Emergency communication resource optimization method for dynamic spectrum allocation

Abstract

The invention provides an emergency communication resource optimization method for dynamic spectrum allocation, which is applied to a self-organizing emergency communication network formed by a plurality of unmanned aerial vehicle nodes and at least one edge aggregation node, and comprises the following steps: an offline pre-training step; a model initialization step; an on-line optimization step: the unmanned aerial vehicle node executes on-line optimization and senses spectrum information of the current environment in real time; on the basis of the frequency spectrum information, performing online fine tuning on the initial local model to update the initial local model; predicting the channel quality of the candidate communication frequency band by using the updated initial local model, and determining a target communication frequency band according to the channel quality; after the target communication frequency band is determined, a consensus mechanism is started, other nodes in the communication network verify selection of the target communication frequency band, and switching to the target communication frequency band is executed after verification is passed. According to the invention, the problems of slow convergence, low decision-making precision and untimely response caused by lack of training samples of the model in an emergency communication task can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to an emergency communication resource optimization method of dynamic spectrum allocation. BACKGROUND

[0002] After natural disasters such as earthquakes and floods, the ground public communication network (such as 4G / 5G base station) often fails in a large area due to physical damage or power interruption, resulting in a "information island" in the rescue site. It has become a key technology in the industry to use a cluster of unmanned aerial vehicles to quickly deploy an air emergency communication network to restore command and dispatch, data backhaul and public communication capabilities.

[0003] However, the electromagnetic environment in the emergency scenario is extremely complex: spectrum unknown: unknown commercial, broadcast or dedicated communication signals in the disaster area, a large number of potential interference sources; spectrum congestion: multiple rescue teams (such as fire, medical, public security) carrying different types of radios working at the same time, which can cause serious congestion and conflict in certain frequency bands; dynamic change: the movement of rescue teams, the addition of new equipment and environmental changes, resulting in a constantly changing spectrum occupation state. The existing technical solution has a key technical bottleneck when applied to the emergency scenario: in the early stage of the task, due to the lack of effective channel samples on site, the local model of each unmanned aerial vehicle node needs to start training from random weights, and the federal learning needs to go through multiple rounds (usually dozens of rounds) of communication and aggregation before the model accuracy converges to a usable level. In the golden time of emergency rescue, this minute-level model initialization delay is unacceptable, which may cause unstable communication links in the early stage, and even miss the best rescue opportunity. SUMMARY

[0004] The technical problem to be solved by the present application is to provide an emergency communication resource optimization method of dynamic spectrum allocation to solve the problem of "cold start" in emergency situations, that is, in emergency communication tasks, the model converges slowly due to the lack of training samples, the decision accuracy is low, and the response is not timely.

[0005] In order to achieve the above technical effects, the present application provides an emergency communication resource optimization method of dynamic spectrum allocation, applied to a self-organizing emergency communication network composed of a plurality of unmanned aerial vehicle nodes and at least one edge aggregation node, comprising the following steps: The offline pre-training step: based on a preset spectrum data set, a general channel feature pre-training model is trained and generated offline, which is used to predict the corresponding channel quality according to the input spectrum information; the model initialization step: when the emergency communication task starts, at least one unmanned aerial vehicle node in the communication network loads the general channel feature pre-training model as the initial local model of the unmanned aerial vehicle node; the online optimization step: the unmanned aerial vehicle node performs online optimization, which includes: real-time sensing of the spectrum information of the current environment; based on the real-time sensed spectrum information, the initial local model is fine-tuned online to update the initial local model; the updated initial local model is used to predict the channel quality of the candidate communication frequency band, and the target communication frequency band is determined according to the predicted channel quality of the candidate communication frequency band; the consensus verification step: after determining the target communication frequency band, a consensus mechanism is started, and the selection of the target communication frequency band is verified by other nodes in the communication network, and the switching to the target communication frequency band is performed after the verification is passed.

[0006] The scheme provided in the application is based on a preset spectrum data set. Before task execution, a general channel feature pre-training model with strong generalization ability is pre-trained offline. In the emergency communication task, the general channel feature pre-training model is deployed as an initial local model in at least one unmanned aerial vehicle node. The unmanned aerial vehicle uses the on-board scanning device to real-time sense the spectrum information of the current environment, generates a three-dimensional spectrum heat map consistent with the input format of the preset spectrum data set, and fine-tunes and updates the initial local model based on the real-time sensed spectrum information, realizes the prediction of the channel quality of the candidate communication frequency band, so as to achieve the effect that the unmanned aerial vehicle has high-precision channel quality prediction ability at the moment of taking off, and determines the target communication frequency band according to the channel quality of the candidate communication frequency band. After determining the target communication frequency band, a consensus mechanism is started, and the selection of the target communication frequency band is verified by other nodes in the communication network, and the switching to the target communication frequency band is performed after the verification is passed, thereby solving the problems of slow convergence, low decision-making precision and untimely response of the model in the emergency communication task due to the lack of training samples, and shortening the rescue time for emergency tasks.

[0007] As an improvement of the technical scheme of the application, the online fine-tuning mode is federated learning; the initial local model is fine-tuned based on the real-time sensed spectrum information to update the initial local model, which includes: the unmanned aerial vehicle node performs local training on the initial local model based on the real-time sensed spectrum information to generate a fine-tuned local model; the unmanned aerial vehicle node interacts with the edge aggregation node in the network to perform federated update on the local model according to the fine-tuned local model and the model information of other unmanned aerial vehicle nodes in the network.

[0008] The scheme provided in the application achieves the effects of rapid adaptation of an existing model and reduction of computing overhead by adopting a federated learning method, using real-time sensed spectrum information of a UAV node to locally train an initial local model to generate a fine-tuned local model, and enabling the local model to quickly adapt to a specific environment in a current task scenario by interacting between the UAV node and an edge aggregation node in the network, and performing federated updating of the local model according to the fine-tuned local model and model information of other UAV nodes in the network.

[0009] As an improvement of the technical scheme of the application, the step of federated updating comprises: the UAV node calculates and encrypts model update information thereof; the encrypted model update information is uploaded to the edge aggregation node; the edge aggregation node aggregates the received model update information to generate global model update parameters; and the UAV node downloads the global model update parameters to update the local model thereof.

[0010] Further, in the above improvement, the model update information is model gradient.

[0011] Further, in the above improvement, the encryption adopts a homomorphic encryption algorithm.

[0012] As an improvement of the technical scheme of the application, further, the preset spectrum data set comprises at least one of the following: a radio signal spectrum data set covering multiple modulation signals and signal-to-noise ratios; and a simulation spectrum data set generated based on a channel modeling technology, the simulation spectrum data set being used to simulate electromagnetic propagation and interference characteristics under a typical disaster area terrain.

[0013] As an improvement of the technical scheme of the application, further, the spectrum information is a three-dimensional spectrum heat map, and the three-dimensional spectrum heat map comprises time, frequency and signal energy dimensions.

[0014] As an improvement of the technical scheme of the application, further, the general channel feature pre-training model is a deep neural network comprising a convolutional neural network (CNN) layer and a long short-term memory network (LSTM) layer.

[0015] As an improvement of the technical scheme of the application, determining the target communication frequency band comprises: using a reinforcement learning algorithm to select the target communication frequency band according to channel quality of the predicted candidate communication frequency band and at least one other preset decision factor.

[0016] Further, the reinforcement learning algorithm is a Q-Learning algorithm; and the at least one other preset decision factor comprises channel capacity of a candidate communication frequency band and / or a task importance level of a UAV node.

[0017] The beneficial effects of the present application are: 1. The scheme provided by the present application is based on a preset spectrum data set. Before task execution, a general channel feature pre-training model with strong generalization ability is pre-trained offline. In an emergency communication task, the general channel feature pre-training model is deployed as an initial local model on at least one unmanned aerial vehicle node. The unmanned aerial vehicle uses an on-board scanning device to realize real-time sensing of the spectrum information of the current environment, generates a three-dimensional spectrum heat map consistent with the input format of the preset spectrum data set, and updates the initial local model based on the real-time sensed spectrum information, realizes prediction of the channel quality of a candidate communication frequency band, so as to achieve the effect that the unmanned aerial vehicle has high-precision channel quality prediction capability at the moment of taking off, and determines a target communication frequency band according to the channel quality of the candidate communication frequency band. After the target communication frequency band is determined, a consensus mechanism is started, and the selection of the target communication frequency band is verified by other nodes in the communication network, and after the verification is passed, switching to the target communication frequency band is performed, thereby solving the problem that in an emergency communication task, the model converges slowly, the decision accuracy is low, and the response is not timely due to the lack of training samples, and shortening the rescue time for emergency tasks.

[0018] 2. The scheme provided by the present application adopts a federated learning method, and based on the spectrum information sensed in real time by the unmanned aerial vehicle node, the initial local model is locally trained to generate a fine-tuned local model, achieving the effects of rapid adaptation of the existing model and reduction of computing overhead. The unmanned aerial vehicle node interacts with the edge aggregation node in the network, updates the local model based on the fine-tuned local model and the model information of other unmanned aerial vehicle nodes in the network, so that the local model can quickly adapt to the specific environment in the current task scenario. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creating laborious work.

[0020] Figure 1 A flowchart of an emergency communication resource optimization method of dynamic spectrum allocation according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] The embodiments of the technical solutions of the present application will be described in detail below. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and therefore only serve as examples, and cannot limit the protection scope of the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0022] The technical scheme provided by the embodiments of the present application is specifically described below with reference to the accompanying drawings. Figure 1 The technical scheme provided by the embodiments of the present application is specifically described below with reference to the accompanying drawings.

[0023] As shown in FIG. 1, which is a flowchart of an emergency communication resource optimization method for dynamic spectrum allocation provided by an embodiment of the present application, applied to a self-organizing emergency communication network composed of a plurality of unmanned aerial vehicle nodes and at least one edge aggregation node, specifically comprising the following steps: Figure 1 S1: offline pre-training step: based on a preset spectrum dataset, an offline training and a general channel feature pre-training model are generated, which is used to predict the corresponding channel quality according to the input spectrum information; In the embodiments of the present application, before performing a task, based on a preset spectrum dataset, an offline training and a general channel feature pre-training model with strong generalization ability are generated. In an implementation mode, in order to realize the high generalization ability of the model, the preset spectrum dataset includes at least one of the following: a radio signal spectrum dataset covering a plurality of modulation signals and signal-to-noise ratios, such as RML2016.10a dataset, and a simulation spectrum dataset generated based on channel modeling technology, wherein the simulation spectrum dataset is used to simulate the electromagnetic propagation and interference characteristics under typical disaster area terrain. S2: model initialization step: at the beginning of the emergency communication task, the general channel feature pre-training model is loaded into at least one unmanned aerial vehicle node in the communication network as the initial local model of the unmanned aerial vehicle node;

[0024] In the embodiments of the present application, at the beginning of the emergency communication task, the general channel feature pre-training model is loaded into the computing unit of at least one unmanned aerial vehicle node in the communication network, so as to serve as the initial local weight model of the unmanned aerial vehicle node. In an implementation mode, in order to construct a deep neural network as a pre-training model, the general channel feature pre-training model is a deep neural network containing a convolutional neural network (CNN) layer and a long short-term memory network (LSTM) layer. Among them, the CNN layer is used to extract the local correlation features of the spectrum data in the frequency dimension and the time dimension, so as to achieve the effect of effectively identifying the "shape" of the interference signal, and the LSTM layer is used to learn the dependence relationship of the correlation features extracted by the CNN layer in the time sequence, so as to achieve the effect of capturing the dynamic change rule of spectrum occupation.

[0025]

[0026] ​Exemplarily, in the training process of the general channel feature pre-training model, mean square error (MSE) can be used as a loss function, an Adam optimizer is used for back propagation training until the model converges on the validation set, so as to obtain the general channel feature pre-training model.

[0027] S3: online optimization step: the UAV node performs online optimization; In an implementation manner, in order to greatly improve the emergency response speed, so that the UAV has high-precision channel quality prediction capability at the moment of taking off, and quickly acquires the specific electromagnetic environment of the current task scene, the step includes the following sub-steps: S31: real-time sensing of spectrum information of the current environment; In the embodiment of the application, the UAV uses the scanning device carried thereby to perform real-time and continuous scanning on a preset frequency range. The data obtained in the scanning process can be converted to generate a three-dimensional spectrum heat map. The format of the three-dimensional spectrum heat map is consistent with the pre-training input format, so as to realize direct data feeding and analysis. In an implementation manner, the three-dimensional spectrum heat map includes time, frequency and signal energy dimensions. The time dimension is used to record time sequence information of the scanning data; the frequency dimension is used to represent frequency components of the scanned signals; and the signal energy dimension is used to quantify the intensity or power of the signals at a specific time and frequency point.

[0028] S32: based on the spectrum information sensed in real time, online fine-tuning is performed on the initial local model to update the initial local model; In an implementation manner, in order to reduce the model computing power and without waiting for multi-round convergence of the federated learning, the step includes the following sub-steps: S321: the UAV node uses the spectrum information sensed in real time to perform local training on the initial local model to generate a fine-tuned local model; In the embodiment of the application, the fine-tuning manner is federated learning. In an implementation manner, the general channel feature pre-training model loaded is quickly trained using the current environment spectrum data collected in real time in the computing unit of the UAV node, that is, the existing spectrum information data is quickly adapted, so as to realize online fine-tuning and updating of the initial local model, and ensure the generalization and specificity of the model in an unknown environment.

[0029] S322: the UAV node interacts with an edge aggregation node in the network to perform federated updating on the local model according to the fine-tuned local model and model information of other UAV nodes in the network.

[0030] In the embodiments of the present application, the local model update information is a model gradient, that is, when a preset global update time node is reached, the unmanned aerial vehicle calculates the gradient increment of the locally fine-tuned model relative to the global model, wherein the federated updating step includes the following substeps: S3221: The unmanned aerial vehicle node calculates and encrypts its model update information; In an implementation manner, in order to encrypt the model gradient and ensure privacy security, specifically, a homomorphic encryption algorithm such as a Paillier algorithm is used.

[0031] S3222: The encrypted model update information is uploaded to the edge aggregation node; Exemplarily, the encrypted model gradient is uploaded to a ground command vehicle or a high-altitude unmanned aerial vehicle serving as an edge aggregation server.

[0032] S3223: The edge aggregation node aggregates the received model update information to generate a global model update parameter; In an implementation manner, the server decrypts and aggregates all model gradients (for example, using a FedAvg algorithm), and further updates the global model.

[0033] S3224: The unmanned aerial vehicle node downloads the global model update parameter to update its local model.

[0034] In an implementation manner, the updated global model parameter is distributed to the unmanned aerial vehicle nodes, and each unmanned aerial vehicle node updates its local model accordingly.

[0035] S33: Using the updated initial local model, the channel quality of the candidate communication frequency band is predicted, and the target communication frequency band is determined according to the predicted channel quality of the candidate communication frequency band; In the embodiments of the present application, the unmanned aerial vehicle uses the updated high-precision initial local model and real-time input spectrum information to predict the channel quality of each selected communication frequency band and determine the target communication frequency band.

[0036] In an implementation manner, determining the target communication frequency band includes: using a reinforcement learning algorithm to select the target communication frequency band according to the predicted channel quality of the candidate communication frequency band and at least one other preset decision factor, wherein the reinforcement learning algorithm is a Q-Learning algorithm; and the at least one other preset decision factor includes the channel capacity of the candidate communication frequency band and / or the task importance level of the unmanned aerial vehicle node.

[0037] Exemplarily, a state space s={channel capacity C of each frequency band, task importance level P of the unmanned aerial vehicle, predicted channel quality I of each frequency band} is constructed. The action space a = {keep the current frequency band, switch to candidate frequency band 1, switch to candidate frequency band 2,...}; Using the Q-Learning algorithm, iteratively solve the optimal action (i.e., the optimal communication frequency band) according to the reward function R(s, a): R(s, a) = C(s, a) + (1 - I(s, a)) - Δ(s, a) α ⋅ C actual + β ⋅(1− I normalized )− gamma ⋅Δ f cost + delta ⋅ P task_bonus Where the reward function R(s, a) is used to quantify the benefits obtained by the UAV after performing action a in state s. The reward function R(s, a) aims to guide the algorithm to minimize unnecessary resource consumption while ensuring communication quality.

[0038] C actual : refers to the actual measured or estimated channel capacity of the current frequency band after performing action a.

[0039] I normalized : refers to the normalized actual interference level (e.g., reverse mapping the signal-to-noise ratio to 0-1, the higher the value, the greater the interference) after performing action a.

[0040] Δ f cost : refers to the overhead generated by the frequency band switching operation, including but not limited to switching delay, instantaneous packet loss rate, energy consumption increment, etc. If the instruction in the task is "keep the current frequency band", this item is zero.

[0041] P task_bonus : based on the importance level of the current task P task Set additional rewards or penalties. For example, when the communication quality reaches the preset threshold in an important task, give additional rewards; conversely, when the important task communication is damaged, give additional penalties.

[0042] α 、 β 、 gamma 、 delta : are all preset weight coefficients, used to adjust the relative importance of each factor in the reward function. These weights can be optimized and adjusted through offline simulation, expert experience or heuristic algorithms to meet the needs of different application scenarios.

[0043] S4: consensus verification step: after determining the target communication frequency band, a consensus mechanism is started, and the selection of the target communication frequency band is verified by other nodes in the communication network, and switching to the target communication frequency band is performed after verification.

[0044] In an implementation manner, after the UAV node selects the optimal frequency band, in order to prevent malicious or faulty nodes from making wrong decisions, a consensus mechanism is triggered, and the selection of the target communication frequency band is verified by other UAV nodes in the communication interaction network, for example, "2f+1" nodes (f is the number of tolerable malicious nodes) are randomly selected from the adjacent UAV nodes as verification nodes, and after more than 2 / 3 of the verification nodes confirm the decision, the frequency band switching is formally performed and the adjacent UAV nodes are announced, and optionally, the consensus mechanism can be a Byzantine fault-tolerant consensus mechanism.

[0045] As can be seen from the above embodiments, the scheme provided by the present application is based on a preset frequency spectrum data set. Before task execution, a general channel feature pre-training model with strong generalization ability is pre-trained offline. In the emergency communication task, the general channel feature pre-training model is deployed as an initial local model in at least one UAV node. The UAV uses the on-board scanning device to realize real-time sensing of the frequency spectrum information of the current environment, generates a three-dimensional frequency spectrum heat map consistent with the input format of the preset frequency spectrum data set, and based on the real-time sensed frequency spectrum information, the initial local model is fine-tuned and updated online, so that it can quickly adapt to the specific environment in the current task scenario, and realize the prediction of the channel quality of the candidate communication frequency band, so as to achieve the effect that the UAV has high-precision channel quality prediction ability at the moment of taking off, and determine the target communication frequency band according to the channel quality of the candidate communication frequency band. After determining the target communication frequency band, a consensus mechanism is started, and the selection of the target communication frequency band is verified by other nodes in the communication network, and switching to the target communication frequency band is performed after verification, thereby solving the problem that in the emergency communication task, the model converges slowly, the decision accuracy is low, and the response is not timely due to the lack of training samples, and shortening the rescue time for emergency tasks.

[0046] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for optimizing emergency communication resources through dynamic spectrum allocation, applied to a self-organizing emergency communication network consisting of several UAV nodes and at least one edge aggregation node, characterized in that, Includes the following steps: Offline pre-training step: Based on a preset spectrum dataset, an offline training and generation of a general channel feature pre-training model is performed. The general channel feature pre-training model is used to predict the corresponding channel quality based on the input spectrum information. Model initialization steps: At the start of the emergency communication mission, the general channel feature pre-trained model is loaded into at least one UAV node in the communication network as the initial local model of the UAV node; Online optimization steps: The drone node performs online optimization, which includes: Real-time perception of the spectrum information of the current environment; Based on the real-time perceived spectrum information, the initial local model is fine-tuned online to update the initial local model; Using the updated initial local model, the channel quality of candidate communication frequency bands is predicted, and the target communication frequency band is determined based on the predicted channel quality of the candidate communication frequency bands. Consensus verification steps: After determining the target communication frequency band, a consensus mechanism is initiated, and other nodes in the communication network verify the selection of the target communication frequency band. After successful verification, the switching to the target communication frequency band is executed.

2. The method according to claim 1, characterized in that, The online fine-tuning method is federated learning; The online fine-tuning of the initial local model based on the real-time sensed spectrum information to update the initial local model includes: The drone node uses the real-time perceived spectrum information to train the initial local model locally, so as to generate a fine-tuned local model. The drone node interacts with edge aggregation nodes in the network to perform federated updates to the local model based on the fine-tuned local model and model information from other drone nodes in the network.

3. The method according to claim 2, characterized in that, The steps of the federal update include: The drone node calculates and encrypts its model update information; The encrypted model update information is uploaded to the edge aggregation node; The edge aggregation node aggregates the received model update information to generate global model update parameters; The drone node downloads the global model update parameters to update its local model.

4. The method according to claim 3, characterized in that, The model update information is the model gradient.

5. The method according to claim 4, characterized in that, The encryption uses a homomorphic encryption algorithm.

6. The method according to claim 1, characterized in that, The preset spectrum dataset includes at least one of the following: A radio signal spectrum dataset covering various modulation signals and signal-to-noise ratios; The simulated spectrum dataset, generated based on channel modeling technology, is used to simulate the electromagnetic propagation and interference characteristics under typical disaster area terrain.

7. The method according to claim 1, characterized in that, The spectrum information is a three-dimensional spectrum heatmap, which includes time, frequency, and signal energy dimensions.

8. The method according to claim 1, characterized in that, The general channel feature pre-training model is a deep neural network containing convolutional neural network (CNN) layers and long short-term memory network (LSTM) layers.

9. The method according to claim 1, characterized in that, Determining the target communication frequency band includes: The target communication frequency band is selected by using a reinforcement learning algorithm based on the channel quality of the predicted candidate communication frequency band and at least one other preset decision factor.

10. The method according to claim 9, characterized in that, The reinforcement learning algorithm is the Q-Learning algorithm; The at least one other preset decision factor includes the channel capacity of the candidate communication frequency band and / or the mission importance level of the UAV node.

Images