A Pairing Method for D2D Mutual Learning System Based on Unauthorized Band Transmission
The method optimizes device pairing and resource allocation in D2D-U systems on unlicensed frequency bands to address coexistence and energy challenges, ensuring fair coexistence with Wi-Fi and reducing computational complexity.
Patent Information
- Application Number
- CN202211562776.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-07
- Publication Date
- 2025-07-15
- Estimated Expiration
- 2042-12-07
AI Technical Summary
In the authorization-free frequency band, D2D-U technology faces fair coexistence problems when sharing the frequency band with Wi-Fi networks, power and spectrum allocation complexity, and channel access uncertainty, resulting in high computational complexity and difficulty in achieving efficient mutual learning tasks.
The equipment list is determined through the base station, and the duty cycle silent mechanism is used to access the authorization-free channel. After the equipment broadcasts information, the base station optimizes the equipment pairing and resource allocation, and combines deep neural network model training to realize wireless model parameter transmission and mutual learning between devices.
On the premise of ensuring fair coexistence with the Wi-Fi network, the average energy consumption of the system is reduced and efficient mutual learning tasks are completed between multiple devices.
Smart Images
Figure CN116170781B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of wireless communication and deep learning. More precisely, it relates to a pairing method for a D2D mutual learning system based on transmission in an unlicensed band. Background Art
[0002] Mutual Learning (ML) is an emerging distributed learning method proposed after federated learning. Its concept originates from the Knowledge Distillation (KD) branch in the field of deep learning. Compared with the simple averaging of all model parameters in traditional federated learning, mutual learning improves the performance of small deep neural networks through mutual knowledge distillation between models. During the training process, each network not only receives supervision from the labels of the local dataset but also refers to the learning experiences of other peer networks to further enhance the generalization ability of the model.
[0003] However, due to the mobility of devices, in most cases, model parameter transmission can only be carried out through wireless transmission rather than wired transmission. The huge amount of data will bring an unprecedented increase in mobile traffic, which poses a great challenge to the existing mobile cellular network. To achieve higher data rates, there are currently two potential research directions: one is to develop revolutionary technologies to improve spectral efficiency, such as Device-to-Device (D2D) communication. As one of the key technologies for improving spectral efficiency in the 5G (5th Generation Mobile Communication Technology) era, D2D technology allows two terminal devices to directly communicate by reusing existing channels without the direct participation of a base station. The other is to utilize as much spectrum as possible, such as unlicensed spectrum and millimeter-wave bands. Since the existing cellular network communication technology has achieved very high spectral efficiency, there is very little room for further effective improvement. At the same time, the unlicensed bands used by Wireless Local Area Network (WLAN) systems are widely underutilized. Therefore, turning to unlicensed bands is an inevitable trend to increase data rates.
[0004] Although D2D technology in licensed bands has been relatively widely studied, the D2D technology in unlicensed bands (Device-to-Device-Unlicensed, D2D-U) is still in its infancy and faces many technical challenges: First, the fair coexistence problem when the D2D-U network shares the unlicensed band with the Wi-Fi (wireless fidelity) network; second, discrete power and spectrum allocation schemes need to be adopted to adapt to the distributed structure of the D2D-U network, so as to avoid excessive signaling overhead; third, the uncertainty of channel access of existing Wi-Fi systems in the unlicensed band makes the performance analysis of D2D-U communication more complex and brings unbearable computational complexity. Summary of the Invention
[0005] The object of the present invention is to overcome the deficiencies in the prior art and provide a pairing method for a D2D mutual learning system based on unlicensed band transmission, including:
[0006] Step 1: Determine the list of mobile devices participating in the mutual learning system;
[0007] Step 2: All devices determine the structure of their local deep neural network models and initialize the parameters of the local deep neural network models. Subsequently, model training is performed according to the local data sets, and the model parameters are updated in each round of training;
[0008] Step 3: After all devices complete local model training, broadcast their own information to the base station;
[0009] Step 4: After the base station receives the information of all devices, aiming at minimizing the average energy consumption of the mutual learning system, give the optimal device pairing scheme and the optimal communication resource allocation scheme, and send the scheme to all devices in the mutual learning system;
[0010] Step 5: All devices in the mutual learning system are paired according to the device pairing scheme. The formed device pairs include two parties, device A and device B; after the communication is established, device A sends the local model parameters to device B in a wireless transmission form;
[0011] Step 6: After device B receives the model parameters of device A, perform several rounds of mutual learning model training based on the local data set and the model parameters of both parties, and update the model parameters in each round of training; when the mutual learning process ends, device B replaces the original local model with the updated model and sends the new model parameters to device A;
[0012] Step 7: After receiving the new model parameters transmitted back by Device B, Device A also performs mutual learning model training for several rounds based on the local dataset and the new model parameters transmitted back by B. The model parameters are updated in each round of training. After the mutual learning process ends, Device A replaces the original local model with the updated model.
[0013] Preferably, in Step 1, the mobile devices on the mobile device list all access the unlicensed channel using the Duty Cycle Muting (DCM) mechanism.
[0014] Preferably, in Step 2, each local dataset has a unique global number and also corresponds to a unique "knowledge".
[0015] Preferably, in Step 3, the information about the device broadcast to the base station includes: the transmitting power of the device antenna, the circuit power of the device, the power of the device sensing the channel state, the fraction of available time, the amount of local model parameter bits held, and the set of "knowledge" held.
[0016] Preferably, in Step 7, it is determined whether there are device pairs in the mutual learning system that have not completed mutual learning. When all device pairs have completed mutual learning, it is regarded as the end of the entire process of the mutual learning system.
[0017] The beneficial effects of the present invention are as follows: The present invention combines the mutual learning technology and the D2D-U technology, focuses on the distributed learning system composed of multiple mobile devices, designs a scheme for joint allocation of device spectrum and power aiming at completing the mutual learning task, and minimizes the average energy consumption of the system with relatively low computational complexity on the premise of ensuring the fair coexistence of D2D-U and Wi-Fi networks, ensuring that all devices can complete the mutual learning task. Description of the Drawings
[0018] Figure 1 It is a flowchart of the D2D mutual learning system pairing method based on unlicensed band transmission provided by the present invention;
[0019] Figure 2 It is a schematic diagram of the D2D mutual learning system model in the embodiment;
[0020] Figure 3 It is a schematic diagram of the data frame structure in the DCM channel access mechanism;
[0021] Figure 4 It is a schematic diagram of mutual learning between the two communication parties of D2D-U after pairing is completed. Detailed Embodiments
[0022] The present invention will be further described below in conjunction with embodiments. The description of the following embodiments is only for helping to understand the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.
[0023] As an embodiment, the present invention provides a D2D mutual learning system pairing method based on unlicensed band transmission, as Figure 1 shown, including:
[0024] Step 1: Determine the list of mobile devices participating in the mutual learning system.
[0025] Exemplarily, the mutual learning system operates in an unlicensed band, and there is a base station and N mobile devices in the system. In this example, N = 4 is set. At the same time, there are also several unknown Wi-Fi users in the unlicensed band for data transmission, as detailed in Figure 2 in the accompanying drawings of the specification. In order to ensure that D2D communication can coexist fairly with Wi-Fi users in the unlicensed band, the present invention stipulates that mobile devices access the unlicensed band with the DCM mechanism. The schematic diagram of the data frame structure of the DCM mechanism is detailed in Figure 3 in the accompanying drawings of the specification. In the DCM mechanism, a data frame will be divided into several sub-data frames (subframe), and a part of the sub-data frames will be used by Wi-Fi users, while D2D devices are only allowed to perform data transmission on the remaining sub-data frames. When Wi-Fi users use the sub-data frames allocated to them for data transmission, the unlicensed sub-channel will be fully occupied by Wi-Fi users, and D2D users must remain silent and continuously monitor the channel to estimate the data traffic of Wi-Fi users, so as to determine how many sub-data frames can be allocated in the next data frame. The proportion of the number of sub-data frames allocated to D2D devices in the total number is called the time fraction, denoted as ρ. The part of the sub-data frames allowed for D2D devices to transmit is collectively referred to as the on-period; similarly, for D2D devices, the sub-data frames only allowed for Wi-Fi users to transmit are collectively referred to as the off-period.
[0026] Step 2: All devices determine the structure of their local deep neural network models and initialize the parameters of the local deep neural network models, and then perform model training according to the local data sets, and update the model parameters in each round of training.
[0027] The present invention defines that each local data set has a unique global number, and at the same time corresponds to a unique "knowledge". When a device completes the training for a certain local data set, it means that it has mastered the "knowledge" corresponding to the data set with this number. If a device has multiple different local data sets, then after all the training is completed, the device has mastered multiple different "knowledges".
[0028] Exemplarily, in step 2, the local model structures of all devices are ResNet32 (on the premise of ensuring that the softmax output layer structures of all models are the same, the models of the devices can be freely replaced with other mainstream neural network structures). Denote the local data set of device i as D i , the device first initializes the model parameters, and then uses the local data set for a total of T1 rounds of training, and continuously updates the model parameters. In this embodiment, T1 = 50 is set. After the training is completed, denote the model parameters of device i as ω i , and at the same time denote the "knowledge" set mastered by device i after the training is completed as M i , the size of the set is positively correlated with the number of local data sets owned by the device, and each local data set uniquely corresponds to a piece of "knowledge" in the set. For example, if device i has data sets numbered 1, 2, and 3, then the "knowledge" set of device i after the local training is completed is denoted as M i = {1, 2, 3}. Similarly, if device j has data sets numbered 1, 3, and 5, then the "knowledge" set of device j after the local training is completed is denoted as M j = {1, 3, 5}. Since there may be data sharing behaviors between some devices, there is a possibility that multiple devices have one or more identical data sets. For example, in the above example, device i and device j both have two data sets numbered 1 and 3 at the same time.
[0029] Step 3: After all devices complete the local model training, they broadcast their own information to the base station.
[0030] The information that the device broadcasts to the base station includes: the transmitting power p of the device antenna, the circuit power p c of the device, the power p s of the device for sensing the channel state, the fractional time ρ available, the number of bits of the local model parameters held, and the "knowledge" set M held.
[0031] Step 4: After the base station receives the information of all devices, with the goal of minimizing the average energy consumption of the mutual learning system, it gives the optimal device pairing scheme and the optimal communication resource allocation scheme (including the optimal transmitting power and the optimal fractional time for each device), and sends the scheme to all mobile devices in the mutual learning system.
[0032] Step 5. All devices in the mutual learning system are paired according to the device pairing scheme, and the formed device pairs include device i and device j; after the communication is established, device i sends its local model parameters to device j in the form of wireless transmission.
[0033] For example, as Figure 4 shown, device i will establish a connection with device j specified by the base station, and according to the optimal transmission power p and optimal time fraction ρ given by the base station, send its local model parameters ω i to device j.
[0034] Step 6. After receiving the model parameters ω i from device i, device j performs a total of T2 rounds of mutual learning model training based on the local dataset D j and the model parameters of both sides. In this example, T2 = 50 is set, and the model parameters are updated in each round of training; when the mutual learning process ends, device j replaces the original local model with the updated model and sends the new model parameters ω′ j to device i.
[0035] Step 7. After receiving the new model parameters ω′ j sent back by device j, device i also performs a total of T2 rounds of mutual learning model training based on the local dataset D i , the new model parameters ω′ j sent back by j, and its local model parameters ω i , and the model parameters are updated in each round of training; after the mutual learning process ends, device i replaces the original local model with the updated model and updates its local model parameters to ω′ i .
[0036] When both sides of a device pair have completed the mutual learning training, it is regarded that the device pair has successfully completed the mutual learning task. Check whether there are device pairs in the system that have not completed the mutual learning task. When all the mutual learning device pairs in the system have completed steps five, six, and seven, it can be regarded that the entire process of the mutual learning system ends.
Claims
1. A pairing method for a D2D mutual learning system based on unlicensed band transmission, characterized in that, Including: Step 1: Determine the list of mobile devices participating in the mutual learning system; Step 2: All devices determine their local deep neural network model structures and initialize the local deep neural network model parameters. Subsequently, model training is performed according to the local dataset, and the model parameters are updated in each round of training; Step 3: After all devices complete local model training, broadcast their own information to the base station; Step 4: After the base station receives the information of all devices, aiming to minimize the average energy consumption of the mutual learning system, give the optimal device pairing scheme and the optimal communication resource allocation scheme, and send the schemes to all devices in the mutual learning system; Step 5: All devices in the mutual learning system are paired according to the device pairing scheme. The formed device pairs include two parties: device A and device B. After communication is established, device A sends the local model parameters to device B in the form of wireless transmission; Step 6: After device B receives the model parameters of device A, perform several rounds of mutual learning model training based on the local dataset and the model parameters of both parties, and update the model parameters in each round of training. When the mutual learning process ends, device B replaces the original local model with the updated model and sends the new model parameters to device A; Step 7: After device A receives the new model parameters sent back by device B, also perform several rounds of mutual learning model training based on the local dataset and the new model parameters sent back by B, and update the model parameters in each round of training; After the mutual learning process ends, device A replaces the original local model with the updated model.
2. The pairing method of the D2D mutual learning system based on unlicensed band transmission according to claim 1, characterized in that, In Step 1, the mobile devices on the mobile device list all use the duty cycle silent mechanism to access the unlicensed channel.
3. The D2D mutual learning system pairing method based on unlicensed band transmission according to claim 1 or 2, characterized in that In Step 2, each local dataset has a unique global number and also corresponds to a unique "knowledge".
4. The method for pairing a D2D mutual learning system based on unlicensed band transmission according to claim 3, characterized in that In Step 3, the information that the device broadcasts to the base station includes: device antenna transmit power, device circuit power, device sensing channel state power, available time fraction, the amount of local model parameter bits held, and the "knowledge" set held.
5. The pairing method of the D2D mutual learning system based on unlicensed band transmission according to claim 4, wherein In Step 7, determine whether there are device pairs in the mutual learning system that have not completed mutual learning. When all device pairs have completed mutual learning, it is regarded as the end of the entire process of the mutual learning system.