Semantic relay communication method, relay device, system

Abstract

The application discloses a semantic relay communication method, a relay device and a system, which are executed by the relay device and include the following steps: collecting target data, outputting a confidence degree of the target data, and comparing the confidence degree with a threshold; when the confidence degree is not less than the threshold, generating a first feature signal through a local feature extractor, and sending the first feature signal to an edge server; when the confidence degree is less than the threshold, sending a target information acquisition request to an edge device, so that the edge device sends a second feature signal; receiving the second feature signal, fusing the first feature signal and the second feature signal to generate a fusion signal, and transmitting the fusion signal to the edge server. The application realizes dynamic trade-off between a signal source quality and communication resources, reduces redundant transmission, and improves system performance.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to semantic relay communication methods, relay devices, and systems. Background Technology

[0002] With the rapid development of the Internet of Things (IoT) and edge computing, the massive amounts of data generated by terminal devices require processing through efficient communication and computing architectures. Edge-device joint inference technology has become a common solution for scenarios with intensive communication data and high timeliness requirements. In edge-device joint inference systems, the AI ​​model based on deep neural networks is typically divided into two parts: one part is assigned to the edge device, and the other part is assigned to the edge server. The edge device uses its internal model to extract semantic features from the local raw data. The semantic features are transmitted to the edge server via a wireless link, where they are further processed to generate the final inference result.

[0003] Compared to uploading large amounts of raw data to edge servers, edge-device joint inference transmits semantic features with reduced dimensionality, thus significantly reducing inference latency in bandwidth-constrained scenarios. In edge-device joint inference systems, efficient feature transmission is crucial on fluctuating wireless channels. As an important means of improving communication system performance, relay technology, by establishing auxiliary transmission links between source and target nodes, can effectively extend communication coverage, improve transmission reliability, and reduce system energy consumption.

[0004] In scenarios such as intelligent transportation, industrial IoT, and smart cities, collaborative communication between edge devices (such as cameras and sensors) and relay devices has become a key technology for achieving real-time decision-making. Traditional relay technologies mainly focus on signal-level forwarding and processing, optimizing transmission performance through physical layer coding and power control, but fail to fully consider the optimization needs at the semantic level. Relay collaboration modes suffer from problems such as wasted communication resources, excessive latency, and high energy consumption. These issues severely restrict the overall system efficiency, especially in high-density device deployment scenarios.

[0005] Furthermore, simply embedding semantic communication modules into existing relay communication frameworks fails to fully consider the special role of relay nodes in semantic feature extraction and forwarding, and lacks a joint optimization mechanism for semantic information and channel state, thus exhibiting significant shortcomings. Therefore, research on semantic relay requires a systematic redesign to optimize performance. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the present invention provides a semantic relay communication method, device and medium to realize the dynamic balance between source quality and communication resources, reduce redundant transmission and improve system performance.

[0007] To address the aforementioned technical problems, the first aspect of this invention discloses a semantic relay communication method, executed by a relay device, comprising:

[0008] Collect target data, output the confidence level of the target data, and compare the confidence level with a threshold;

[0009] When the confidence level is not less than the threshold, a first feature signal is generated by the local feature extractor and sent to the edge server.

[0010] When the confidence level is less than the threshold, a target information acquisition request is sent to the edge device, and the target information acquisition request causes the edge device to emit a second feature signal.

[0011] The second feature signal is received, and the first feature signal and the second feature signal are fused to generate a fused signal, which is then transmitted to the edge server.

[0012] In some implementations, the confidence level is calculated in the following manner:

[0013] The first feature signal is input into the confidence judgment module based on a deep neural network, and the class probability distribution is output through the Softmax function, with the maximum probability value being used as the confidence level.

[0014] In some implementations, the threshold is a dynamically adjustable parameter; the threshold is determined by a traversal method based on changes in the source data format and the channel environment.

[0015] In some implementations, the threshold decreases as the task accuracy of the edge server increases, and the threshold decreases as the real-time signal-to-noise ratio decreases.

[0016] In some implementations, fusing the first feature signal and the second feature signal to generate a fused signal includes:

[0017] The first and second feature signals are weighted and denoised, and the feature dimensions are compressed through a fully connected layer to generate a fused signal.

[0018] In some implementations, the dimension of the fused signal is not greater than the dimension of the first feature signal.

[0019] Secondly, a relay device is disclosed that performs the semantic relay communication method as described above; the relay device includes a confidence judgment module, a feature extractor, and a fusion module;

[0020] The feature extractor is used to generate a first feature signal from the target data;

[0021] The confidence level determination module is used to output a confidence level and compare the confidence level with a threshold; when the confidence level is not less than the threshold, it sends the first feature signal to the edge server; when the confidence level is less than the threshold, it sends a target information acquisition request to the edge device, and the target information acquisition request causes the edge device to emit a second feature signal.

[0022] The fusion module is used to generate a fused signal based on the first feature signal and the second feature signal, and transmit it to the edge server.

[0023] In some implementations, the confidence judgment module, feature extractor, and fusion module employ deep neural networks, and the loss function for training the deep neural network is:

[0024]

[0025] Among them, l k It is the label of the k-th category corresponding to the input data s; It is the prediction result The k-th element in.

[0026] Thirdly, a semantic relay communication system is disclosed, comprising:

[0027] The relay device as described above;

[0028] When an edge device receives a request to acquire target information, it collects the target information and sends out a second characteristic signal.

[0029] The edge server receives the fusion signal.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0031] This invention provides a semantic relay communication method, relay device, and system. The relay device intelligently selects which relay device to directly transmit or requests supplementary feature signals from edge devices based on the output confidence level, effectively reducing redundant data transmission. While maintaining task performance similar to traditional multi-stage transmission, it reduces transmission latency and energy consumption. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the structure of a semantic communication system provided by the present invention.

[0033] Figure 2 This invention provides a timing diagram for the transmission of a relay scheduling strategy in a semantic communication system.

[0034] Figure 3 A flowchart of a transmission scheduling method provided by the present invention;

[0035] Figure 4 This is a schematic diagram of the structure of a feature extractor in a semantic communication system provided by the present invention;

[0036] Figure 5 This is a schematic diagram of the task processing model structure in a semantic communication system provided by the present invention;

[0037] Figure 6 This is a schematic diagram of the confidence judgment module in a semantic communication system provided by the present invention;

[0038] Figure 7 This is a schematic diagram of the structure of a feature fusion module in a semantic communication system provided by the present invention;

[0039] Figure 8 This is a schematic diagram of the structure of a traditional relay multi-stage transmission;

[0040] Figure 9 A schematic diagram illustrating the structure of data transmission from edge devices to edge servers;

[0041] Figure 10a This is a diagram illustrating the blurring effect provided by the present invention.

[0042] Figure 10b This is an illustration of the occlusion processing effect provided by the present invention;

[0043] Figure 11 This is a comparison diagram of the distribution of features before and after fusion provided by the present invention;

[0044] Figure 12 The graph showing the variation of MNIST classification accuracy with SNR provided by this invention;

[0045] Figure 13 The graph showing the change in the request edge device transmission ratio as a function of the confidence threshold provided by this invention;

[0046] Figure 14a The accuracy versus threshold curve provided by this invention;

[0047] Figure 14b This is a line graph showing the transmission delay as a function of a threshold, provided by the present invention.

[0048] Figure 14c The energy consumption versus threshold curve provided by this invention;

[0049] Figure 15 Line graph comparing the accuracy of the method and random request scheme provided by this invention under different β values. Detailed Implementation

[0050] To better understand and implement this invention, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0051] The terms “comprising” and “having” and any variations thereof in this invention are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0052] like Figure 1 As shown, the relay-based semantic communication system provided in this application includes an edge device, a relay device, and an edge server, forming an edge-device joint inference system. The edge server needs to obtain target information from the edge device and the relay device to complete the final inference task. Specifically, the edge device is located near the target and has a high ability to acquire object information, enabling it to obtain relatively good relevant information about the target object, such as image data. The relay device is located further away from the target and is situated between the edge device and the edge server, serving as an auxiliary communication transmission tool, further transmitting the signals from the edge device to the edge server. The relay device also contains monitoring equipment and can obtain object information, but due to its distance from the target, the quality of the obtained images is lower, potentially resulting in low clarity or occlusion.

[0053] In traditional communication models, edge devices typically transmit their acquired information (x1) directly to a relay device. The relay device then amplifies the raw signal sent by the edge device and transmits it to the edge server. If the relay device includes a monitoring device, it will also send a feature signal extracted by the relay device. Therefore, the edge server will receive two feature signals, which are then processed to complete the target task. In this model, the communication process is divided into three time periods: the edge device transmits a signal to the relay device, the relay device transmits the signal sent by the edge device to the edge server, and the relay device transmits its own signal to the edge server. In task-oriented semantic communication systems, the final inference result often depends on the compressed and extracted feature signal, rather than necessarily sending all the information completely to the edge server. Therefore, if the information extracted by the relay device already meets the inference requirements, the edge device no longer needs to send the target feature information extracted by its own sensors.

[0054] like Figure 1 As shown, the relay-based semantic communication system provided in this application includes an edge device (ED), a relay device, and an edge server. The relay device contains a feature extractor, a confidence level assessment module, and a fusion module. The feature extractor extracts a first feature signal from the target data. The confidence level assessment module compares the confidence level with a threshold and intelligently selects whether to directly transmit the feature signal via the relay device or request the edge device to supplement it, effectively reducing redundant data transmission. While maintaining task performance similar to traditional multi-stage transmission, it reduces transmission latency and energy consumption.

[0055] This application also provides a relay device that executes a semantic relay scheduling method. The relay device independently determines whether to request information transmission from the edge device and performs scheduling accordingly.

[0056] Specifically, a semantic relay scheduling method includes:

[0057] Step S1: Collect target data and output the confidence level of the target data;

[0058] Step S2: Compare the confidence level with the threshold. If the confidence level is not less than the threshold, proceed to step S31; if the confidence level is less than the threshold, proceed to step S32.

[0059] Step S31: When the confidence level is not less than the threshold, generate a first feature signal through the local feature extractor and send the first feature signal to the edge server;

[0060] Step S32: When the confidence level is less than the threshold, a target information acquisition request is sent to the edge device, and the target information acquisition request causes the edge device to emit a second feature signal;

[0061] The second feature signal is received, and the first feature signal and the second feature signal are fused to generate a fused signal, which is then transmitted to the edge server.

[0062] By determining the confidence level to decide whether to send a message requesting the edge device, the process of information transmission to the edge device is further simplified, resulting in a reduction in overall communication latency and energy consumption. Finally, the fused signal is sent to the edge server to complete subsequent tasks. The overall system transmission timing flow is shown below. Figure 2 Transmission timing diagram of relay scheduling strategy.

[0063] The relay device first acquires the target data. The acquired target data can be obtained by the relay device through its own detection function, or it can be obtained through the edge device.

[0064] A confidence judgment module based on a neural network is built into the relay device. c Taking the target data s0 as input, the output is the confidence level α:

[0065] α=f c (s0;θ c (1)

[0066] θ c For the confidence level judgment module f c The model parameters. The confidence level α and the threshold... Compare the results. If the confidence level is greater than the threshold... This indicates that the information obtained from the relay device is of sufficient quality to support the completion of the target task. At this point, the relay device has already used a feature extractor f based on a deep neural network to extract the target data s0. r Extracting the first feature signal x0, we obtain the following:

[0067] x0 = f r (s0;θ r (1)

[0068] θ r For relay equipment feature extractor f s The model parameters. The relay device sends x0 to the edge server. The signal received at the edge server is:

[0069] y0=h0x0+n0 (2)

[0070] Where h0 is the channel coefficient of the relay device-edge server link, and n0~CN(0, σ 2The signal y0 represents Gaussian white noise in the channel. At the edge server, the received signal y0 is input to the task processing module f. d Complete the subsequent inference task:

[0071]

[0072] θ d For task processing module f d Model parameters.

[0073] If the confidence level is less than the threshold This indicates that the information quality at the relay device is low. For example, the acquired information is blurry or heavily obscured, requiring continued transmission from the edge device to improve task performance. In this case, the relay device sends a short target information acquisition request x to the edge device. req The requirement is for edge devices to transmit information to relay devices. After the monitoring device at the edge device obtains the raw target data s1, it passes through the feature extractor f inside the edge device. s Extract the second feature signal x1:

[0074] x1=f s (s1;θ s (4)

[0075] θ s It is a feature extractor inside the edge device. s The model parameters. When the edge device receives x... req Then, the edge device extracts the second feature signal x1 and transmits it to the relay device. The relay device receives the following signal:

[0076] y1=h1x1+n1 (5)

[0077] h1 is the channel coefficient of the edge device-relay device link. The noise is Gaussian white noise on the channel. After the relay device receives the second feature signal y1 from the edge device, it fuses it with its own extracted first feature signal x0. Feature fusion includes denoising the signal sent by the edge device, extracting feature information from x0 and y1 and fusing them into a higher-quality feature vector, further reducing the dimensionality of the output signal and compressing the data volume. The signal is then processed by the relay device's fusion module f. com We obtain a fused signal x2 with the same dimension as the first feature signal x0:

[0078] x2=f com (x0;y1;θ com (6)

[0079] θ com It is f comThe model parameters. After obtaining the fused signal x2, the relay device sends it to the edge server:

[0080] y2=h2x2+n2 (7)

[0081] h2 is the channel coefficient of the relay device-edge server link. It is Gaussian white noise on the channel.

[0082] At the edge server, the received fused signal y2 is input to the task processing module f. d Complete the subsequent inference task:

[0083] During training, the hyperparameters of each module are trained using the following loss function:

[0084]

[0085] Among them, l k It is the label of the k-th category corresponding to the input data s; It is the prediction result The k-th element in.

[0086] In this system, multiple modules in the relay devices, edge devices, and edge servers are all implemented based on neural networks. First, the feature extractors in the edge devices and edge servers are trained individually. Then, the trained feature extractors are used to train the confidence judgment module and fusion module in the relay devices. Finally, the feature extractors and confidence judgment module in the task processing module of the edge server and the overall system are trained. The feature extractors use the same DNN network model.

[0087] Specifically, initialize the dataset Model parameters θ s θ r θ c θ com θ d First, train a local feature extractor for the edge device using the dataset. Select training batch Model parameters θ s Through such Figure 4 The feature extractor model structure shown is trained, and the optimized parameters θ are output. s , regarding the original target data Extract features; input x1 into the classifier to obtain the prediction result. Calculate the training loss function L, and update θ through backpropagation. s ; until parameter θ s Convergence complete, training finished.

[0088] Similarly, train the feature extractor for the relay device. From the dataset... Select training batch Model parameters θ r Through such Figure 4 The feature extractor model structure shown is trained, and the optimized parameters θ are output. r , regarding the original target data Extract features; input x0 into the classifier to obtain the prediction result. Calculate the training loss function L, and update θ through backpropagation. r ; until parameter θ r Convergence complete, training finished.

[0089] The confidence assessment module for training relay devices is derived from the dataset. Select training batch Model parameters θ c Through such Figure 6 The model structure shown is trained, and the optimized parameters θ are output. c , regarding the original target data Extract features x0 and confidence level α; input x0 into the classifier to obtain the prediction result. Calculate the training loss function L, and update θ through backpropagation. c ; until parameter θ c Convergence achieved; training complete. The confidence assessment module outputs the distribution of each class using softmax, takes the maximum distribution value α, and compares it with the confidence threshold. Compare them.

[0090] The fusion module for training relay devices is derived from the dataset. Select training batch Model parameters θ com Through such Figure 7 The model structure shown is trained, and the optimized parameters θ are output. com , regarding the original target data Extract features x0 and x1, fuse x0 and x1 to obtain y1; input x0 and y1 into the fusion module to obtain the fused signal x2; input the fused signal x2 into the classifier to obtain the prediction result. Calculate the training loss function L, and update θ through backpropagation. com ; until parameter θ com Convergence has occurred, and training is complete. The fusion module outputs a feature vector with merged dimensions.

[0091] Training the edge server's task processing module and the overall system, from the dataset Select training batch Model parameters θ d Through such Figure 5The model structure shown is trained, and the optimized parameters θ are output. d , regarding the original target data Extract features x0 and confidence level α; if y0 is obtained from formula (3), and the prediction result is obtained using formula (4). like y2 is obtained from formulas (3), (5)-(8), and the prediction result is obtained using formula (9). The training loss function L is calculated using formula (10), and θ is updated via backpropagation. d ; until parameter θ d Convergence complete, training finished.

[0092] By training edge devices, relay devices, and edge servers, the confidence judgment module of the relay devices compares the confidence level with a threshold, intelligently selecting whether to directly transmit the signal to the relay device or request the edge device to supplement the feature signal, effectively reducing redundant data transmission. While ensuring task performance is similar to traditional multi-stage transmission, it reduces transmission latency and energy consumption.

[0093] The following explanation, in conjunction with traditional relay transmission methods, clarifies this application:

[0094] like Figure 8 As shown, Scheme 1 is a traditional multi-stage relay transmission method, where all information acquired by the edge device and the relay device is sent to the edge server. The specific system model is as follows: Figure 6 As shown, data transmission is divided into three main stages: In the first stage, the source data s1 obtained by the edge device is processed to obtain a feature signal x1, which is then transmitted to the relay device. In the second stage, the relay device amplifies the signal sent from the edge device and directly transmits it to the edge server. In the third stage, the relay device transmits the feature signal x0 obtained from processing data source s0 to the edge server. After receiving signals y0 and y2, the edge server inputs the two feature vectors into its task processing module f. d In the middle. It is worth noting that f d The input dimension length at this point is the sum of the dimensions of the two received signals. After passing through the task processing module with the same depth neural network structure, the final result is obtained.

[0095] like Figure 9 As shown, Scheme 2 is for comparative verification, transmitting only the information obtained from the edge device. In this mode, data is directly transmitted from the edge device to the edge server. The specific process is as follows: The source data s1 obtained by the edge device is processed by a feature extractor to obtain the feature signal x1, which is then transmitted by the edge device to the relay device. After passing through the wireless channel, the signal y3 is received by the edge server and input into the task processing module f of the edge server.d In this case, the final result is obtained. In Solution 2, the transmission distance is the distance from the edge device to the edge server, and there is no assistance from a relay during the transmission.

[0096] Solution 3 is compared with the method proposed in this application in terms of the relay decision method. Different from this application, in Solution 3, a random determination is used to judge whether to request the edge device to transmit. The probability of transmission is set as β. In this way, the edge device has a probability of β to request the edge device to transmit. After receiving the request signal, the edge device sends its own feature signal and fuses it with the features extracted by the feature extractor of the relay device at the relay device. At this time, the transmission timing is the same as the timing process of the transmission in this application. Additionally, the direct transmission of the relay device signal is performed with a probability of (1 - β). The time consumption of the transmission can be calculated in the following way after statistical averaging:

[0097] t = T×(1 - β) + 2T×β (11)

[0098] This application is Solution 4, and its transmission timing is as Figure 2 shown. When it is necessary to count the transmission delay and energy consumption, in the actual communication process, at a preset confidence threshold , the number of times determined to request the edge device to transmit is counted, and the corresponding transmission probability β is calculated. Then, the corresponding transmission time and energy consumption are calculated through formula (11). In the actual application scenario, the proportion of requests for transmission can be changed by changing the value of.

[0099] For the above 4 comparison solutions, the feature extractors in the edge devices, the feature extractor modules in the relay devices, and the task processing modules in the edge servers all adopt the same deep neural network structure to ensure a fair comparison only in terms of the relay scheduling strategy. During the transmission process, let the basic unit of time transmission be T, representing the transmission delay of the edge device signal, and the transmission delays of various solutions are expressed as multiples of T. Let the basic unit of energy consumption be E, representing the energy consumed by the edge device to send the signal to the wireless channel, and the total transmission energy consumption of various solutions is expressed as multiples of E.

[0100] The distance between the edge device and the edge server is set to 100 meters, marked as d0 = 100m. The relay device is deployed on the straight line between the edge device and the edge server, and the distance from it to the edge device is marked as d1, where 0 < d1 < 100. Correspondingly, the distance from the relay device to the edge server is d2 = d0 - d1. Considering the path loss of the channel, the channel gain is set as |h i |2 = G(3×10 8 / 4πf c d i ) εWhere i = 0, 1, 2. Where G = 4.11 is the antenna gain, f c =915MHz is the carrier frequency, and ε=3 is the path loss index.

[0101] MNIST was used as the dataset in the experiments. The MNIST dataset was divided into three parts for testing: 50k training samples, 10k validation samples, and 10k test samples. The Adam optimizer was used with a learning rate of 0.001. Furthermore, the batch size for each training set was set to 128, and training was performed for 30 iterations. All simulations were conducted on an Intel(R) Xeon(R) Silver4110 (2.10GHz) machine and an NVIDIA 2080ti GPU.

[0102] Two contrasting methods were used to process the relay equipment's dataset: Gaussian blurring and masking. Gaussian blurring is a common image processing technique that smooths an image by applying a Gaussian kernel function, reducing noise and detail. In the experiment, Gaussian blurring was applied to the test images in the MNIST dataset, and the blur level followed a normal distribution across the test set; that is, the blur level of each test image was random, but overall it followed a normal distribution. Masking is also a common image processing technique that simulates data loss or occlusion by randomly masking some pixels on an image. In the experiment, random masking was applied to the test images in the MNIST dataset, and the masking ratio followed a normal distribution across the test set; that is, the masking ratio of each test image was random, but overall it followed a normal distribution. The masking ratio refers to the proportion of masked pixels to the total number of pixels in the image. By controlling the distribution of the masking ratio, different degrees of occlusion can be simulated. Both methods were used to simulate situations where the source data obtained at the relay point is of low quality, such as low device resolution or pedestrian obstruction, with processing effects as follows: Figure 10a , 10b As shown.

[0103] As shown in Table 1, when the SNR is -3dB, The performance comparison results of Method 1, Method 2, Method 3, and Method 4 are as follows:

[0104] Table 1. Performance comparison of different schemes for classification tasks

[0105]

[0106] As shown in Table 1, among all the schemes, Scheme 1 has the best task performance because the transmitted data contains all data related to the target. In the scheme with only edge device transmission, although the original source s1 obtained at the edge device is of relatively high quality, the impact of channel fading on the signal is much greater than in the method with relay equipment without relay equipment to assist transmission, thus the performance is relatively poor. Because the source and equipment at the relay equipment are not used, the results are the same for the data sources under both fuzzing and masking processing methods. In Scheme 4 provided in this application, when At that time, statistical analysis showed that the proportion requiring edge device transmission was β = 0.4842, and the performance under this method was far superior to the mode with only edge device transmission. Simultaneously, the performance was close to that of the traditional relay-assisted transmission method. Correspondingly, the performance of the random transmission scheme with β = 0.4842 was lower than that of Scheme Four. Regarding transmission latency, the calculated average latency was 1.48T, a 50% reduction compared to the traditional Scheme One. Regarding communication energy consumption, the calculated average energy consumption was 1.48E, a 50% reduction compared to the traditional Scheme One. Therefore, Scheme Four proposed in this application, through relay scheduling, reduces transmission latency and energy consumption while maintaining system performance close to that of traditional schemes.

[0107] Figure 11 In the comparison diagram of feature distribution before and after feature fusion, the leftmost image shows the feature signal x1 after passing through the feature processor at the edge device; the middle image shows the edge device signal x2 sent at the relay device in Scheme 1; and the rightmost image shows the fused signal x2 sent at the relay device in the invention. Figure 11 As can be seen, the feature signals extracted from the edge devices experience a decrease in channel noise quality, while the fusion module at the relay device reduces the impact of noise on the feature signals.

[0108] Figure 12 This display shows the performance of Schemes 1 through 4 for a classification task on the MNIST dataset, under different signal-to-noise ratios. The image processing method at the relay point is occlusion, and a confidence threshold is set. As shown in the figure, the scheme with only edge device transmission exhibits the lowest performance. Classification accuracy gradually increases with improving channel quality. The scheme based on relay decision-driven random transmission performs worse than the scheme of this invention because the image source at the relay device is affected by occlusion. Therefore, after the SNR increases to a certain level, the highest accuracy of scheme three is lower than other schemes. The semantic relay scheduling strategy of this invention consistently maintains an inference accuracy close to that of traditional multi-stage transmission schemes at different SNRs, with the accuracy difference not exceeding 3%. Compared to the random request edge device transmission scheme, it shows a maximum improvement of 6.8%, and compared to the no-relay scheme, it shows a maximum improvement of 12%.

[0109] To more intuitively illustrate the confidence threshold The impact of the value on relay scheduling performance and resource optimization was statistically analyzed in the test results of the MNIST dataset, with SNR = -3dB. The graph showing the relationship between the retransmission ratio β and the retransmission rate β is as follows: Figure 13 It can be seen that, with As the value increases, the proportion of data transmission required by edge devices gradually increases. This can be achieved by changing... The value is adjusted to determine the proportion of data transmitted by the edge device in the invention, thereby achieving a trade-off between performance and communication resources.

[0110] As the confidence level increases, the judgment of image quality at the relay device becomes more stringent, leading to more requests from edge devices. This improves the accuracy of the inference results, but the increased proportion of requests from edge devices also increases the average transmission time and average transmission energy consumption. Figures 14(a), (b), and (c) show the relationship between accuracy, latency, and energy consumption as the confidence level changes. As can be seen from the figures, As the speed increases, mission accuracy gradually improves, but the corresponding transmission latency and energy consumption also increase accordingly. Performance improvements will gradually become apparent in the future. At around 0.9, the task performance has reached 93.32%, but the transmission latency and energy consumption are only half that of multi-stage relay transmission.

[0111] At the same time, Figure 15 The graph shows a performance comparison between the proposed solution and the random request transmission solution under different retransmission ratios. The horizontal axis represents the proportion β of requests from edge devices, and the vertical axis represents the task accuracy. According to... The β values ​​corresponding to the range of 0.1-0.9 were statistically analyzed, and the corresponding task precision was recorded. The graph shows that by adding a judgment module at the relay point to determine the request transmission, the performance is higher than that of random requests, given the same proportion of requests transmitted to edge devices. Especially in… That is, when β = 0.347, the inference accuracy has a difference of up to 4%.

[0112] This invention intelligently selects relay devices to directly transmit or request supplementary feature signals from edge devices based on the output confidence level, effectively reducing redundant data transmission. While maintaining task performance similar to traditional multi-stage transmission, it reduces transmission latency and energy consumption. Furthermore, the adjustability of the confidence threshold allows for a flexible trade-off between task performance and resource consumption, significantly improving system efficiency and resource utilization in high-density device scenarios.

[0113] Based on the same inventive concept, the present invention also provides a computer device, comprising: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the steps of the semantic relay communication method described above.

[0114] The processing methods for computer devices can be referred to the description of the methods above, and will not be repeated here.

[0115] This application also provides a non-transitory machine-readable storage medium storing an executable program, which, when run by a microprocessor, causes the processor to execute the method provided in the above embodiments.

[0116] This invention discloses a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform the described methods.

[0117] This invention discloses a computer program product including a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform the described method.

[0118] The embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0119] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0120] Finally, it should be noted that the embodiments disclosed in this invention are merely preferred embodiments of this invention and are only used to illustrate the technical solutions of this invention, not to limit it. Although this invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention.

Claims

1. A method of semantic relay communication, characterized by, The method is performed by a relay device, comprising: collecting target data, generating a first feature signal by a local feature extractor; outputting a confidence level of the target data, comparing the confidence level with a threshold value; when the confidence level is not less than the threshold value, sending the first feature signal to an edge server; when the confidence level is less than the threshold value, sending a target information acquisition request to an edge device, the target information acquisition request causing the edge device to send a second feature signal; receiving the second feature signal, fusing the first feature signal and the second feature signal to generate a fused signal, and transmitting the fused signal to the edge server.

2. The semantic relay communication method of claim 1, wherein, The confidence level is calculated by: inputting the first feature signal into a confidence level judgment module based on a deep neural network, outputting a class probability distribution by a Softmax function, and taking the maximum probability value as the confidence level.

3. The semantic relay communication method of claim 1, wherein, The threshold value is a dynamic adjustable parameter; according to the changes of source data form and channel environment, the threshold value is determined by traversal method.

4. The semantic relay communication method according to claim 3, characterized in that, The threshold value decreases with the increase of the task accuracy of the edge server, and the threshold value decreases with the decrease of the real-time signal-to-noise ratio.

5. The semantic relay communication method of claim 1, wherein, Fusing the first feature signal and the second feature signal to generate a fused signal, comprising: weighting and denoising the first feature signal and the second feature signal, compressing the feature dimension by a fully connected layer to generate a fused signal.

6. The semantic relay communication method of claim 5, wherein, The dimension of the fused signal is not greater than the dimension of the first feature signal.

7. A relay device, characterized by, The method for semantic relay communication is performed according to any one of claims 1-6; the relay device comprises a confidence level judgment module, a feature extractor, and a fusion module; The feature extractor is used to generate a first feature signal from target data; The confidence level judgment module is used to output a confidence level, compare the confidence level with a threshold value, and send the first feature signal to an edge server when the confidence level is not less than the threshold value; When the confidence level is less than the threshold value, a target information acquisition request is sent to an edge device, which causes the edge device to send a second feature signal; The fusion module is used to generate a fused signal according to the first feature signal and the second feature signal, and transmit the fused signal to an edge server.

8. The relay device of claim 7, wherein, The confidence level judgment module, the feature extractor, and the fusion module use a deep neural network, and the loss function of the deep neural network is: wherein, is for input data a label of the corresponding k-th class; is a prediction result the k-th element in.

9. A semantic relay communication system, characterized by comprising: The relay device according to any one of claims 7-8; The edge device receives the target information acquisition request, collects target information, and sends a second feature signal; The edge server receives the fused signal.

Images