The beneficial effect of the above embodiments is that when the main UAV receives a semantic communication message from the client, since the semantic communication message only contains key semantic content within the semantic information corresponding to the original instruction, it cannot reflect the complete semantic content of the original instruction. If the corresponding task is executed correctly directly based on the above semantic communication message, it is necessary to complete the semantic content of the semantic communication message. At this time, the decoder of the main UAV performs contextual and semantic reasoning on all key semantic content in the semantic communication message through, but not limited to, the DeepSC model, and completes the corresponding non-key semantic content based on the key semantic content, so that the converted restored instruction completely reflects the semantic content of the original instruction in textual logic. Among them, the contextual and semantic reasoning of key semantic content by the DeepSC model is a conventional technical means in this field, and the above contextual and semantic reasoning process will not be described in detail here.

[0070] In another embodiment, the restoration instruction and the original instruction are verified to determine whether they satisfy a preset semantic consistency condition, including:

[0071] Based on the BLEU metric, semantic content verification is performed on the restored instruction and the original instruction to obtain the semantic content similarity between the restored instruction and the original instruction. The semantic content similarity is compared with a preset similarity threshold. If the semantic content similarity is greater than or equal to the preset similarity threshold, it is determined that the restored instruction and the original instruction meet the preset semantic consistency condition; otherwise, it is determined that the restored instruction and the original instruction do not meet the preset semantic consistency condition.

[0072] In end-to-end communication systems, the sender and receiver typically use bit error rate (BER) as the training objective, neglecting other communication-related objectives. For text transmission, BER does not accurately reflect communication performance. Besides manually judging and establishing sentence similarity, bilingual evaluation BLUE scores are used to measure machine translation results. However, BLUE scores can only compare differences between words in two sentences, not their semantic information. Therefore, this invention proposes a new metric—sentence similarity—to describe the degree of similarity between two sentences from a semantic information perspective.

[0073] For length of The transmitted sentence s and its length are Decoding sentences The BLUE score is represented as follows:

[0074] (1)

[0075] In the above formula, The weight of an n-gram is indicated. The score represents the score of an n-gram, where N represents the total number of n-grams.

[0076] Among them, the score of n-grams It is expressed as follows:

[0077] (2)

[0078] In the above formula, C k () represents the frequency counting function for the k-th element in an n-gram.

[0079] The BLUE score ranges from 0 to 1, representing the similarity between the decoded and transmitted sentences. A BLUE score of 1 indicates the highest similarity. In practice, achieving a BLUE score of 1 through manual translation is difficult because a single error can lead to semantic differences between sentences. For communication command transmission, using synonyms does not change the semantics but will result in different BLUE scores for the two commands. To meet the requirements, this invention proposes a new metric at the sentence level—sentence similarity.

[0080] Specifically, a word can have different meanings in different contexts. Traditional methods like word2vec cannot recognize this polysemy. The problem lies in how to represent a word using numerical vectors, which differ across contexts. Based on semantic similarity, this paper proposes calculating the sentence similarity (match) between the source and retrieved instructions. , ),

[0081] (3)

[0082] In the above formula, This refers to the pre-trained BERT model, which uses parameters (e.g., billions of parameters) to extract semantic information. `match(` , The match() represents the similarity between the decoded sentence and the transmitted sentence, with a value ranging from 0 to 1. , A match value of 1 indicates the highest similarity between the decoded sentence and the transmitted sentence. , A value of 0 indicates that the decoded sentence is completely dissimilar to the transmitted sentence.

[0083] Compared to BLUE scores, the pre-trained BERT model has been trained on billions of sentences, thus learning semantic information from them and effectively generating different semantic vectors in different contexts. Using the pre-trained BERT model, the semantic information behind the transmitted sentence can be represented as c, while the estimated semantic information of the transmitted sentence can be represented as... Therefore, it can be accessed through Calculate sentence similarity.

[0084] The beneficial effects of the above embodiments are that, when the DeepSC model is not effectively trained, the restored instructions obtained after contextualizing and semantically reasoning the key semantic content may have a large semantic content deviation from the original instructions. This makes it impossible for the restored instructions to correctly and completely reflect the true semantic situation of the original instructions. If the task is executed directly based on the restored instructions, task execution errors will occur. In order to ensure the accuracy and reliability of the decoder's contextualization and semantic reasoning of the key semantic content, the semantic content of the restored instructions and the original instructions is verified based on the Bilingual Evaluation Understudy (BLEU) metric. The semantic content similarity between the restored instructions and the original instructions is obtained, and a threshold comparison is performed on the semantic content similarity to determine whether the restored instructions and the original instructions meet the preset semantic consistency conditions. The semantic content verification of the restored instructions and the original instructions based on the BLEU metric can be implemented by, but is not limited to, the large language model all-MinLM-L12-v2, which is a conventional technique in this field and will not be described in detail here. When the restored command and the original command do not meet the preset semantic consistency condition, it indicates that the DeepSC model has not been effectively trained and cannot guarantee that the restored command and the original command are completely consistent in semantic content. At this time, the DeepSC model should continue to be trained until the preset semantic consistency condition is met. When the restored command and the original command meet the preset semantic consistency condition, it indicates that the DeepSC model has been effectively trained and can guarantee that the restored command and the original command are completely consistent in semantic content. That is, the encoder on the client and the decoder on the master drone can extract and supplement the effective semantic content of the command at the semantic level. At this time, the same settings can be applied to the encoder on the master drone and the decoder on the slave drone based on the encoder on the client and the decoder on the master drone.

[0085] In another embodiment, the preset similarity threshold is dynamically adjusted in real time based on the signal-to-noise ratio of the communication channel, including:

[0086] The signal-to-noise ratio (SNR) of the communication link is monitored in real time. If the absolute value of the difference between the current SNR and the previous SNR exceeds a preset fluctuation threshold, the semantic information entropy of the current instruction to be verified is determined. The semantic information entropy of the current instruction to be verified is obtained using the following formula:

[0087]

[0088] Among them, H cmd The semantic information entropy of the instruction to be verified is represented by n; the total number of words appearing in the instruction is represented by p. iThe probability of the i-th word in the instruction appearing in the UAV communication instruction library can be obtained through a large language model or statistical word frequency analysis.

[0089] The semantic information entropy of the current instruction to be verified is compared with the semantic information entropy of the previous instruction to obtain the absolute difference between the semantic information entropy of the current instruction to be verified and the semantic information entropy of the previous instruction.

[0090] The absolute difference between the semantic information entropy of the current instruction to be verified and the semantic information entropy of the previous instruction is compared with a preset entropy change threshold.

[0091] When the absolute difference exceeds the preset entropy change threshold, the semantic information entropy of the current instruction to be verified and the signal-to-noise ratio of the current communication channel are retrieved.

[0092] The information entropy of the current communication channel is obtained using the signal-to-noise ratio (SNR); wherein the information entropy of the current communication channel is obtained using the following formula:

[0093]

[0094] Among them, H ch The SNR represents the information entropy of the current communication channel;

[0095] The current similarity threshold is adjusted using the semantic information entropy of the current instruction to be verified and the signal-to-noise ratio of the current communication channel; wherein the adjusted similarity threshold is obtained by the following formula:

[0096]

[0097] Among them, P m λ represents the adjusted similarity threshold; λ and x represent the first adjustment parameter and the second adjustment coefficient, respectively, with values ​​ranging from [0.25, 1.0] to [0.1, 0.5], used to control the steepness and offset of the sigmoid curve; SNR0 represents the preset reference signal-to-noise ratio.

[0098] The beneficial effects of the above embodiments are that, by constructing a multi-dimensional adaptive correlation model of semantic information entropy and channel signal-to-noise ratio (SNR), dynamic and precise control of the communication command verification threshold is achieved, significantly improving the robustness and verification efficiency of command recognition in complex communication environments. The scheme innovatively quantifies the uncertainty of command semantics into an information entropy index, accurately characterizes the complexity of command semantics using a probabilistic statistical model, and derives channel information entropy by combining channel SNR, constructing a similarity threshold calculation mechanism that reflects the coupling relationship between source information and channel state. Regarding triggering conditions, by real-time monitoring of SNR fluctuations and entropy differences, threshold adjustment is only initiated when the system state changes significantly, effectively avoiding the consumption of ineffective computing resources. The dynamic adjustment of the threshold achieves a positive correlation between command semantic complexity and the threshold, i.e., the more complex the semantics, the more stringent the verification. Simultaneously, negative feedback compensation is achieved by combining channel quality; the better the channel quality, the more lenient the threshold, and vice versa. Through the nonlinear mapping characteristics of the sigmoid function, the threshold adjustment process possesses a smooth and controllable dynamic response capability, capable of adapting to fluctuations in verification requirements under different channel conditions. This solution effectively addresses the problem of poor adaptability of traditional fixed thresholds in dynamic communication environments. It improves the anti-interference capability and recognition accuracy of command verification in low signal-to-noise ratio and strong interference environments, and enhances the system response speed by lowering the verification threshold in scenarios with excellent channel quality. Overall, it optimizes the balance between verification accuracy and processing efficiency, ensuring the reliability and real-time performance of command interaction in dynamic communication scenarios such as UAVs.

[0099] In another embodiment, when a preset semantic consistency condition is met, a large language model matching setting is performed on the encoder of the master UAV and the decoder of the UAV swarm, including:

[0100] When the preset semantic consistency condition is met, the large language model parameter matching settings are performed on the encoder of the main drone and the decoder of the drone swarm based on the large language model parameters of the encoder of the client and the decoder of the main drone.

[0101] As can be seen from the above analysis, when the preset semantic consistency condition is met, it indicates that the DeepSC model has been effectively trained and can guarantee that the restored instruction is completely consistent with the original instruction in terms of semantic content. That is, the encoder on the client and the decoder on the master drone can extract and supplement the effective semantic content of the instruction at the semantic level. At this time, based on the encoder on the client and the decoder on the master drone, the same settings can be applied to the encoder on the master drone and the decoder on the slave drone. That is, based on the large language model parameters of the encoder on the client and the decoder on the master drone, the large language model parameter matching settings can be applied to the encoder on the master drone and the decoder on the slave drone swarm. This ensures that the encoder on the master drone and the decoder on the slave drone swarm have the same instruction semantic extraction and context and reasoning performance as the encoder on the client and the decoder on the master drone, respectively. This achieves the transmission of key semantic content within the instruction between the master drone and the slave drone, rather than the overall data transmission.

[0102] In another embodiment, instructing the encoder of the master drone to convert the control commands issued by the client into semantic control messages and send the semantic control messages to the swarm of slave drones includes:

[0103] The encoder of the main UAV is instructed to perform semantic encoding on the control commands issued by the client to obtain the semantic information of the control commands;

[0104] The semantic information is subjected to text recognition to extract key semantic content, which is then converted into a semantic control message and sent to the corresponding slave drone in the drone swarm.

[0105] The beneficial effect of the above embodiment is that when the client needs to control the working status of the slave drone, the client first sends a control command to the master drone. The encoder of the master drone performs semantic encoding on the control command sent by the client to obtain the semantic information of the control command. Then, the semantic information is subjected to text recognition to extract the key semantic content within the semantic information, thereby generating a semantic control message containing only the key semantic content and sending it to the corresponding slave drone in the slave drone group, so as to realize the client's command control of the slave drone.

[0106] In another embodiment, the decoder from the drone swarm is instructed to convert the semantic control message into a restoring control command, including:

[0107] The decoder of the corresponding drone in the drone swarm is instructed to perform contextual and semantic reasoning on all key semantic content in the semantic control message and convert the semantic control message into a restored control command.

[0108] The beneficial effects of the above embodiments are that when the slave drone receives a semantic control message from the master drone, it uses the decoder of the slave drone to perform contextual and semantic reasoning on all key semantic content in the semantic control message, and completes the key semantic content into a restored control command. On the basis of reducing the amount of data transmission between the master drone and the slave drone, it ensures that the control command can be completely transmitted to the slave drone at the semantic level, thereby achieving precise control of the slave drone.

[0109] In another embodiment, adjusting the operating state of a corresponding slave drone within the drone swarm according to the restore control command includes:

[0110] The restore control command is parsed and processed, and the corresponding task content is added to the task queue of the slave drone, thereby adjusting the working status of the slave drone.

[0111] The UAV semantic communication framework comprises a semantic layer and a transport layer. The semantic layer processes semantic information, encoding and decoding it to extract semantic information. The transport layer ensures the correct exchange of semantic information over the transmission medium. The input to the DeepSC encoder is a communication command s = ,in This represents the first word in a communication command. The sending end consists of two parts: a semantic encoder and a channel encoder, used to extract the semantic information of the communication command and ensure its successful transmission over the physical channel. The encoded symbol stream x is shown below:

[0112] (4)

[0113] In the above formula, S β () denotes the semantic encoder network with parameter set β, C α () represents the channel encoder of parameter set α.

[0114] To simplify the analysis, let the correlation time be M. If symbol stream x is transmitted, the signal y received by the receiver is as follows:

[0115] (5)

[0116] In the above formula, h represents and The channel coefficients of the Rayleigh fading channels, where n represents the number of Rayleigh fading channels. This represents a complex standard normal distribution. The variance is expressed as It follows a normal distribution.

[0117] For end-to-end training of the encoder and decoder, the physical channel must allow backpropagation. The physical channel can be represented using neural networks; for example, simple neural networks can be used to model AWGN channels, multiplicative Gaussian noise channels, or erasing channels. For fading channels, more complex neural networks are required, primarily considering the simplicity of AWGN and Rayleigh fading channels while focusing on semantic encoding and decoding.

[0118] The receiver also includes a channel decoder and a semantic decoder, which recover the signal sent by the encoder. The decoded channel can be represented as follows:

[0119] (6)

[0120] In the above formula, Sentences indicating restoration, C δ -1 () indicates that there is a parameter set The channel encoder, S χ -1 () indicates that there is a parameter set A semantic decoder network.

[0121] The goal of a semantic communication system is to reduce the number of symbols to be transmitted while minimizing semantic errors.

[0122] In the UAV communication framework, joint semantic channel coding needs to be designed, and semantic transmission issues not considered in traditional communication systems also need to be taken into account. Traditional communication systems can achieve low bit error rates, but some bits exceed their error correction capabilities under noise interference, leading to the loss of partial semantic information of the entire command. To achieve successful semantic recovery, semantic coding and channel coding are jointly designed to preserve the transmitted sentence. and decoding sentences The meaning between them remains unchanged, and they are implemented in a new deep neural network framework, using cross-entropy (CE) as the loss function for measurement. and The difference between them can be expressed as follows:

[0123] (7)

[0124] In the above formula, Let represent the true probability that the l-th word appears in the estimated transmitted sentence s. Indicates the l-th word Appears in the decoded sentence The predicted probability in the source sentence. Cross-entropy (CE) is used to measure the difference between two probability distributions. By reducing the loss value of cross-entropy CE, the network can learn the word distribution in the source sentence. This demonstrates that the network can learn the grammar, phrases, and meanings of words in context. Jointly designing and training semantic channel codes allows the entire network to learn knowledge for a specific target. Channel coding can place greater emphasis on preserving semantic information relevant to the transmission target while ignoring secondary information.

[0125] To maximize the transmission rate of a semantic communication system, mutual information can provide additional information for training the decoder. The mutual information I(x;y) between the transmitted symbol x and the received symbol y is calculated as follows:

[0126] (8)

[0127] In the above formula, Let represent a pair of random variables whose values ​​are distributed as follows: Spatially, This represents the marginal probability of sending x. This represents the marginal probability of receiving y. Let x represent the joint probability of x and y. Mutual information is equivalent to the KL divergence between the marginal probabilities and the joint probabilities, as shown below.

[0128] (9)

[0129] Furthermore, according to Theorem 1 below: KL divergence allows the following dual representation:

[0130] (10)

[0131] In the above formula, E p Let E represent the expectation under probability distribution P. Q Let represent the expectation under probability distribution Q, T represent the function defined on the input space, R represent the joint distribution of random variables or the sample pair space, Ω represent the function space used to constrain the range of values ​​of function T, and sup represents the supremum, that is, to choose the function T in the function space Ω that maximizes the objective function value.

[0132] Combining Theorem 1 above, the KL divergence can also be expressed as follows:

[0133] (11)

[0134] In the above formula, E p(x,y) Let E represent the expectation under the probability distribution P(x,y). p(x)p(y) Let represent the expectation under the probability distribution P(x)p(y), and T represent the function defined on the input variables x and y.

[0135] The lower bound of mutual information I(x;y) can be obtained through the above equations (9) and (11).

[0136] To obtain an upper bound for mutual information I(x;y), an unsupervised method is used to train network T, which can be approximated by a neural network.

[0137] Meanwhile, the expectation in equation (11) above can be obtained through sampling calculation, and converges to the true value as the number of samples increases. The encoder is then optimized by maximizing the mutual information defined in equation (11) above, with the relevant loss function as follows:

[0138] (12)

[0139] In the above formula, This represents a neural network, where the input is... and The sample, x represents the function and The generation is where E represents the expectation operation and T represents the function used to estimate mutual information.

[0140] Therefore, the loss function can Let it be represented as follows.

[0141] (13)

[0142] In the above formula, the loss function is used to train the neural network to obtain... , And T. For example, when the encoder and When the mutual information is fixed, it can be estimated by training the network T; simultaneously, given the mutual information, it can be estimated by training... and To optimize the encoder.

[0143] The beneficial effects of the above embodiments are that after the decoder of the drone obtains the restored control command, the restored control command is parsed and processed to obtain the task information to be executed contained in the restored control command. In this way, the corresponding task content is added to the task queue of the drone, so that the drone can adjust its working state according to the control command issued by the client, ensuring the real-time performance and reliability of the client's control over the drone.

[0144] In summary, this semantic communication control method for UAVs based on a large language model adds encoders and decoders to the communication network framework to form a semantic communication system for semantic-level command transmission. During this transmission, only key semantic content within the command is transmitted, effectively reducing the amount of data transmitted and avoiding redundant transmission of invalid commands, thus improving communication efficiency between UAVs. Furthermore, it performs semantic comparison between the commands sent by the sender and received by the receiver, assessing their semantic consistency to prevent errors in command communication. Encoding, decoding, and error correction are performed only on key semantic content within the command, reducing computational load and avoiding the processing of large amounts of redundant data, thereby lowering computational resource consumption. Transmitting key semantic content within the command between UAVs, rather than transmitting the entire data, and relying on context and semantic reasoning to recover non-key semantic content, enables a higher fault tolerance rate in the communication process, improving the reliability and stability of communication between UAVs.

[0145] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. This disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims. Thus, if these modifications and variations of the invention fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A semantic communication control method for unmanned aerial vehicles (UAVs) based on a large language model, characterized in that, include: The encoder of the client is instructed to convert the original instructions issued by the client into semantic communication messages and send the semantic communication messages to the main UAV; The decoder of the main UAV is instructed to convert the semantic communication message into a restoration instruction, and to verify the restoration instruction and the original instruction to determine whether the restoration instruction and the original instruction meet the preset semantic consistency conditions. When the preset semantic consistency condition is met, large language model matching is set for the encoder of the master UAV and the decoder of the slave UAV swarm; the encoder of the master UAV is instructed to convert the control command issued by the client into a semantic control message and send the semantic control message to the slave UAV swarm. The decoder of the drone swarm is instructed to convert the semantic control message into a restoration control command, thereby adjusting the working state of the corresponding drone within the drone swarm according to the restoration control command.

2. The semantic communication control method for unmanned aerial vehicles based on a large language model as described in claim 1, characterized in that: The encoder of the client is instructed to convert the raw instructions issued by the client into semantic communication messages, including: The encoder of the client is instructed to perform semantic encoding on the original instructions issued by the client to obtain the semantic information of the original instructions; The semantic information is subjected to text recognition to extract key semantic content, which is then converted into semantic communication messages.

3. The semantic communication control method for unmanned aerial vehicles based on a large language model as described in claim 2, characterized in that: Before sending the semantic communication message to the main drone, the process includes: Search for all drones in the vicinity of the client to obtain the identity information of all drones; based on the identity information, identify the master drone and establish a communication channel between the client and the master drone.

4. The semantic communication control method for UAVs based on a large language model as described in claim 3, characterized in that: Sending the semantic communication message to the main UAV includes: Based on the available bandwidth information of the communication channel between the client and the main drone, the semantic communication message is packaged and sent to the main drone.

5. The semantic communication control method for unmanned aerial vehicles based on a large language model as described in claim 1, characterized in that: The decoder of the main UAV is instructed to convert the semantic communication message into a reconstruction instruction, including: The decoder of the main UAV is instructed to perform contextual and semantic reasoning on all key semantic content within the semantic communication message, and convert the semantic communication message into a restoration instruction.

6. The semantic communication control method for unmanned aerial vehicles based on a large language model as described in claim 5, characterized in that: Verify the restoration instruction and the original instruction to determine whether they meet a preset semantic consistency condition, including: Based on the BLEU metric, semantic content verification is performed on the restored instruction and the original instruction to obtain the semantic content similarity between the restored instruction and the original instruction; the semantic content similarity is compared with a preset similarity threshold. If the semantic content similarity is greater than or equal to the preset similarity threshold, it is determined that the restored instruction and the original instruction meet the preset semantic consistency condition; otherwise, it is determined that the restored instruction and the original instruction do not meet the preset semantic consistency condition.

7. The semantic communication control method for unmanned aerial vehicles based on a large language model as described in claim 6, characterized in that: Based on the signal-to-noise ratio of the communication channel, the preset similarity threshold is dynamically adjusted in real time, including: The signal-to-noise ratio (SNR) of the communication link is monitored in real time. If the absolute value of the difference between the current SNR and the previous SNR exceeds a preset fluctuation threshold, the semantic information entropy of the current instruction to be verified is determined. The semantic information entropy of the current instruction to be verified is obtained using the following formula: Among them, H cmd The semantic information entropy of the instruction to be verified is represented by n; the total number of words appearing in the instruction is represented by p. i This represents the probability that the i-th word in the instruction appears in the UAV communication instruction library; The semantic information entropy of the current instruction to be verified is compared with the semantic information entropy of the previous instruction to obtain the absolute difference between the semantic information entropy of the current instruction to be verified and the semantic information entropy of the previous instruction. The absolute difference between the semantic information entropy of the current instruction to be verified and the semantic information entropy of the previous instruction is compared with a preset entropy change threshold. When the absolute difference exceeds the preset entropy change threshold, the semantic information entropy of the current instruction to be verified and the signal-to-noise ratio of the current communication channel are retrieved. The information entropy of the current communication channel is obtained using the signal-to-noise ratio (SNR); wherein the information entropy of the current communication channel is obtained using the following formula: Among them, H ch The SNR represents the information entropy of the current communication channel; The current similarity threshold is adjusted using the semantic information entropy of the current instruction to be verified and the signal-to-noise ratio of the current communication channel; wherein the adjusted similarity threshold is obtained by the following formula: Among them, P m λ represents the adjusted similarity threshold; λ and x represent the first adjustment parameter and the second adjustment coefficient, respectively, with values ​​ranging from [0.25, 1.0] to [0.1, 0.5], used to control the steepness and offset of the sigmoid curve; SNR0 represents the preset reference signal-to-noise ratio.

8. The semantic communication control method for unmanned aerial vehicles based on a large language model as described in claim 1, characterized in that: When the preset semantic consistency condition is met, large language model matching settings are performed on the encoder of the master UAV and the decoder of the UAV swarm, including: When the preset semantic consistency condition is met, the large language model parameters of the encoder of the client and the decoder of the master drone are matched and set based on their respective large language model parameters.

9. The semantic communication control method for unmanned aerial vehicles based on a large language model as described in claim 8, characterized in that: The encoder of the master drone is instructed to convert the control commands issued by the client into semantic control messages, and to send the semantic control messages to the swarm of slave drones, including: The encoder of the main UAV is instructed to perform semantic encoding on the control commands issued by the client to obtain the semantic information of the control commands; The semantic information is subjected to text recognition to extract key semantic content, which is then converted into a semantic control message and sent to the corresponding slave drone in the drone swarm.

10. The semantic communication control method for unmanned aerial vehicles based on a large language model as described in claim 1, characterized in that: The decoder from the drone swarm is instructed to convert the semantic control message into a restoring control command, including: The decoder of the corresponding slave drone in the drone swarm is instructed to perform contextual and semantic reasoning on all key semantic content in the semantic control message, and convert the semantic control message into a restored control command; According to the restore control command, adjust the working status of the corresponding slave drones in the drone swarm, including: The restore control command is parsed and processed, and corresponding task content is added to the task queue of the slave drone, thereby adjusting the working status of the slave drone.

Images