Anti-interference aerial computing power allocation method and device, equipment and medium

Abstract

The application relates to an anti-interference aerial computing power distribution method and device, equipment and medium. The method comprises the following steps: inputting a composite feature vector into an aerial computing power distribution model which has been trained to convergence; performing high-dimensional feature nonlinear mapping on a hidden layer of the aerial computing power distribution model to obtain an intermediate feature vector; adopting a preset hard saturation activation function to map the intermediate feature vector into a preset normalization interval on an output layer of the aerial computing power distribution model; scaling the normalized output value according to the maximum transmission power of an edge device to obtain a target transmission power of each edge device; and outputting the target transmission power so that each edge device performs precoding according to the target transmission power and a channel phase thereof to concurrently transmit data at the same time and in the same frequency band. The aerial computing time of the application almost does not increase with the increase of the scale of the edge device, and the speed is accelerated by several orders of magnitude.

Description

Technical Field

[0001] This application relates to the field of in-flight computing, and more particularly to an interference-resistant in-flight computing power allocation method, corresponding apparatus, electronic equipment, and computer-readable storage medium. Background Technology

[0002] To meet the processing needs of massive amounts of sensing data in modern Internet of Things (IoT) and edge computing applications, AirComp technology has emerged as an efficient solution for distributed data aggregation. The main idea behind AirComp is to utilize the natural superposition characteristics of wireless channels, allowing signals from different uplink terminals to be superimposed during wireless transmission to calculate global statistics. The receiver only needs to scale and restore the superimposed signal to directly obtain the calculation result, thus completely breaking through the latency bottleneck caused by traditional Orthogonal Multiple Access (OMA) technology.

[0003] However, due to the use of simulated uncoded concurrent transmission, airborne computing systems are extremely sensitive to wireless channel fading and external electromagnetic interference. Especially in complex wireless environments with multiple malicious interference sources, the precise amplitude alignment of the receiver signal is severely disrupted, leading to a sharp deterioration in the mean square error (MSE). Currently, to address this issue, the industry often employs traditional non-convex optimization methods (such as alternating optimization algorithms and block coordinate descent methods) to dynamically adjust the transmit power of each device to combat interference. However, when faced with massive node access, these traditional optimization methods typically involve extremely complex high-dimensional matrix operations and multiple iterative convergence processes, resulting in high computational complexity and significant latency. On the other hand, most existing pure data-driven deep learning solutions directly use raw channel information as a black-box input, lacking explicit mining and fusion of communication physical characteristics, leading to insufficient model generalization ability and a high likelihood of generating invalid solutions that do not conform to physical laws such as device transmission limitations. How to balance low latency, low complexity, and high anti-interference computational accuracy remains a technical challenge for existing airborne computing systems.

[0004] In summary, among the anti-interference power allocation methods adapted to existing airborne computing systems, traditional non-convex optimization schemes suffer from high computational complexity and large online inference latency, while pure data-driven deep learning schemes lack physical feature constraints and are prone to generating invalid solutions. The applicant has made corresponding explorations to address these issues. Summary of the Invention

[0005] The purpose of this application is to solve the above-mentioned problems by providing an anti-interference airborne computing power allocation method, corresponding apparatus, electronic equipment and computer-readable storage medium.

[0006] To achieve the various objectives of this application, the following technical solution is adopted:

[0007] An anti-jamming airborne computing power allocation method proposed for one of the purposes of this application includes:

[0008] The over-the-air computing system acquires raw channel parameters and interference parameters. The raw channel parameters include the direct channel gain from each edge device to the receiver, and the interference parameters include the interference channel gain from each malicious interference source to the receiver, as well as the additive white Gaussian noise power at the receiver.

[0009] Global statistical features are extracted from the direct channel gain, and the direct channel gains are sorted in descending order of amplitude, with the top-ranked direct channel gain values ​​used as sorting features. The total interference noise power is calculated based on the interference channel gain and the additive white Gaussian noise power, and used as an environmental interference feature. The global statistical features, the sorting features, and the environmental interference features are combined to construct a composite feature vector.

[0010] The composite feature vector is input into the air computing power allocation model that has been trained to convergence, and the high-dimensional feature nonlinear mapping is performed through the hidden layer of the air computing power allocation model to obtain the intermediate feature vector.

[0011] In the output layer of the air computing power allocation model, a preset hard saturation activation function is used to map the intermediate feature vector to a preset normalization interval. The normalized output value is scaled according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device.

[0012] The target transmit power is output so that each edge device can precode the target transmit power and its own channel phase, and then transmit data concurrently at the same time and in the same frequency band.

[0013] Optionally, the step of calculating the total interference noise power as an environmental interference characteristic quantity based on the interference channel gain and the additive white Gaussian noise power includes:

[0014] Calculate the squared modulus of the interference channel gain from each malicious interference source to the receiver, and calculate the first product between the squared modulus of the interference channel gain and the transmission variance of the corresponding interference signal to obtain the interference power component contributed by each malicious interference source.

[0015] The interference power components contributed by all malicious interference sources are summed to determine a first sum. The first sum is then added to the additive white Gaussian noise power to determine a second sum. The second sum is used as the total interference noise power to determine the environmental interference characteristic quantity.

[0016] Optionally, the step of mapping the intermediate feature vector to a preset normalization interval using a preset hard saturation activation function, and scaling the normalized output value according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device includes:

[0017] The intermediate feature vector is input to the output layer of the air computing power allocation model. After linear transformation by the output layer, an initial power feature vector corresponding to each edge device is output.

[0018] The initial power feature vector is hard-saturated by using a preset hard-saturation activation function, and each element value in the initial power feature vector is constrained to the normalization interval [0, 1] to generate a normalized power coefficient vector, wherein each coefficient in the normalized power coefficient vector corresponds one-to-one with each edge device;

[0019] Calculate the second product between each coefficient in the normalized power coefficient vector and the maximum transmit power limit of the edge device, and use the second product as the target transmit power of the corresponding edge device.

[0020] Optionally, the steps for training the in-flight computational power allocation model include:

[0021] Multiple historical samples are acquired under different channel conditions and interference distributions. The historical samples include historical channel parameters and historical interference parameters. The basic network architecture of the air computing power allocation model is a multilayer perceptron. The multilayer perceptron includes an input layer, multiple fully connected hidden layers, and an output layer. Each fully connected hidden layer is followed by a batch normalization layer, a ReLU activation layer, and a random deactivation layer.

[0022] Based on the historical samples, the alternating optimization algorithm is used to iteratively solve the mean square error minimization model for aerial computation to obtain the optimal power allocation scheme corresponding to each historical sample, and the optimal power allocation scheme is used as the training label.

[0023] The composite feature vectors corresponding to each of the historical samples are input into the air computing power allocation model to be trained, and the predicted power vector is output after forward propagation.

[0024] Using the mean square error between the predicted power vector and the training label as the loss value, the AdamW optimizer algorithm is used to update the network parameters of the air computing power allocation model according to the loss value until the loss value meets the preset convergence condition, thus obtaining the air computing power allocation model that has been trained to convergence.

[0025] Optionally, the step of inputting the composite feature vector into a trained and converged airborne computing power allocation model, and performing high-dimensional feature nonlinear mapping through the hidden layer of the airborne computing power allocation model to obtain an intermediate feature vector includes:

[0026] The composite feature vector is input into the air computing power allocation model via the input layer;

[0027] In the multiple fully connected hidden layers, each hidden layer performs a linear transformation on the input features of the previous level. After batch normalization and nonlinear activation by the ReLU activation function, the linear transformation results are randomly discarded by a random deactivation layer at a preset discard rate, and the processing results are passed to the next hidden layer to complete the nonlinear mapping of high-dimensional features layer by layer.

[0028] The feature vector output by the last fully connected hidden layer is used as the intermediate feature vector.

[0029] Optionally, the steps of constructing an aerial computation mean square error minimization model include:

[0030] Obtain the number of edge devices, the direct channel gain from each edge device to the receiver, the total interference noise power, and the maximum transmit power of each edge device;

[0031] Using the transmit power of each edge device and the denoising factor of the receiver as optimization variables, the number of edge devices, the direct channel gain, the total interference noise power, and the maximum transmit power as parameters, the first constraint is that the transmit power of each edge device is between zero and the maximum transmit power, and the second constraint is that the denoising factor is greater than zero. The optimization objective is to minimize the mean square error of the over-the-air computation.

[0032] Optionally, after the step of outputting the target transmit power so that each of the edge devices performs precoding based on the target transmit power and its own channel phase, and then concurrently transmits data at the same time and in the same frequency band, the method includes:

[0033] The receiver receives superimposed wireless signals transmitted concurrently by all edge devices on the same frequency band, as well as interference signals emitted by various malicious interference sources.

[0034] The third product between the target transmit power of each edge device and the square of the corresponding direct channel gain is calculated to obtain the weighted channel power value of each edge device.

[0035] The weighted channel power values ​​of all edge devices are summed and then added to the total interference noise power to obtain the first accumulated value;

[0036] The square root of the target transmit power of each edge device is calculated and the fourth product between the corresponding direct channel gain is used to obtain the amplitude alignment value of each edge device.

[0037] The amplitude alignment values ​​of all edge devices are summed to obtain the second accumulated value;

[0038] Calculate the square of the ratio of the first accumulated value to the second accumulated value, and use the squared value as the optimal denoising factor;

[0039] Divide the received superimposed signal by the optimal denoising factor to suppress interference signals and additive noise contained in the superimposed signal, and recover the global aggregation result of the local data of each edge device.

[0040] An anti-interference airborne computing power distribution device provided for another purpose of this application includes:

[0041] The data acquisition module is configured to acquire raw channel parameters and interference parameters from the over-the-air computing system. The raw channel parameters include the direct channel gain from each edge device to the receiver, and the interference parameters include the interference channel gain from each malicious interference source to the receiver, as well as the additive white Gaussian noise power at the receiver.

[0042] The composite feature construction module is configured to extract global statistical features from the direct channel gain, and after sorting the direct channel gains in descending order of amplitude, extract the top-ranked direct channel gain values ​​as sorting features; calculate the total interference noise power based on the interference channel gain and the additive white Gaussian noise power as environmental interference features; and combine the global statistical features, the sorting features, and the environmental interference features to construct a composite feature vector.

[0043] The intermediate feature output module is configured to input the composite feature vector into the air computing power allocation model that has been trained to convergence, and perform high-dimensional feature nonlinear mapping through the hidden layer of the air computing power allocation model to obtain the intermediate feature vector.

[0044] The target power determination module is set as the output layer of the power allocation model calculated in the air. It uses a preset hard saturation activation function to map the intermediate feature vector to a preset normalization interval, and scales the normalized output value according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device.

[0045] The data transmission module is configured to output the target transmit power so that each edge device can transmit data concurrently at the same time and in the same frequency band after precoding the target transmit power and its own channel phase.

[0046] An electronic device provided for another purpose of this application includes a central processing unit and a memory, the central processing unit being configured to invoke and run a computer program stored in the memory to perform the steps of the anti-interference over-the-air computing power allocation method described in this application.

[0047] A computer-readable storage medium is provided for another purpose of this application, which stores, in the form of computer-readable instructions, a computer program implemented according to the anti-jamming airborne computing power allocation method, which, when invoked by a computer, executes the steps included in the corresponding method.

[0048] Compared to existing technologies, this application addresses the problems in existing anti-interference power allocation methods for in-flight computing systems, such as the high computational complexity and large online inference latency of traditional non-convex optimization schemes, and the lack of physical feature constraints and the tendency to generate invalid solutions in pure data-driven deep learning schemes. This application provides, but is not limited to, the following beneficial effects:

[0049] Firstly, this application replaces the traditional approach of directly using raw channel information as a black-box input by constructing an enhanced feature vector that includes channel statistics and interference distribution. Specifically, this feature fusion strategy extracts global statistical features such as the mean, standard deviation, maximum, minimum, and sum of squares energy of the channel gain, representing the overall distribution of channel conditions; it extracts the top few strongest channel values ​​in descending order as ranking features, injecting the "prioritizing users with strong channels" principle into the neural network as a priori value; and it incorporates the total interference noise power as an environmental interference feature, enabling the model to perceive the overall interference intensity in the current time slot. The fusion of these multi-dimensional physical features allows the neural network to optimize within a subspace constrained by physical laws, rather than blindly searching in a large space of invalid solutions, significantly reducing the learning difficulty of the model and improving its generalization ability.

[0050] Secondly, this application innovatively deploys the Hardsigmoid activation function at the network output end. Utilizing its hard saturation characteristics at both ends, it forcibly fits the nonlinear physical boundary of "full-power transmission" or "physical truncation silence" for devices in anti-interference scenarios at the network structure level. In contrast, traditional smooth activation functions such as Sigmoid or Tanh can only output non-zero decimals that infinitely approximate the boundary values, failing to reach strict 0 or 1 boundaries. This results in devices with extremely poor channel conditions still transmitting at very low power, wasting power resources and polluting superimposed signals. This application fundamentally avoids such ineffective power allocation schemes through the hard-saturated activation function, ensuring that the decision result naturally conforms to the optimal strategy form for anti-interference communication.

[0051] Thirdly, this application adopts an "offline training, online deployment" framework design. In the offline training stage, an alternating optimization algorithm is used to generate high-precision training labels, and the optimal solution knowledge of the traditional optimization algorithm is efficiently transferred to the neural network weights through the mean square error loss function and the AdamW optimizer. In the online inference stage, only one forward propagation is required to complete the power allocation decision. Compared with the traditional iterative alternating optimization algorithm, the on-the-air computation time of this application hardly increases with the increase of the scale of edge devices, achieving an order of magnitude speedup (exhibiting an approximately O(1) time complexity advantage, that is, no matter how many edge devices are in the system, the online inference time of this application remains basically unchanged and always remains extremely short). Moreover, the mean square error calculation performance of the system is highly close to the optimal solution, successfully balancing the deployment requirements of optimal accuracy and ultra-low latency. Attached Figure Description

[0052] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0053] Figure 1 This is a flowchart illustrating the anti-interference over-the-air computing power allocation method in the embodiments of this application;

[0054] Figure 2 This is a graph showing the variation of the mean square error of different methods in the embodiments of this application with the number of devices;

[0055] Figure 3 This is a graph showing the variation of the average computation time of different methods under micro-benchmark testing in the embodiments of this application;

[0056] Figure 4 This is a schematic block diagram of the anti-interference airborne computing power distribution device in the embodiments of this application;

[0057] Figure 5 This is a schematic diagram of the structure of the computer device in the embodiments of this application. Detailed Implementation

[0058] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0059] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0060] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0061] Those skilled in the art will understand that the terms "client," "terminal," and "terminal device" as used herein include both devices that receive wireless signals, devices that only possess wireless signal receiver capabilities without transmission capabilities, and devices with receiving and transmitting hardware, devices that have receiving and transmitting hardware capable of bidirectional communication over a bidirectional communication link. Such devices may include: cellular or other communication devices such as personal computers or tablets, having single-line displays, multi-line displays, or cellular or other communication devices without multi-line displays; PCS (Personal Communications Service) that can combine voice, data processing, fax, and / or data communication capabilities; PDAs (Personal Digital Assistants) that may include radio frequency receivers, pagers, internet / intranet access, web browsers, notebooks, calendars, and / or GPS (Global Positioning System) receivers; and conventional laptops and / or handheld computers or other devices that have and / or include radio frequency receivers. As used herein, "client," "terminal," and "terminal device" can be portable, transportable, installed in a means of transportation (air, sea, and / or land), or suitable and / or configured to operate locally and / or in a distributed manner, operating in any other location on Earth and / or in space. "Client," "terminal," and "terminal device" as used herein can also be a communication terminal, an internet access terminal, or a music / video playback terminal, such as a PDA, a MID (Mobile Internet Device), and / or a mobile phone with music / video playback capabilities, or a smart TV, set-top box, etc.

[0062] The hardware referred to by the names "server," "client," and "service node" in this application is essentially an electronic device with the equivalent capabilities of a personal computer. It is a hardware device with the necessary components revealed by the von Neumann architecture, such as a central processing unit (including an arithmetic logic unit and a control unit), memory, input devices, and output devices. The computer program is stored in its memory, and the central processing unit loads the program stored in the secondary storage into the main memory to run it, executes the instructions in the program, and interacts with the input and output devices to complete specific functions.

[0063] It should be noted that the concept of "server" used in this application can also be extended to the case of server clusters. Based on the network deployment principles understood by those skilled in the art, the servers should be logically divided. Physically, these servers can be independent of each other but accessible through interfaces, or they can be integrated into a single physical computer or a computer cluster. Those skilled in the art should understand this flexibility and should not use it to constrain the implementation of the network deployment method in this application.

[0064] One or more of the technical features of this application, unless explicitly specified herein, can be deployed on a server and accessed by a client remotely calling the online service interface provided by the server, or can be directly deployed and run on a client for access.

[0065] Unless otherwise specified, the neural network models referenced or potentially referenced in this application may be deployed on a remote server and invoked remotely on the client, or deployed on a client with the capability to invoke directly. In some embodiments, when running on the client, the corresponding intelligence may be acquired through transfer learning in order to reduce the requirements on the client's hardware resources and avoid excessive consumption of the client's hardware resources.

[0066] Unless otherwise specified, all data involved in this application may be stored remotely on a server or on a local terminal device, as long as it is suitable for use by the technical solution of this application.

[0067] Those skilled in the art will understand that although the various methods in this application are described based on the same concept and thus present commonality among them, they can be performed independently unless otherwise specified. Similarly, the various embodiments disclosed in this application are all based on the same inventive concept; therefore, concepts expressed in the same way, as well as concepts that are appropriately changed for convenience but are expressed differently, should be understood equivalently.

[0068] Unless otherwise expressly stated, the various embodiments disclosed in this application can be combined in a cross-cutting manner to flexibly construct new embodiments, as long as such combination does not depart from the inventive spirit of this application and can meet the needs of the prior art or solve a certain deficiency in the prior art. Those skilled in the art should be aware of such modifications.

[0069] Based on the above exemplary scenarios, please refer to Figure 1 In one embodiment of the anti-jamming over-the-air computing power allocation method of this application, the method includes:

[0070] Step S10: The over-the-air computing system obtains the original channel parameters and interference parameters. The original channel parameters include the direct channel gain from each edge device to the receiver, and the interference parameters include the interference channel gain from each malicious interference source to the receiver, as well as the additive white Gaussian noise power at the receiver.

[0071] The over-the-air computing system in the terminal device can obtain raw channel parameters and interference parameters. The raw channel parameters include the direct channel gain from each edge device to the receiver, and the interference parameters include the interference channel gain from each malicious interference source to the receiver, as well as the additive white Gaussian noise power at the receiver.

[0072] It should be noted that the over-the-air computing system consists of K single-antenna edge devices, one single-antenna receiver, and N malicious interference sources. The over-the-air computing system first obtains the... Direct channel gain from edge device to receiver , No. The interference channel gain from a malicious interference source to the receiver and the additive white Gaussian noise power at the receiving end. .

[0073] In some embodiments, the direct channel gain represents the amplitude attenuation coefficient of the wireless transmission channel between the k-th edge device and the receiver, reflecting the amplitude change of the signal after it is emitted from the device and propagates through space to reach the receiver; the interference channel gain represents the amplitude attenuation coefficient of the interference channel between the n-th malicious interference source and the receiver, reflecting the amplitude change of the interference signal after it is emitted from the interference source and propagates through space to reach the receiver; the additive white Gaussian noise power is the variance of the additive white Gaussian noise, representing the inherent thermal noise power inside the receiver, which is independent of external signals and always exists in the receiver.

[0074] Step S20: Extract global statistical features from the direct channel gain, and after arranging the direct channel gains in descending order of amplitude, extract the top-ranked direct channel gain values ​​as sorting features; calculate the total interference noise power based on the interference channel gain and the additive white Gaussian noise power, as an environmental interference feature; combine the global statistical features, the sorting features, and the environmental interference features to construct a composite feature vector;

[0075] After the airborne computing system acquires the original channel parameters and interference parameters, it extracts global statistical features from the direct channel gain, and after arranging the direct channel gains in descending order of amplitude, it extracts the top-ranked direct channel gain values ​​as sorting features. It calculates the total interference noise power based on the interference channel gain and the additive white Gaussian noise power, which is used as environmental interference features. It combines the global statistical features, the sorting features, and the environmental interference features to construct a composite feature vector.

[0076] Specifically, instead of directly using the original channel gain as model input, data cleaning and feature enhancement are performed in conjunction with communication physical characteristics: First, the direct channel gain from each edge device to the receiver is normalized, and global statistical features of the channel gain of all edge devices are extracted. These global statistical features include the mean, standard deviation, maximum, minimum, and sum of squares of the channel gain of all edge devices. Second, the direct channel gains of all edge devices are sorted in descending order of amplitude, and the top M strongest channel gain values ​​are selected as sorting features to characterize the logic of prioritizing strong channel users in the over-the-air computing system. Finally, the total interference noise power is calculated based on the interference channel gain and additive white Gaussian noise power from each malicious interference source to the receiver, serving as an environmental interference feature to characterize the current interference intensity experienced by the over-the-air computing system. The global statistical features, the sorting features, and the environmental interference features are combined to construct a composite feature vector. .

[0077] In some embodiments, the step of calculating the total interference noise power as an environmental interference characteristic based on the interference channel gain and the additive white Gaussian noise power includes:

[0078] Step S201: Calculate the squared modulus of the interference channel gain from each malicious interference source to the receiver, and calculate the first product between the squared modulus of the interference channel gain and the transmission variance of the corresponding interference signal to obtain the interference power component contributed by each malicious interference source.

[0079] For the There are several malicious interference sources, and the interference channel gain to the receiver is... First, regarding the first The squared modulo of the interference channel gain from a malicious interference source to the receiver. The power gain values ​​of each interference channel are obtained. Then, the power gain values ​​of each interference channel are compared with the signal transmission variance of the corresponding interference source. Multiplying these components yields the interference power components contributed by each malicious interference source. .

[0080] Step S202: Summate the interference power components contributed by all malicious interference sources to determine a first sum value, calculate and add the first sum value to the additive white Gaussian noise power to determine a second sum value, and use the second sum value as the total interference noise power to determine the environmental interference characteristic quantity.

[0081] Interference power components contributed by all malicious interference sources Perform cumulative summation to determine the first sum value. Calculate and determine the first sum and the additive white Gaussian noise power. Add them to determine the second sum. , the second sum value As the total interference noise power In order to determine the characteristic quantities of environmental disturbance.

[0082] More specifically, the formula for calculating the total interference noise power is expressed as follows:

[0083]

[0084] in, Indicates the total interference noise power; This represents the power of additive white Gaussian noise; Indicates the first The interference channel gain from a malicious interference source to the receiver; This represents the transmission variance of the interference signal; Indicates the first One source of malicious interference; This indicates the total number of malicious interference sources.

[0085] Step S30: Input the composite feature vector into the air computing power allocation model that has been trained to convergence, and perform high-dimensional feature nonlinear mapping through the hidden layer of the air computing power allocation model to obtain the intermediate feature vector;

[0086] Global statistical features are extracted from the direct channel gain, and the direct channel gains are sorted in descending order of amplitude, with the top-ranked direct channel gain values ​​used as sorting features. The total interference noise power is calculated based on the interference channel gain and the additive white Gaussian noise power, serving as an environmental interference feature. The global statistical features, the sorting features, and the environmental interference features are combined to construct a composite feature vector, which is then input into a converged airborne computing power allocation model. The model undergoes high-dimensional feature nonlinear mapping through its hidden layer to obtain an intermediate feature vector.

[0087] In some embodiments, the step of training an in-flight computational power allocation model includes:

[0088] Step S1100: Obtain multiple historical samples under different channel conditions and interference distributions, wherein the historical samples include historical channel parameters and historical interference parameters. The basic network architecture of the air computing power allocation model is a multilayer perceptron, wherein the multilayer perceptron includes an input layer, multiple fully connected hidden layers and an output layer; each fully connected hidden layer is followed by a batch normalization layer, a ReLU activation layer and a random deactivation layer in sequence.

[0089] The multilayer perceptron includes an input layer, four hidden layers, and an output layer. The neuron dimensions of the hidden layers are configured as 512, 256, 128, and 64, respectively. Each hidden layer is non-linearly activated by the ReLU activation function after matrix multiplication, and the data is normalized by batch normalization (BatchNorm1D). It is also connected to a random deactivation layer with a dropout rate of 0.05 to prevent the model from overfitting.

[0090] Step S1200: Based on each historical sample, the alternating optimization algorithm is used to iteratively solve the mean square error minimization model for aerial computation to obtain the optimal power allocation scheme corresponding to each historical sample, and the optimal power allocation scheme is used as the training label.

[0091] Step S1300: Input the composite feature vector corresponding to each of the historical samples into the air computing power allocation model to be trained, and output the predicted power vector after forward propagation;

[0092] Step S1400: Using the mean square error between the predicted power vector and the training label as the loss value, the AdamW optimizer algorithm is used to update the network parameters of the air computing power allocation model according to the loss value until the loss value meets the preset convergence condition, thus obtaining the air computing power allocation model that has been trained to convergence.

[0093] In a further embodiment, the step of constructing an airborne computation mean square error minimization model includes:

[0094] Step S301: Obtain the number of edge devices, the direct channel gain from each edge device to the receiver, the total interference noise power, and the maximum transmit power of each edge device;

[0095] Step S302: Using the transmit power of each edge device and the denoising factor of the receiver as optimization variables, the number of edge devices, the direct channel gain, the total interference noise power and the maximum transmit power as parameters, the first constraint is that the transmit power of each edge device is between zero and the maximum transmit power, and the second constraint is that the denoising factor is greater than zero, and the optimization objective is to minimize the mean square error of the air computation.

[0096] Specifically, the calculation formula for the mean square error minimization model in aerial computation is expressed as follows:

[0097]

[0098] The first constraint condition is expressed as follows:

[0099]

[0100] The second constraint is expressed as follows:

[0101]

[0102] Where Q is the mean square error of the calculation in the air; K is the total number of edge devices; Let k be the transmit power of the k-th edge device; The direct channel gain from the k-th edge device to the receiver is estimated. denoised by the noise reduction factor at the receiver; C is the total noise plus interference power; This represents the maximum transmit power of the edge device.

[0103] In some embodiments, during the offline training phase, firstly, multiple historical samples under different channel conditions and interference distributions are acquired, each historical sample including historical channel parameters and historical interference parameters; then, for each historical sample, an alternating optimization algorithm based on block coordinate descent is used to iteratively solve the mean square error minimization model for airborne computation to obtain the optimal power allocation scheme corresponding to each historical sample, and the optimal power allocation scheme is used as high-precision training label data. Next, the composite feature vectors corresponding to each historical sample are input into the air computing power allocation model to be trained, and the predicted power vector is output after forward propagation. Finally, the mean square error between the predicted power vector and the training label data is used as the loss value. The AdamW optimizer algorithm is used to update the network parameters of the air computing power allocation model according to the loss value, and an early stopping strategy is used to determine whether the training has converged. Training is stopped when the loss value meets the preset convergence condition, and the trained air computing power allocation model is obtained.

[0104] The expression for the mean square error between the predicted power vector and the training label data is:

[0105]

[0106] During the online inference phase, the composite feature vectors acquired in real time are input into the trained air computing power allocation model. After one forward propagation calculation, the target transmit power of each edge device is directly output without performing any iterative optimization operations.

[0107] In a further embodiment, the step of inputting the composite feature vector into a trained and converged airborne computing power allocation model, and performing high-dimensional feature nonlinear mapping through the hidden layer of the airborne computing power allocation model to obtain an intermediate feature vector includes:

[0108] Step S3001: Input the composite feature vector into the air computing power allocation model via the input layer;

[0109] Step S3002: In the multiple fully connected hidden layers, each hidden layer performs a linear transformation on the input features of the previous level. After batch normalization and nonlinear activation by the ReLU activation function, the linear transformation results are randomly discarded by a random deactivation layer at a preset discard rate, and the processing results are passed to the next hidden layer to complete the nonlinear mapping of high-dimensional features layer by layer.

[0110] The preset discard rate can be 0.05, etc. Those skilled in the art can determine the preset discard rate as needed according to the actual application scenario, and no limitation is made here.

[0111] Step S3003: Use the feature vector output by the last fully connected hidden layer as the intermediate feature vector.

[0112] Step S40: In the output layer of the air computing power allocation model, a preset hard saturation activation function is used to map the intermediate feature vector to a preset normalization interval, and the normalized output value is scaled according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device.

[0113] The composite feature vector is input into the air computing power allocation model that has been trained to convergence. After high-dimensional feature nonlinear mapping is performed in the hidden layer of the air computing power allocation model to obtain the intermediate feature vector, the intermediate feature vector is mapped to the preset normalization interval by a preset hard saturation activation function in the output layer of the air computing power allocation model. The normalized output value is scaled according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device.

[0114] The steps of mapping the intermediate feature vector to a preset normalization interval using a preset hard saturation activation function, and scaling the normalized output value according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device include:

[0115] Step S401: Input the intermediate feature vector into the output layer of the air computing power allocation model, and output the initial power feature vector corresponding to each edge device after linear transformation of the output layer;

[0116] The intermediate feature vector is a high-dimensional feature representation output from the last fully connected hidden layer of the over-the-air computing power allocation model. This intermediate feature vector is input to the output layer of the over-the-air computing power allocation model. After a linear transformation (i.e., matrix multiplication and bias addition) of the output layer, an initial power feature vector corresponding one-to-one with each edge device is output. The values ​​of each element in the initial power feature vector... This is the raw linear output value before activation function processing, and its value range is any real number.

[0117] Step S402: Use a preset hard saturation activation function to perform hard saturation mapping on the initial power feature vector, constrain the value of each element in the initial power feature vector to the normalization interval [0, 1], and generate a normalized power coefficient vector, wherein each coefficient in the normalized power coefficient vector corresponds one-to-one with each edge device;

[0118] The target transmit power is determined by physical boundary constraint mapping of intermediate dimension features using a preset hard saturation activation function. The hard saturation activation function can be a Hardsigmoid activation function, etc. A preset Hardsigmoid activation function is used to apply physical boundary constraints to the values ​​of each element in the initial power feature vector. Perform hard saturation mapping to constrain the values ​​of each element to the normalized interval [0, 1].

[0119] It should be noted that traditional neural network output layers often use smooth activation functions such as Sigmoid or Tanh, which are difficult to output absolute extreme values. However, in real-world anti-interference communication games, to maximize overall computational accuracy and suppress interference, the system needs to force devices with poor channel conditions to "silence" (power strictly zero) or allow devices with normal channel conditions to "transmit at full power" (power strictly zero). ).

[0120] In anti-interference communication game theory, the optimal power allocation strategy often presents an extreme "all or nothing" form: devices with excellent channel quality should transmit at maximum power to suppress interference, while devices with extremely poor channel quality should completely shut down transmission to avoid contaminating superimposed signals. Traditional smooth activation functions such as Sigmoid or Tanh cannot output precise 0 or 1 boundary values, causing the model output to naturally deviate from the optimal strategy form. This application, through the hard saturation characteristic of the Hardsigmoid activation function, enforces this physical boundary constraint at the network structure level, fundamentally eliminating the generation of such invalid power allocation schemes.

[0121] Step S403: Calculate the second product between each coefficient in the normalized power coefficient vector and the maximum transmit power limit value of the edge device, and use the second product as the target transmit power of the corresponding edge device.

[0122] The coefficients in the normalized power coefficient vector generated in step S402 are... Multiply by the maximum transmit power limit of the edge device respectively That is, to calculate each coefficient in the normalized power coefficient vector. Maximum transmit power limit of edge devices The second product between them is used as the target transmit power of the corresponding edge device, which is expressed as: This allows us to obtain the accurate K-dimensional power allocation decision vector. , which correspond to the target transmit power of K edge devices respectively.

[0123] Step S50: Output the target transmit power so that each edge device can transmit data concurrently at the same time and in the same frequency band after precoding according to the target transmit power and its own channel phase.

[0124] In the output layer of the air computing power allocation model, a preset hard saturation activation function is used to map the intermediate feature vector to a preset normalization interval. The normalized output value is scaled according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device. Then, the target transmit power is output so that each edge device can transmit data concurrently at the same time and in the same frequency band after precoding according to the target transmit power and its own channel phase.

[0125] Specifically, after each edge device acquires its corresponding target transmit power, it uses a precoding strategy to pre-compensate the phase of its local data to be transmitted based on the channel phase information from its own transmission to the receiver. This compensates for the phase rotation caused by the wireless channel, ensuring that the signals from each edge device are phase-aligned when they reach the receiver. After precoding, each edge device concurrently transmits its baseband signal at the same time and on the same frequency band.

[0126] In some embodiments, after the step of outputting the target transmit power so that each of the edge devices performs precoding based on the target transmit power and its own channel phase, and then concurrently transmits data at the same time and in the same frequency band, the method includes:

[0127] Step S501: The receiving end receives the superimposed wireless signals transmitted concurrently by all edge devices in the same frequency band, as well as the interference signals emitted by each malicious interference source.

[0128] Each edge device precodes its data according to its target transmit power and channel phase, and then transmits data concurrently on the same frequency band at the same time. The receiver receives the superimposed wireless signals from all the edge devices through its antenna. At the same time, the superimposed signal is also mixed with the interference signals emitted by various malicious interference sources, forming a complete received signal.

[0129] Step S502: Calculate the third product between the target transmit power of each edge device and the square of the corresponding direct channel gain to obtain the weighted channel power value of each edge device;

[0130] For the k-th edge device, its target transmit power is... With corresponding direct channel gain Multiplying the squares of the values ​​yields the weighted channel power value when the signal from the k-th edge device arrives at the receiver. For all Each edge device performs the above calculation to obtain the weighted channel power value contributed by each edge device.

[0131] Step S503: Sum the weighted channel power values ​​of all edge devices and add them to the total interference noise power to obtain the first accumulated value;

[0132] Step S504: Calculate the fourth product between the square root of the target transmit power of each edge device and the corresponding direct channel gain to obtain the amplitude alignment value of each edge device;

[0133] Step S505: Sum the amplitude alignment values ​​of all edge devices to obtain the second accumulated value;

[0134] Step S506: Calculate the square of the ratio of the first accumulated value to the second accumulated value, and use the squared value as the optimal denoising factor;

[0135] Step S507: Divide the received superimposed signal by the optimal denoising factor to suppress the interference signal and additive noise contained in the superimposed signal, and recover the global aggregation result of the local data of each edge device.

[0136] The superimposed signal is a synthesized radio frequency signal received by the receiver, which is composed of useful signals transmitted by each edge device, interference signals transmitted by each malicious interference source, and additive noise from the receiver. The amplitude of the useful signal is determined by the transmit power of each edge device and the corresponding direct channel gain, while the amplitude of the interference signal is determined by the interference channel gain from each malicious interference source to the receiver.

[0137] Specifically, the formula for calculating the optimal denoising factor is expressed as follows:

[0138]

[0139] in, A denoising factor that achieves the optimal trade-off between signal amplitude alignment and interference suppression under the current power allocation; Let the target transmit power be that of the k-th edge device; For direct channel gain; The total noise plus interference power, where k represents the k-th edge device. This indicates the total number of edge devices.

[0140] In some embodiments, Figure 2 The curves showing the variation of the mean squared error (MSE) of different methods with the number of devices are presented. Experimental results show that the deep learning-based power allocation method proposed in this application is close to the traditional optimization method in terms of MSE calculation performance, and the gap between the two remains at a low level. In contrast, although the random forest regression method outperforms the maximum power allocation method, it falls short in high-dimensional input (large... Under certain conditions, it is difficult to capture accurate power allocation strategies, resulting in high mean square errors in calculations. The maximum power allocation method has the worst performance because it ignores channel differences and interference effects.

[0141] In some embodiments, Figure 3 The average computation time of different methods under micro-benchmark tests is shown. Traditional optimization methods, due to complex iterative optimization, have computation times that increase significantly to tens of milliseconds with K. In contrast, the anti-interference aerial computing power allocation method proposed in this application benefits from the non-iterative nature of neural network forward propagation and highly optimized tensor operations, resulting in extremely short inference time and minimal sensitivity to changes in the value of K. Compared with traditional methods, the proposed scheme not only achieves a speedup of tens of times but also demonstrates an advantage of approximately O(1) time complexity.

[0142] As can be seen from the above embodiments, compared with the prior art, this application addresses the problems of high computational complexity and large online inference latency in the traditional non-convex optimization schemes of existing airborne computing system anti-interference power allocation methods, while pure data-driven deep learning schemes lack physical feature constraints and are prone to generating invalid solutions. This application has, but is not limited to, the following beneficial effects:

[0143] Firstly, this application replaces the traditional approach of directly using raw channel information as a black-box input by constructing an enhanced feature vector that includes channel statistics and interference distribution. Specifically, this feature fusion strategy extracts global statistical features such as the mean, standard deviation, maximum, minimum, and sum of squares energy of the channel gain, representing the overall distribution of channel conditions; it extracts the top few strongest channel values ​​in descending order as ranking features, injecting the "prioritizing users with strong channels" principle into the neural network as a priori value; and it incorporates the total interference noise power as an environmental interference feature, enabling the model to perceive the overall interference intensity in the current time slot. The fusion of these multi-dimensional physical features allows the neural network to optimize within a subspace constrained by physical laws, rather than blindly searching in a large space of invalid solutions, significantly reducing the learning difficulty of the model and improving its generalization ability.

[0144] Secondly, this application innovatively deploys the Hardsigmoid activation function at the network output end. Utilizing its hard saturation characteristics at both ends, it forcibly fits the nonlinear physical boundary of "full-power transmission" or "physical truncation silence" for devices in anti-interference scenarios at the network structure level. In contrast, traditional smooth activation functions such as Sigmoid or Tanh can only output non-zero decimals that infinitely approximate the boundary values, failing to reach strict 0 or 1 boundaries. This results in devices with extremely poor channel conditions still transmitting at very low power, wasting power resources and polluting superimposed signals. This application fundamentally avoids such ineffective power allocation schemes through the hard-saturated activation function, ensuring that the decision result naturally conforms to the optimal strategy form for anti-interference communication.

[0145] Thirdly, this application adopts an "offline training, online deployment" framework design. In the offline training stage, an alternating optimization algorithm is used to generate high-precision training labels, and the optimal solution knowledge of the traditional optimization algorithm is efficiently transferred to the neural network weights through the mean square error loss function and the AdamW optimizer. In the online inference stage, only one forward propagation is required to complete the power allocation decision. Compared with the traditional iterative alternating optimization algorithm, the on-the-air computation time of this application hardly increases with the increase of the scale of edge devices, achieving an order of magnitude speedup (exhibiting an approximately O(1) time complexity advantage, that is, no matter how many edge devices are in the system, the online inference time of this application remains basically unchanged and always remains extremely short). Moreover, the mean square error calculation performance of the system is highly close to the optimal solution, successfully balancing the deployment requirements of optimal accuracy and ultra-low latency.

[0146] Please see Figure 4An anti-interference over-the-air computing power allocation device, provided to meet one of the purposes of this application, includes a data acquisition module 1100, a composite feature construction module 1200, an intermediate feature output module 1300, a target power determination module 1400, and a data transmission module 1500. The data acquisition module 1100 is configured to acquire raw channel parameters and interference parameters for the over-the-air computing system. The raw channel parameters include the direct channel gain from each edge device to the receiver, and the interference parameters include the interference channel gain from each malicious interference source to the receiver, and the additive white Gaussian noise power at the receiver. The composite feature construction module 1200 is configured to extract global statistical features from the direct channel gains, and after arranging the direct channel gains in descending order of amplitude, extract the top-ranked direct channel gain values ​​as sorting features; calculate the total interference noise power based on the interference channel gains and the additive white Gaussian noise power as environmental interference features; and combine the global statistical features, the sorting features, and the environmental interference features to construct a composite feature vector. The intermediate feature output module 1300 is configured to input the composite feature vector into a trained and converged air computing power allocation model, and perform high-dimensional feature nonlinear mapping through the hidden layer of the air computing power allocation model to obtain the intermediate feature vector; the target power determination module 1400 is configured to map the intermediate feature vector to a preset normalization interval using a preset hard saturation activation function in the output layer of the air computing power allocation model, and scale the normalized output value according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device; the data transmission module 1500 is configured to output the target transmit power so that each edge device can transmit data concurrently at the same time and in the same frequency band after precoding according to the target transmit power and its own channel phase.

[0147] Based on any embodiment of this application, please refer to Figure 5 Another embodiment of this application also provides an electronic device, which can be implemented by a computer device, such as... Figure 5The diagram shows the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable storage medium stores an operating system, a database, and computer-readable instructions. The database may store control information sequences. When the computer-readable instructions are executed by the processor, they enable the processor to implement an anti-interference over-the-air computing power allocation method. The processor of the computer device provides computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When these computer-readable instructions are executed by the processor, they enable the processor to execute the anti-interference over-the-air computing power allocation method of this application. The network interface of the computer device is used for communication with a terminal. Those skilled in the art will understand that… Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0148] In this embodiment, the processor is used to execute... Figure 4 The specific functions of each module are defined within the device, and the memory stores the program code and various data required to execute these modules. The network interface is used for data transmission between the user terminal and the server. In this embodiment, the memory stores the program code and data required to execute all modules in the anti-interference over-the-air computing power distribution device of this application, and the server can call the server's program code and data to execute the functions of all modules.

[0149] This application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the anti-interference over-the-air computing power allocation method described in any embodiment of this application.

[0150] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the anti-interference over-the-air computing power allocation method described in any embodiment of this application.

[0151] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0152] The above description is only a partial embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for anti-interference aerial computing power allocation, characterized in that, include: The over-the-air computing system acquires raw channel parameters and interference parameters. The raw channel parameters include the direct channel gain from each edge device to the receiver, and the interference parameters include the interference channel gain from each malicious interference source to the receiver, as well as the additive white Gaussian noise power at the receiver. Global statistical features are extracted from the direct channel gain, and the direct channel gains are sorted in descending order of amplitude, with the top-ranked direct channel gain values ​​used as sorting features. The total interference noise power is calculated based on the interference channel gain and the additive white Gaussian noise power, and used as an environmental interference feature. The global statistical features, the sorting features, and the environmental interference features are combined to construct a composite feature vector. The composite feature vector is input into the air computing power allocation model that has been trained to convergence, and the high-dimensional feature nonlinear mapping is performed through the hidden layer of the air computing power allocation model to obtain the intermediate feature vector. In the output layer of the air computing power allocation model, a preset hard saturation activation function is used to map the intermediate feature vector to a preset normalization interval. The normalized output value is scaled according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device. The target transmit power is output so that each edge device can precode the target transmit power and its own channel phase, and then transmit data concurrently at the same time and in the same frequency band.

2. The anti-interference airborne computing power allocation method according to claim 1, characterized in that, The step of calculating the total interference noise power based on the interference channel gain and the additive white Gaussian noise power, as an environmental interference characteristic quantity, includes: Calculate the squared modulus of the interference channel gain from each malicious interference source to the receiver, and calculate the first product between the squared modulus of the interference channel gain and the transmission variance of the corresponding interference signal to obtain the interference power component contributed by each malicious interference source. The interference power components contributed by all malicious interference sources are summed to determine a first sum. The first sum is then added to the additive white Gaussian noise power to determine a second sum. The second sum is used as the total interference noise power to determine the environmental interference characteristic quantity.

3. The anti-interference airborne computing power allocation method according to claim 1, characterized in that, The steps of mapping the intermediate feature vector to a preset normalization interval using a preset hard saturation activation function, and scaling the normalized output value according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device include: The intermediate feature vector is input to the output layer of the air computing power allocation model. After linear transformation by the output layer, an initial power feature vector corresponding to each edge device is output. The initial power feature vector is hard-saturated by using a preset hard-saturation activation function, and each element value in the initial power feature vector is constrained to the normalization interval [0, 1] to generate a normalized power coefficient vector, wherein each coefficient in the normalized power coefficient vector corresponds one-to-one with each edge device; Calculate the second product between each coefficient in the normalized power coefficient vector and the maximum transmit power limit of the edge device, and use the second product as the target transmit power of the corresponding edge device.

4. The anti-interference airborne computing power allocation method according to claim 1, characterized in that, The steps for training an in-flight computational power allocation model include: Multiple historical samples are acquired under different channel conditions and interference distributions. The historical samples include historical channel parameters and historical interference parameters. The basic network architecture of the air computing power allocation model is a multilayer perceptron. The multilayer perceptron includes an input layer, multiple fully connected hidden layers, and an output layer. Each fully connected hidden layer is followed by a batch normalization layer, a ReLU activation layer, and a random deactivation layer. Based on the historical samples, the alternating optimization algorithm is used to iteratively solve the mean square error minimization model for aerial computation to obtain the optimal power allocation scheme corresponding to each historical sample, and the optimal power allocation scheme is used as the training label. The composite feature vectors corresponding to each of the historical samples are input into the air computing power allocation model to be trained, and the predicted power vector is output after forward propagation. Using the mean square error between the predicted power vector and the training label as the loss value, the AdamW optimizer algorithm is used to update the network parameters of the air computing power allocation model according to the loss value until the loss value meets the preset convergence condition, thus obtaining the air computing power allocation model that has been trained to convergence.

5. The anti-interference airborne computing power allocation method according to claim 4, characterized in that, The steps of inputting the composite feature vector into a trained and converged airborne computing power allocation model, and performing high-dimensional feature nonlinear mapping through the hidden layer of the airborne computing power allocation model to obtain intermediate feature vectors include: The composite feature vector is input into the air computing power allocation model via the input layer; In the multiple fully connected hidden layers, each hidden layer performs a linear transformation on the input features of the previous level. After batch normalization and nonlinear activation by the ReLU activation function, the linear transformation results are randomly discarded by a random deactivation layer at a preset discard rate, and the processing results are passed to the next hidden layer to complete the nonlinear mapping of high-dimensional features layer by layer. The feature vector output by the last fully connected hidden layer is used as the intermediate feature vector.

6. The anti-interference airborne computing power allocation method according to claim 4, characterized in that, The steps for constructing an aerial computation mean square error minimization model include: Obtain the number of edge devices, the direct channel gain from each edge device to the receiver, the total interference noise power, and the maximum transmit power of each edge device; Using the transmit power of each edge device and the denoising factor of the receiver as optimization variables, the number of edge devices, the direct channel gain, the total interference noise power, and the maximum transmit power as parameters, the first constraint is that the transmit power of each edge device is between zero and the maximum transmit power, and the second constraint is that the denoising factor is greater than zero. The optimization objective is to minimize the mean square error of the over-the-air computation.

7. The anti-interference airborne computing power allocation method according to claim 1, characterized in that, After the step of outputting the target transmit power so that each edge device performs precoding based on the target transmit power and its own channel phase, and then transmits data concurrently at the same time and in the same frequency band, the following steps are included: The receiver receives superimposed wireless signals transmitted concurrently by all edge devices on the same frequency band, as well as interference signals emitted by various malicious interference sources. The third product between the target transmit power of each edge device and the square of the corresponding direct channel gain is calculated to obtain the weighted channel power value of each edge device. The weighted channel power values ​​of all edge devices are summed and then added to the total interference noise power to obtain the first accumulated value; The square root of the target transmit power of each edge device is calculated and the fourth product between the corresponding direct channel gain is used to obtain the amplitude alignment value of each edge device. The amplitude alignment values ​​of all edge devices are summed to obtain the second accumulated value; Calculate the square of the ratio of the first accumulated value to the second accumulated value, and use the squared value as the optimal denoising factor; Divide the received superimposed signal by the optimal denoising factor to suppress interference signals and additive noise contained in the superimposed signal, and recover the global aggregation result of the local data of each edge device.

8. An anti-interference airborne computing power distribution device, characterized in that, include: The data acquisition module is configured to acquire raw channel parameters and interference parameters from the over-the-air computing system. The raw channel parameters include the direct channel gain from each edge device to the receiver, and the interference parameters include the interference channel gain from each malicious interference source to the receiver, as well as the additive white Gaussian noise power at the receiver. The composite feature construction module is configured to extract global statistical features from the direct channel gain, and after sorting the direct channel gains in descending order of amplitude, extract the top-ranked direct channel gain values ​​as sorting features; calculate the total interference noise power based on the interference channel gain and the additive white Gaussian noise power as environmental interference features; and combine the global statistical features, the sorting features, and the environmental interference features to construct a composite feature vector. The intermediate feature output module is configured to input the composite feature vector into the air computing power allocation model that has been trained to convergence, and perform high-dimensional feature nonlinear mapping through the hidden layer of the air computing power allocation model to obtain the intermediate feature vector. The target power determination module is set as the output layer of the power allocation model calculated in the air. It uses a preset hard saturation activation function to map the intermediate feature vector to a preset normalization interval, and scales the normalized output value according to the maximum transmit power of the edge device to obtain the target transmit power of each edge device. The data transmission module is configured to output the target transmit power so that each edge device can transmit data concurrently at the same time and in the same frequency band after precoding the target transmit power and its own channel phase.

9. An electronic device comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images