A Federated Learning Semantic Communication Method for Unmanned Aerial Vehicle Swarms

Abstract

This invention discloses a federated learning semantic communication method for drone swarms. First, a multi-scale joint source-channel encoder is designed based on the characteristics of traffic aerial images to achieve efficient extraction and compression of semantic features. Then, a low-bit quantization weighted convolutional neural network is trained locally on each drone, uploading only the quantized values ​​of parameter changes to reduce energy consumption and protect privacy. Next, an edge server aggregates and globally synchronizes the parameters uploaded by each drone, improving system coordination efficiency. Finally, through energy consumption modeling and experimental verification, it is demonstrated that this method effectively improves reconstruction quality and significantly reduces overall energy consumption, exhibiting excellent robustness. This invention enables efficient semantic communication in multi-drone traffic scenarios with low energy consumption and high privacy protection, providing technical support for the design and optimization of future AI-assisted 6G intelligent transportation communication systems.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a federated learning semantic communication method for unmanned aerial vehicle (UAV) swarms. Background Technology

[0002] With the rapid development of 6G mobile communication and intelligent transportation, drones are widely used in traffic monitoring, road condition reconnaissance, and other fields. Traditional drone traffic reconnaissance systems typically require transmitting large amounts of aerial image data collected via wireless links to a central server in real time for analysis and processing. This method, which relies primarily on uploading raw data, not only puts immense pressure on communication bandwidth but also results in extremely high energy consumption during data transmission due to the instability of wireless channels, severely impacting the drone's endurance.

[0003] In recent years, semantic communication (SemCom) has received widespread attention as a novel communication paradigm. This technology significantly improves communication efficiency and effectively reduces energy consumption by directly transmitting semantic information from data. Meanwhile, federated learning (FL), as a distributed machine learning method, allows terminal devices to train models locally and only upload parameters or model updates while retaining data locally, protecting data privacy and reducing data transmission volume. Some scholars have combined federated learning and semantic communication to explore the joint design of edge-side intelligent coding and distributed optimization models. However, existing federated learning-based UAV semantic communication systems still have the following shortcomings:

[0004] Existing solutions have limited ability to extract semantic features from aerial traffic images, making it difficult to fully capture the information structure in complex traffic scenarios and affecting reconnaissance and decision-making effectiveness. Secondly, existing models generally use full-precision parameters, resulting in high computational and transmission energy consumption, which is not suitable for the actual deployment of energy-constrained terminals such as drones. In addition, complex channel environments and the heterogeneity of multiple terminals lead to insufficient system robustness, semantic information is easily interfered with, and overall performance degrades significantly.

[0005] Therefore, how to achieve an efficient, low-power, and robust federated learning semantic communication method for multi-UAV traffic reconnaissance scenarios while ensuring data privacy and controllable system energy consumption has become a key technical problem that urgently needs to be solved. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a federated learning semantic communication method for UAV swarms. First, a multi-scale joint source-channel encoder is designed based on the characteristics of traffic aerial images to achieve efficient extraction and compression of semantic features. Then, a low-bit quantization weighted convolutional neural network is trained locally on each UAV, uploading only the quantized values ​​of parameter changes to reduce energy consumption and protect privacy. Next, an edge server aggregates and globally synchronizes the parameters uploaded by each UAV, improving system coordination efficiency. Finally, through energy consumption modeling and experimental verification, it is demonstrated that this method effectively improves reconstruction quality and significantly reduces overall energy consumption, exhibiting excellent robustness. This invention enables efficient semantic communication in multi-UAV traffic scenarios with low energy consumption and high privacy protection, providing technical support for the design and optimization of future AI-assisted 6G intelligent transportation communication systems.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows:

[0008] Step 1: Construct a collaborative semantic communication system architecture consisting of multiple UAV terminals and edge servers. Using end-to-end image reconstruction error as the global optimization target, uniformly model the global loss function and overall energy consumption of the system, and clarify the energy consumption and performance constraints of each link in the system.

[0009] Step 2: Design of an energy-efficient multi-scale semantic communication system, including the design of key modules such as multi-scale convolutional semantic feature extraction and quantization encoding;

[0010] Step 3: Federated learning training and overall system optimization, including local quantization training, differential parameter uploading, server aggregation, and distributed system training convergence optimization.

[0011] Further, step 1 specifically includes:

[0012] Step 1-1: Construct a multi-user collaborative semantic communication system consisting of multiple UAV terminals (equipment side) and edge servers (server side); each UAV terminal collects local traffic aerial images, performs multi-scale joint source-channel coding feature extraction and model training on the data only locally, and does not upload the original image data; the edge server undertakes the task of model aggregation and distribution.

[0013] Steps 1-2: The system uses end-to-end image reconstruction error as the optimization objective to establish joint loss modeling for multi-UAV, distributed semantic communication scenarios. The global loss objective is:

[0014]

[0015] in, It is the global loss after server aggregation. This is the global dataset size, per user. Each has a containing Local dataset of data samples , User have The input sample vector of each feature, It corresponds to the sample input value. The tag, This indicates that all users train the same set of shared encoder-decoder model parameters. The number of parameters in the model is represented by the user set. It indicates that its size is ;

[0016] Steps 1-3: The energy consumption for local computation during federated learning training modeling and the energy consumption for communication between the UAV and the uploaded parameters are as follows: The energy consumption for local computation of the device is:

[0017]

[0018] in User The number of iterations is calculated locally at that location. For users The effective capacitance coefficient of the chipset is calculated. User Calculate the number of CPU cycles required to obtain one sample of data; For users Its computing power is measured in CPU cycles per second.

[0019] After completing local calculations, each drone terminal uploads its local FL model parameters to the edge server via the OFDMA channel for parameter updates. The upload speed is:

[0020]

[0021] in It is the first Assigned to device during round communication bandwidth, It is equipment The average transmit power, It is equipment Channel gain between the server and the server; It is the power spectral density of Gaussian noise, and Shannon capacity is used as the upper limit of transmission rate;

[0022] Because each user trains a neural network with the same structure, the size of the data uploaded by each user is constant, represented as... The unit is bit, the first During round-robin communication, the user The time for uploading model parameters is The formula for calculating uplink communication transmission energy consumption is:

[0023]

[0024] The number of global iterations in the federated learning training process is denoted as... Therefore, the total energy consumption of all terminals participating in FL is expressed as:

[0025] .

[0026] Furthermore, step 2 specifically includes:

[0027] Step 2-1: Employ the Multi-Scale Joint Source-Channel Encoder (MSJSCC), which includes a multi-scale parallel convolution module and an attention mechanism, to extract the input image. Multi-level semantic features are used to achieve efficient compression and enhanced channel robustness; the encoding process expression is:

[0028]

[0029] in As a semantic encoder implemented in deep learning models, These are the relevant learning parameters of the model, where This indicates that the input image is mapped into complex numerical symbols suitable for transmission over a wireless channel;

[0030] The decoding process expression is:

[0031]

[0032] in As a semantic decoder implemented in deep learning models, These are the relevant learning parameters of the model. This represents a complex signal affected by the combined effects of channel fading and noise. This represents an estimate of the original input image;

[0033] The decoder structure introduces multi-scale convolutional layers to reconstruct multi-scale features of different feature maps. Finally, the output is limited to the range [0,1] by the Sigmoid activation function to ensure that the reconstructed image conforms to the standard image format.

[0034] Step 2-2: Perform low-bit quantization on the neural network parameters and gradients, using a pass-through estimator (STE) to quantize the weights. Fixed point value, input real number Quantified as Real number at fixed point STE is defined as follows:

[0035]

[0036]

[0037] in Represent the objective function;

[0038] The gradient quantization formula is:

[0039]

[0040]

[0041] in yes One of the numbers, The maximum value was taken from all weights in each layer. Will Quantization to a range of values of Fixed-point number, affine transformation will The range is mapped to [-1, 1];

[0042] Steps 2-3: After the drone is trained locally, only the quantized values ​​of the changes in model parameters are uploaded to reduce communication bandwidth and energy consumption, while ensuring data privacy.

[0043] Furthermore, step 3 specifically includes:

[0044] Step 3-1: The server periodically receives the quantized model difference values ​​uploaded by each terminal and uses an aggregation algorithm to calculate the updated global model.

[0045]

[0046] in This represents the difference between the locally updated model and the latest global model calculated by the user. The quantization result of bit, Indicates that the drone is in The model parameters received at the start of each training round. Indicates that the drone is in The model parameters received at the start of each training round, where This represents the set of users selected by the central server to participate in training in each round of communication. Indicates the size of the user set;

[0047] Step 3-2: The global model is synchronously distributed to each UAV to achieve distributed collaborative optimization and model consistency, thereby improving the overall system performance.

[0048] Preferably, the pass-through estimator STE is .

[0049] Preferably, the aggregation algorithm is FedAvg.

[0050] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to enable the electronic device to perform the above-described federated learning semantic communication method.

[0051] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above-described federated learning semantic communication method.

[0052] A chip includes a processor for retrieving and running a computer program from a memory, causing a device on which the chip is mounted to perform the above-described federated learning semantic communication method.

[0053] A computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the above-described federated learning semantic communication method.

[0054] The beneficial effects of this invention are as follows:

[0055] This invention significantly reduces the computational and communication energy consumption of UAVs during model training and parameter uploading by using multi-scale joint source-channel coding and low-bit quantization neural networks. It enables efficient semantic communication in multi-UAV traffic scenarios with low energy consumption and high privacy protection, significantly extending the endurance of a single UAV flight mission and providing technical support for the design and optimization of future AI-assisted 6G intelligent transportation communication systems. Attached Figure Description

[0056] Figure 1 This is a system model diagram of a federated learning semantic communication system for multi-UAV traffic reconnaissance.

[0057] Figure 2 This invention is a multi-scale joint source-channel encoder.

[0058] Figure 3 This is a comparison chart of image reconstruction quality using different schemes and quantization levels.

[0059] Figure 4 This is a comparison chart of total energy consumption for different schemes and different quantization levels. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] This invention provides a federated learning semantic communication method for drone swarms. First, a multi-scale joint source-channel encoder is designed based on the characteristics of traffic aerial images to achieve efficient extraction and compression of semantic features. Then, a low-bit quantization weighted convolutional neural network is trained locally on each drone, uploading only the quantized values ​​of parameter changes, reducing energy consumption and protecting privacy. Furthermore, an edge server aggregates and globally synchronizes the parameters uploaded by each drone, improving system coordination efficiency. Finally, through energy consumption modeling and experimental verification, it is demonstrated that this method effectively improves reconstruction quality and significantly reduces overall energy consumption, exhibiting excellent robustness. This invention enables efficient semantic communication in multi-drone traffic scenarios with low energy consumption and high privacy protection, providing technical support for the design and optimization of future AI-assisted 6G intelligent transportation communication systems.

[0062] This invention proposes a federated learning semantic communication method for multi-UAV traffic reconnaissance. Combining the characteristics of traffic aerial images, a multi-scale joint source-channel encoder is designed to achieve efficient extraction and compression of semantic features. Then, a low-bit quantization convolutional neural network is trained locally on each UAV, uploading only the quantized values ​​of parameter changes. This invention enables efficient semantic communication in multi-UAV traffic scenarios with low power consumption and high privacy protection.

[0063] Step S1: Construct a collaborative semantic communication system architecture consisting of multiple UAV terminals and edge servers. Using end-to-end image reconstruction error as the global optimization target, uniformly model the global loss function and overall energy consumption of the system, and clarify the energy consumption and performance constraints of each link in the system.

[0064] Step S101: System architecture setup. The system involved in this invention includes an edge server and... Each drone terminal collects local traffic aerial images and performs distributed model training. An edge server is responsible for global model aggregation and distribution, enabling collaborative intelligent communication among multiple terminals. For ease of subsequent explanation, the drone ensemble is referred to as... It indicates that its size is Each drone Each has a containing Local dataset of data samples ,in It is a terminal have The input sample vector of each feature, It corresponds to the sample input value. The tag.

[0065] Step S102: Establishing the end-to-end loss function and optimization objective. For semantic communication systems, semantic information extraction and reconstruction are considered crucial to the system's success. Therefore, this invention uses a UAV local model with the original image. With reconstructed images The mean squared error (MSE) is the loss function, which is defined as:

[0066]

[0067] in It is the first Local training loss of each drone It is the original image vector. It reconstructs the image vector. This represents the length of the image vector. The smaller the MSE, the closer the reconstructed image is to the original image, and the better the image reconstruction quality.

[0068] Therefore, the global federated training loss can be obtained by weighting the loss of each terminal based on the size of each user's dataset. The specific calculation formula is as follows:

[0069]

[0070] in, It is the global loss after server aggregation. It is the size of the global dataset. The goal of FL training is to find the globally optimal model that minimizes the sum of the training losses of all data.

[0071] Each terminal trains the same set of shared encoding / decoding model parameters, represented as: ,in Indicates the number of parameters in the model. Let Representation and in the model Data samples calculated above The corresponding objective function. Therefore, the first... The objective function for each terminal is defined by the following equation:

[0072]

[0073] in .

[0074] In this case, the entire federated optimization model is defined as:

[0075]

[0076] in Indicates the first The weight of each user , .

[0077] During the model training phase, each terminal needs to repeatedly perform gradient calculations and model updates, involving a large number of local computations and communication operations. The inference phase typically only requires one model inference operation, with lower computation and communication requirements. By optimizing the computation and communication energy consumption during the training phase, the need for frequent terminal charging can be directly reduced, providing greater ease of use for resource-constrained devices. Therefore, this invention only considers the energy consumption of each device during federated learning training, which is divided into local computation energy consumption and transmission energy consumption.

[0078] Assumption (cycles / sample) is the user's The number of CPU cycles required to calculate a sample of data for the user The number of CPU cycles required to perform one local iteration on all data samples is .make For users Its computing power is measured in CPU cycles per second (cycles / s). It is equipment The number of local iterations at the location, user In the During round-robin communication, the energy consumption for local model computation is:

[0079]

[0080] in User The number of iterations is calculated locally at that location. For users The effective capacitance coefficient of the chipset is calculated.

[0081] After completing local calculations, each user's terminal uploads its local FL model parameters to the edge server via an Orthogonal Frequency Division Multiple Access (OFDMA) channel for parameter updates. The upload speed is:

[0082]

[0083] in It is the first Assigned to device during round communication bandwidth, It is equipment The average transmit power, It is equipment Channel gain between the server and the server It is the power spectral density of Gaussian noise, and Shannon capacity is used as the upper limit of transmission rate.

[0084] Because each user trains a neural network with the same structure, the size of the data uploaded by each user is constant, represented as... The unit is bit. During round-robin communication, the user The time for uploading model parameters is The formula for calculating uplink communication transmission energy consumption is:

[0085]

[0086] The number of global iterations in the federated learning training process is denoted as... Therefore, the total energy consumption of all users participating in FL is expressed as:

[0087]

[0088] Step S2: Through the design of a multi-scale semantic communication system architecture, efficient extraction, compression, and channel adaptation of image features are achieved. Based on this, a low-bit quantization mechanism is employed to achieve efficient collaborative design of local computation and quantized uploading of model parameters.

[0089] Step S201: Construction of the Multi-Scale Joint Source-Channel Coding (MSJSCC) uses a cross-spatial learning method to fuse the outputs of two parallel sub-networks, thereby reshaping and grouping feature maps at different scales along the channel dimension. Distributing spatial semantic features evenly across each feature group effectively captures the global contextual information of all pixels in the image and improves image quality.

[0090] The encoder contains multi-scale parallel convolutional modules and attention mechanisms to extract input images. The encoder utilizes multi-level semantic features to achieve efficient compression and enhanced channel robustness. Mapped into complex numerical symbols suitable for wireless channel transmission The decoder is responsible for processing the received complex-valued signal that is interfered with by channel noise. Recover the original image The entire system is designed so that the encoder and decoder can be jointly optimized during training to minimize the difference between the recovered image and the original input image.

[0091] The statistical data of the input image is usually unknown at the decoder. The pixel values ​​in the range [0,1] are generated by normalizing the maximum pixel value of 255. Then, multi-scale features of the image are extracted by modules such as layer-by-layer convolution, feature fusion, and attention mechanism to further reduce the amount of information transmitted and realize the encoding of the image.

[0092] The encoder uses a deterministic coding function Will 3D input image Mapping to complex-valued channel input samples of Length vector:

[0093]

[0094] in As a semantic encoder implemented in deep learning models, These are the relevant learning parameters of the model.

[0095] 3x3 convolutional kernels excel at capturing subtle features, while 5x5 convolutional kernels can perceive a wider range of contextual information. The encoder's initial convolutional layer uses 5x5 convolutional kernels to extract basic image features, and then uses 3x3 and 5x5 convolutional kernels in parallel to extract features. An efficient multi-scale attention (EMA) module reshapes and groups feature maps of different scales along the channel dimension, ensuring that spatial semantic features are evenly distributed within each feature group. A cross-spatial learning method is used to fuse the output feature maps of the two parallel sub-networks, effectively capturing global contextual information of all pixels in the image, improving the model's feature extraction efficiency, and thus effectively conveying image semantic information.

[0096] The Multi-Scale Attention (EMA) module integrates global and local information by aggregating the feature responses of different convolutional layers, effectively improving the model's ability to jointly represent large objects and detailed structures. In multi-drone scenarios, the EMA module enables the model to focus on the overall traffic environment while capturing subtle road changes or abnormal targets, improving the robustness of semantic information compression and recognition in communication images.

[0097] The Convolutional Block Attention Module (CBAM) is used during feature extraction to enhance important features and suppress useless features, thereby improving the model's semantic feature extraction efficiency. CBAM automatically learns which channels or spatial locations are most important to the current task by sequentially embedding channel attention and spatial attention mechanisms.

[0098] Channel attention first aggregates spatial information for each channel and dynamically adjusts the importance of features in different channels to filter important features. Spatial attention then further highlights "what is useful" in the image based on the aggregated information from the channels, strengthening salient regions.

[0099] The output of the last convolutional layer is 2 It consists of several units, followed by a normalization layer, and the output of the encoder's last convolutional layer. Normalization is performed according to the following formula to achieve average power constraint. :

[0100]

[0101] in, yes The conjugate transpose of .

[0102] After the encoding operation, the output of the normalization layer is combined into One complex-valued channel input sample When transmitted over a communication channel, channel noise introduces random interference into the transmitted symbols. The communication channel can be modeled as a series of untrainable layers, defined by a transfer function. express.

[0103] The channel signal-to-noise ratio is defined as:

[0104]

[0105] in The channel input signal after applying a power normalization layer at the encoder of the proposed MSJSCC scheme. The average power, the variance of the variable Power corresponding to channel noise.

[0106] The receiver is implemented using a joint source-channel decoder to achieve high-quality image reconstruction, employing a decoding function. The complex-valued signal that has been affected by the combined effects of channel fading and noise Estimation mapped to the original input image :

[0107]

[0108] in As a semantic decoder implemented in deep learning models, These are the relevant learning parameters of the model.

[0109] The decoder structure design focuses on the recovery of semantic features. It introduces multi-scale transposed convolutional layers to reconstruct multi-scale features of different feature maps. Finally, the output is limited to the range [0,1] by the Sigmoid activation function to ensure that the reconstructed image conforms to the standard image format.

[0110] Step S202: Perform low-bit quantization on the neural network parameters and gradients, using a straight-through estimator (STE) to quantize the weights. The fixed-point value, the through-pass estimator widely used in this invention is It will input real numbers Quantified as Real number at fixed point Using fixed-point integer dot products in equations and applying appropriate scaling can efficiently compute the dot product of two sequences of real numbers.

[0111]

[0112]

[0113] in This represents the objective function.

[0114] In the model, each terminal Training with the same structure Bit quantization neural network Quantizing the network to Before the position, use The function restricts the range of network weights to [-1, 1]. This is done at the cost of increased energy consumption. This allows for higher precision. Quantization is a non-differentiable operation; without STE, gradients cannot be propagated back through quantization during backpropagation, thus preventing updates to the network weights.

[0115]

[0116]

[0117] in yes One of the numbers, The maximum value was taken from all weights in each layer. Will Quantization to a range of values of Fixed-point number, affine transformation will The range is mapped to [-1, 1].

[0118] For users , will the The quantization weights of the layer network are represented as ,in It is the first The parameters of the layer. Therefore, the first... The output of the layer will be:

[0119]

[0120] in It is the output of the previous layer. It is the first Layer operations on input, Layer output It is passed forward as input to the next layer.

[0121] Gradient quantization is performed only during the backpropagation of the neural network, applying the following STE to the output of each convolutional layer:

[0122]

[0123]

[0124] The gradient is unbounded, and the values ​​are passed to a differentiable nonlinear function to map the weight range to... To compensate for potential biases caused by gradient quantization, a simulated noise function is introduced. ,in Finally, the gradient will be used to convert... The expression for the number of digits is as follows:

[0125]

[0126] in It is the backpropagation gradient of the output r of a certain layer, which is maximized over all axes (except the batch axis) of the gradient tensor dr.

[0127] Step S3: Federated learning training and overall system optimization. Design the server-side federated aggregation algorithm and the overall system optimization process to complete the multi-round synchronization and distributed update of the global model.

[0128] Step S301: The training process of federated learning involves multiple rounds of communication, the first... Wheel communication involves the following three steps:

[0129] Model distribution: The central server samples and selects users to participate in each round of training from all devices; the selected users are denoted as... ;

[0130] Local training: Each selected device Received the latest model After that, firstly, it was carried out. Bit quantization yields the model Based on this model, multiple rounds of local training are performed using local data:

[0131]

[0132] in For learning rate, The local iteration count is [number], and the local optimization on each device uses gradient descent, which means that the local loss function is minimized through multiple iterations. .

[0133] The user calculates the difference between the locally updated model and the latest global model. - For energy-constrained devices, transmitting with full precision It's impractical to proceed. bit quantization Uploaded to the central server.

[0134] Central aggregation: After receiving the model updates uploaded by each device, the server performs global aggregation.

[0135]

[0136] This process is repeated until the model converges, completing the training, updating, distribution, and deployment of the semantic communication system.

[0137] Example:

[0138] Dataset: In this invention, the VisDrone2019 dataset is used to evaluate the performance of the proposed federated learning semantic communication system for multi-UAV traffic reconnaissance. The VisDrone2019 dataset was collected by the AISKYEYE team at the Machine Learning and Data Mining Laboratory of Tianjin University, and was captured by various UAV cameras, covering a wide range of subjects. The VisDrone2019 dataset contains 7019 images, divided into 10 categories: pedestrians, people, cars, vans, buses, trucks, motorcycles, bicycles, chauffeured tricycles, and tricycles. The training and validation sets contain 6,471 and 548 images, respectively. This dataset mainly contains small to medium-sized targets. The dataset is distributed uniformly, randomly, and non-overlappingly to each UAV. During training, the learning rate and minibatch size are set to 0.00005 and 64, respectively. The number of users is 20, the number of local iterations is 5, the compression ratio is 1 / 6, and the number of communication rounds is 80. The computing power of each device is also considered. capacitance coefficient Calculate the number of periods required for one sample. In respectively cycles / s , Randomly selected from clcles / sample, total bandwidth MHz.

[0139] To demonstrate the superiority of the method of this invention, a typical Deep Joint Source-Channel Code (JSCC) method and a traditional communication algorithm are used as comparison algorithms. The traditional communication method uses JPEG to compress the source, LDPC as the channel code, and the modulation and demodulation method is Quadrature Amplitude Modulation (QAM) with a modulation order of 4.

[0140] Without compromising universality, the effectiveness of the proposed system in reducing energy consumption will be investigated. The quantization level of the neural network is... , The system was trained at a medium-to-high signal-to-noise ratio (SNR) level of 19 dB, reflecting typical wireless channel conditions. This was compared with full-precision training of the neural network, full-precision uploading of model parameter results, and traditional communication methods. Figure 3 It can be seen that the JPEG+LDPC+QAM method performs poorly at low signal-to-noise ratios (SNR) and its performance saturates at high SNRs. In contrast, the two JSCC algorithms preserve the semantic information of the image to the greatest extent, ensuring effective image transmission. The system proposed in this invention can extract image details at different levels and granularities, thus outperforming JSCC by approximately 6 dB, which helps the model to understand the image content more comprehensively. Furthermore, the quantization weights limit the model's capacity, reducing overfitting and distortion, and improving image quality.

[0141] Figure 4 Three schemes and comparison charts of total energy consumption for different quantization levels are presented. Traditional communication methods focus on the accuracy of each bit transmission, resulting in higher energy consumption. The MSJSCC system in this invention can tolerate transmission errors, significantly reducing energy consumption by 48.17% compared to full-precision training, while the JSCC method reduces energy consumption by 48.30%, demonstrating the significant energy-saving effect of quantization. The MSJSCC scheme can extract local and global features in parallel and dynamically fuse them with the EMA module to reduce computational redundancy caused by repeatedly stacking convolutional layers. Simultaneously, spatial semantic features are uniformly distributed through feature grouping.

Claims

1. A federated learning semantic communication method for a UAV swarm, characterized in that, Includes the following steps: Step 1: Construct a collaborative semantic communication system architecture consisting of multiple UAV terminals and edge servers. Using end-to-end image reconstruction error as the global optimization target, uniformly model the global loss function and overall energy consumption of the system, and clarify the energy consumption and performance constraints of each link in the system. Step 2: Design of an energy-efficient multi-scale semantic communication system, including multi-scale convolutional semantic feature extraction and design of key modules for quantization encoding; Step 2-1: Employ the Multi-Scale Joint Source-Channel Encoder (MSJSCC), which includes a multi-scale parallel convolution module and an attention mechanism, to extract the input image. Multi-level semantic features are used to achieve efficient compression and enhanced channel robustness; the encoding process expression is: ； in As a semantic encoder implemented in deep learning models, These are the relevant learning parameters of the model, where This indicates that the input image is mapped into complex numerical symbols suitable for transmission over a wireless channel; The decoding process expression is: ； in As a semantic decoder implemented in deep learning models, These are the relevant learning parameters of the model. This represents a complex signal affected by the combined effects of channel fading and noise. This represents an estimate of the original input image; The decoder structure introduces multi-scale convolutional layers to reconstruct multi-scale features of different feature maps. Finally, the output is limited to the range [0,1] by the Sigmoid activation function to ensure that the reconstructed image conforms to the standard image format. Step 2-2: Perform low-bit quantization on the neural network parameters and gradients, using a pass-through estimator (STE) to quantize the weights. Fixed point value, input real number Quantified as Real number at fixed point STE is defined as follows: ； ； in Represent the objective function; The gradient quantization formula is: ； ； in yes One of the numbers, The maximum value was taken from all weights in each layer. Will Quantization to a range of values of Fixed-point number, affine transformation will The range is mapped to [-1, 1]; Steps 2-3: After the drone is trained locally, only the quantized values ​​of the changes in model parameters are uploaded to reduce communication bandwidth and energy consumption, while ensuring data privacy. Step 3: Federated learning training and overall system optimization, including local quantization training, differential parameter uploading, server aggregation, and distributed system training convergence optimization.

2. The federated learning semantic communication method for UAV swarms according to claim 1, characterized in that, Step 1 specifically involves: Step 1-1: Construct a multi-user collaborative semantic communication system consisting of multiple UAV terminals (equipment side) and edge servers (server side); each UAV terminal collects local traffic aerial images, performs multi-scale joint source-channel coding feature extraction and model training on the data only locally, and does not upload the original image data; the edge server undertakes the task of model aggregation and distribution. Steps 1-2: The system uses end-to-end image reconstruction error as the optimization objective to establish joint loss modeling for multi-UAV, distributed semantic communication scenarios. The global loss objective is: ； in, It is the global loss after server aggregation. This is the global dataset size, per user. Each has a containing Local dataset of data samples , User have The input sample vector of each feature, It corresponds to the sample input value. The tag, This indicates that all users train the same set of shared encoder-decoder model parameters. The number of parameters in the model is represented by the user set. It indicates that its size is ; Steps 1-3: The energy consumption for local computation during the federated learning training process modeling and the energy consumption for communication between the UAV and the uploaded parameters. The energy consumption for local computation on the device is as follows: ； in User The number of iterations is calculated locally at that location. For users The effective capacitance coefficient of the chipset is calculated. User Calculate the number of CPU cycles required to obtain one sample of data; For users Its computing power is measured in CPU cycles per second; After completing local calculations, each drone terminal uploads its local FL model parameters to the edge server via the OFDMA channel for parameter updates. The upload speed is: ； in It is the first Assigned to device during round communication bandwidth, It is equipment The average transmit power, It is equipment Channel gain between the server and the server; It is the power spectral density of Gaussian noise, and Shannon capacity is used as the upper limit of transmission rate; Because each user trains a neural network with the same structure, the size of the data uploaded by each user is constant, represented as... The unit is bit, the first During round-robin communication, the user The time for uploading model parameters is The formula for calculating uplink communication transmission energy consumption is: ； The number of global iterations in the federated learning training process is denoted as... Therefore, the total energy consumption of all terminals participating in FL is expressed as: 。 3. The federated learning semantic communication method for UAV swarms according to claim 2, characterized in that, Step 3 specifically involves: Step 3-1: The server periodically receives the quantized model difference values ​​uploaded by each terminal and uses an aggregation algorithm to calculate the updated global model. ； in This represents the difference between the locally updated model and the latest global model calculated by the user. The quantization result of bit, Indicates that the drone is in The model parameters received at the start of each training round. Indicates that the drone is in The model parameters received at the start of each training round, where This represents the set of users selected by the central server to participate in training in each round of communication. Indicates the size of the user set; Step 3-2: The global model is synchronously distributed to each UAV to achieve distributed collaborative optimization and model consistency, thereby improving the overall system performance.

4. The federated learning semantic communication method for UAV swarms according to claim 3, characterized in that, The pass-through estimator STE is .

5. A federated learning semantic communication method for unmanned aerial vehicle (UAV) swarms according to claim 4, characterized in that, The aggregation algorithm is FedAvg.

6. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

8. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the method as described in any one of claims 1 to 5.

Images