Pseudo-neural network transmission method based on air computing

Abstract

The invention discloses a pseudo-neural network transmission method based on air computing, and mainly solves the problems of large information transmission error and large resource overhead in the prior art. According to the implementation scheme, an air computing quasi-neural network transmission platform comprising a source node, a relay node and a destination node is built; initializing parameters of the pseudo-neural network; each node regulates and controls the signal according to the processing weight parameter, transmits a training symbol from a source node to a relay node and then forwards the training symbol to a destination node; the destination node calculates an error and an error gradient, reversely propagates the error and the error gradient to the relay node, and forwards the error and the error gradient to the source node to update processing weight parameters of each node; the forward transmission of the training symbols and the back propagation of the error gradient are repeatedly executed until the conditions are met; and transmitting the to-be-transmitted information from the source node to the relay node according to the current processing weight parameter, and forwarding the to-be-transmitted information to the destination node to recover the to-be-transmitted information. The method can reduce information transmission errors and node resource overhead, and can be used for wireless ad hoc network information transmission under the condition that node resources are limited.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication, and in particular relates to a quasi-neural network transmission method, which can be used for reliable information transmission in a wireless self-organizing network with limited node computing resources and time slot resources. Background Art

[0002] A wireless self-organizing network is a distributed network established by multiple nodes with wireless transceiver functions using wireless communication technology. It does not require any infrastructure support. Nodes form and maintain multi-hop wireless networks through self-organization and adaptive mechanisms without centralized management. Compared with traditional communication networks, wireless self-organizing networks have some outstanding characteristics, such as no center, self-organization, multi-hop, dynamic network topology, temporary, etc. These characteristics enable it to break through the geographical limitations of traditional wireless cellular networks and can be deployed more quickly, conveniently and efficiently. Therefore, it is suitable for communication needs in emergency situations such as emergency rescue.

[0003] The network nodes of wireless ad hoc networks generally use sensors, unmanned devices, etc., with weak computing power and short service life. Communication resources such as codewords, frequency bands, channels, etc. are limited. The physical characteristics of the wireless channel itself determine that the network bandwidth it can provide is much less than that of the wired channel. Due to factors such as collisions, signal fading, noise interference, and inter-channel interference caused by competing for shared wireless channels, its actual bandwidth is far less than the theoretical value. Due to the existence of these factors, limited resources are needed in wireless ad hoc networks to perform complex operations such as encoding and decoding, channel estimation, and channel equalization, which poses a challenge to the computing power and communication resources of network nodes. Therefore, how to use limited resources for reliable transmission is an urgent problem to be solved in wireless ad hoc networks.

[0004] Wireless ad hoc networks have attracted much attention from researchers due to their importance in practical applications. Wireless ad hoc networks are very different from traditional communication networks and require special technologies to provide communication guarantees. Domestic and foreign scholars have done a lot of research on how to improve the effectiveness and reliability of wireless transmission in wireless ad hoc networks.

[0005] In their paper "A New Multicast Network Based on Clustering", Lu Ling and other authors introduced clustering algorithms into relay networks to address the problems of multiple construction of effective routes, high overhead and delay, and low network transmission efficiency in relay networks. They proposed a BRN transmission mechanism based on node clustering, C-BRN. This transmission mechanism combines the routing algorithm in traditional MANET that completely relies on topology change information with the characteristics of BRN that does not rely on topology information changes. The network node structure is hierarchical and the CBR established between cluster heads is repeatedly used to transmit information, effectively avoiding the delay and overhead caused by the reconstruction of CBR in a simple BRN transmission mechanism and improving the routing efficiency of the network. However, each node needs to perform a large number of encoding and decoding, channel estimation and equalization operations at the physical layer, which requires a lot of computing resources. When the resources of the wireless self-organizing network are limited, it cannot improve the efficiency of information transmission.

[0006] Huang Chen and other authors published a paper titled "Data Distribution Mechanism for Wireless Self-Organizing Networks Based on Network Coding" in which they believed that the energy of mobile stations in wireless self-organizing networks is limited and the number of data packet exchanges between intermediate nodes needs to be reduced. They proposed a data distribution mechanism CRNC based on network coding, which combines source information through network coding, reduces the dependence of the destination on a single copy of data, and reduces the cache requirements for intermediate nodes. Compared with the traditional data distribution mechanism based on flooding, this mechanism greatly reduces the number of transmissions while only increasing a small amount of delay, but increases the computational complexity of the receiving node, posing a challenge to the computing power of a single node. When the computing power of a single node is insufficient, it is easy to cause low information transmission efficiency.

[0007] In his paper "Design and Implementation of End-to-End Transmission Protocol Based on Fountain Codes in Self-Organizing Networks", author Geng Jiabo proposed a feedback-based staged fountain code transmission scheme to address the high packet loss rate in mobile self-organizing networks. The sending end is divided into two stages: original packet sending and robust solitary wave distribution sending. In the original packet sending stage, only the original packet is sent without encoding. After receiving the original packet, the receiving end feeds back the decoded original packet sequence number; in the robust solitary wave distribution sending stage, the sending end sends the fountain code packet according to the robust solitary wave distribution, and uses the feedback information to speed up the transmission of the fountain code packet and reduce the decoding overhead of the receiving end. However, this scheme cannot avoid the large transmission delay of the fountain code, is not sensitive to real-time requirements, reduces the information transmission rate when the transmission time slot is tight, and requires the receiving node to have a high computing power, which increases the difficulty of node deployment. Summary of the invention

[0008] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a quasi-neural network transmission method based on air computing to reduce resource overhead and bit error rate in the end-to-end signal transmission process of the self-organizing network, alleviate the computing pressure of the nodes, and improve the information transmission rate.

[0009] The technical idea to achieve the purpose of the present invention is to use the universal software radio peripheral platform USRP to build a quasi-neural network transmission platform based on air computing, and train the quasi-neural network processing weight parameters through the two processes of forward transmission of training symbols and back propagation of error gradients to adapt to unknown channel characteristics, so as to reliably transmit information without complex encoding and decoding operations, channel estimation and channel equalization operations, and only simple linear calculations are required for a single node. The signal fusion of nodes in the wireless self-organizing network is transferred to the wireless transmission process through air computing technology to reduce the computing pressure of the nodes, reduce the node information fusion steps, increase the information transmission rate, and reduce the resource overhead and bit error rate in the end-to-end signal transmission process of the wireless self-organizing network.

[0010] According to the above ideas, the technical solution of the present invention includes the following steps:

[0011] (1) Use the universal software radio peripheral USRP and the host computer to build an air computing quasi-neural network transmission platform including a source node S, n relay nodes R, and a destination node D, and synchronize the time of each USRP through a clock synchronizer;

[0012] (2) Construct the quasi-neural network module, air computing module and USRP control module:

[0013] 2a) establishing a quasi-neural network module including a source signal processing submodule, a relay signal processing submodule, a destination signal processing submodule, a source parameter updating submodule, a relay parameter updating submodule and a destination parameter updating submodule, for regulating the signal and updating the processing weight parameters;

[0014] 2b) establishing an air computing module including an air superposition pre-processing submodule and an air superposition post-processing submodule, which is used to transfer the position where the multi-channel signals are fused from the receiving end to the air channel;

[0015] 2c) establishing a USRP control module including a sending submodule, a receiving submodule and a synchronization submodule for sending and receiving signals at a specified time;

[0016] (3) Deploy all submodules in the quasi-neural network module, the aerial computing module, and the USRP control module to each host computer;

[0017] (4) Initialize the pseudo-neural network parameters of each node, including the source node processing weight parameter u and the i-th relay node processing weight parameter v i , the destination node processes the weight parameter w, the learning rate lr, the error threshold λ, the training round e=0, and at the same time initializes the USRP timestamp to 0, sets the time slot length T, and the random number generator seed seed,i=1,…,n;

[0018] (5) Use the training symbol forward transmission and error gradient back propagation to update the processing weight parameters of each node:

[0019] (5a) Process the training symbol p generated by the source node to obtain the baseband signal y of the source node and send it;

[0020] (5b) The i-th relay node receives the baseband signal y from the source node through the wireless channel and obtains the relay signal r i And process it to get the baseband signal c of the relay node i and send;

[0021] (5c) The destination node receives the baseband signal c from the relay node after the wireless channel is superimposed i , get the fused signal d, and process it to get the predicted symbol x;

[0022] (5d) Calculate the error J between the predicted symbol x and the training symbol p, broadcast it to the source node and the relay node, and compare the error signal J with the set error threshold λ:

[0023] If J>λ, then execute step (5e),

[0024] Otherwise, execute step (6);

[0025] (5e) Update the destination node processing weight parameter w, calculate and send the error gradient of the destination signal z

[0026] (5f) The i-th relay node receives the error gradient of the target signal and obtains the target gradient signal Using the target gradient signal Update the processing weight parameter v of the i-th relay node i , and the relay signal r i The error gradient Send it;

[0027] (5g) The source node receives the relay signal error gradient, obtains the relay gradient signal φ, uses the relay gradient signal φ to update the source node processing weight parameter u, updates the training round e=e+1, and returns to step (5a);

[0028] (6) The source node, relay node, and destination node stop the reverse propagation process;

[0029] (7) According to the current processing weight parameters of each node, the source node sends the information to be transmitted, the relay node receives and forwards it to the destination node, and the destination node receives and restores the information to be transmitted.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] Firstly, the present invention builds a pseudo-neural network in a wireless ad hoc network of relay cooperative communication, and only uses the processing weight parameters of each node to regulate the signal, and performs nonlinear processing to transmit the signal. The node does not need to perform complex encoding and decoding and channel estimation operations, so the computing power requirements of the node are low, which can reduce resource overhead.

[0032] Secondly, the present invention introduces air computing, which transfers the location where multiple signals in the pseudo-neural network are fused from the receiving end to the air channel, thereby further alleviating the node computing pressure, reducing the node's dependence on computing resources, and saving the time for signal fusion at the receiving end, thereby improving the information transmission rate.

[0033] Thirdly, since the present invention uses USRP to build a quasi-neural network transmission platform based on air computing, it effectively verifies the practicality, flexibility and scalability of the present invention in a practical environment. People can quickly deploy and configure communication nodes according to different scenarios, providing a new and feasible technical approach for reliable information transmission in wireless ad hoc networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the implementation of the present invention;

[0035] Figure 2 This is a diagram of a quasi-neural network transmission platform based on air computing built using USRP in the present invention;

[0036] Figure 3 It is a transmission effect simulation diagram of the present invention;

[0037] Figure 4 It is a time slot consumption simulation diagram of the present invention. DETAILED DESCRIPTION

[0038] The examples and effects of the present invention are further described in detail below in conjunction with the accompanying drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and do not constitute improper limitations on the present invention.

[0039] The implementation of the present invention is based on USRP and a host computer installed with LabVIEW NXG software to build a quasi-neural network transmission platform based on air computing.

[0040] Reference Figure 1 , the implementation steps of this example are as follows:

[0041] Step 1: Build a quasi-neural network transmission platform based on air computing.

[0042] like Figure 2 As shown, this example uses four USRPs and their host computers. One of the USRPs and their host computers is selected as the source node, two USRPs and their host computers are selected as relay nodes, and one USRP and its host computer is selected as the destination node. Each USRP is connected to the corresponding host computer through a Gigabit Ethernet cable or a PCIe high-speed cable.

[0043] Install the USRP driver in the host computer, and install the ANT antenna on the TX1 interface of channel 0 and the RX1 interface of channel 1 of each USRP;

[0044] The clock synchronizer inputs two synchronization signals PPS and REF to each USRP through SMA cables.

[0045] Step 2: Build the quasi-neural network module, air computing module and USRP control module.

[0046] (2.1) Establishing a quasi-neural network module including a source signal processing submodule, a relay signal processing submodule, a destination signal processing submodule, a source parameter updating submodule, a relay parameter updating submodule and a destination parameter updating submodule, for regulating the signal and updating the processing weight parameters:

[0047] The source signal processing submodule, whose input is the training symbol p, is used to adjust p according to the processing weight parameter u to obtain the initial signal: s=up, and nonlinearly process the signal s through the nonlinear transformation function σ, and output the baseband signal to be sent by the source node:

[0048] The relay signal processing submodule has a relay signal r as input. i , used to process the weight parameter v i R i Control and obtain the transfer signal: a i =v i H r i , and transforms the signal a through a nonlinear transformation function σ i Perform nonlinear processing and output nonlinear transformation signal: Where H represents the conjugate transpose operation;

[0049] The target signal processing submodule, whose input is the target signal z, is used to adjust z according to the processing weight parameter w to obtain the predicted signal x=w H z, where H represents the conjugate transpose operation;

[0050] The source parameter update submodule, whose input is error J, processing weight parameter u, training symbol p, initial signal s, source node to-be-sent baseband signal y, relay gradient signal φ and pseudo-neural network learning rate lr, is used to calculate the error gradient of the conjugate of processing weight parameter u: And update the source node processing weight parameter according to the error gradient conjugate of the processing weight parameter u: in, * indicates the conjugation operation;

[0051] The relay parameter update submodule takes as input the error J and the processing weight parameter v i , relay signal r i , transfer signal a i , nonlinear transformation signal b i , target gradient signal and the quasi-neural network learning rate lr, which is used to calculate the processing weight parameter v i Conjugate error gradient: Calculate the relay signal r i The error gradient is: And according to the processing weight parameter v i The conjugate error gradient updates the weight parameter v i ′:

[0052] in

[0053] The target parameter update submodule, whose input is error J, processing weight parameter w, target signal z, prediction signal x, training symbol p and pseudo-neural network learning rate lr, is used to calculate the error gradient of the conjugate of processing weight parameter w: Calculate the error gradient of the target signal z: Update the processing weight parameter according to the error gradient conjugated with the processing weight parameter w: in,

[0054] (2.2) Establishing an air computing module including an air superposition pre-processing submodule and an air superposition post-processing submodule, which is used to transfer the location where the multi-channel signals are fused from the receiving end to the air channel:

[0055] The air superposition preprocessing submodule has a nonlinear transformation signal b as input. i , used to preprocess the nonlinear transformation signal to obtain the baseband signal to be sent by the relay node in

[0056] The aerial superposition post-processing submodule, whose input is the fusion signal d, is used to post-process the fusion signal to obtain the target signal

[0057] (2.3) Establish a USRP control module including a sending submodule, a receiving submodule and a synchronization submodule to send and receive signals at a specified time:

[0058] The sending submodule is used to open the USRP sending session, configure the USRP's sending channel, sending antenna, carrier frequency, IQ rate and antenna gain, and write the signal to the USRP for sending;

[0059] The receiving submodule is used to open the USRP receiving session, configure the USRP receiving channel, receiving antenna, carrier frequency, IQ rate, sampling time and antenna gain, and return after receiving the signal;

[0060] The synchronization submodule is used to set the USRP internal timestamp to zero at the same time according to the synchronization signal and the system time, and to set the trigger time of the USRP sending and receiving actions.

[0061] Step 3: deploy all submodules in the quasi-neural network module, the air computing module, and the USRP control module to each host computer for subsequent signal processing and transmission and reception.

[0062] Step 4: Initialize the quasi-neural network parameters of each node.

[0063] (4.1) Assume that the pseudo-neural network parameters of each node include: source node processing weight parameter u, i-th relay node processing weight parameter v i , the destination node processes the weight parameter w, the learning rate lr, the error threshold λ, the training round e=0, and at the same time initializes the USRP timestamp to 0, sets the time slot length T, and the random number generator seed seed, i=1,2;

[0064] In this example, the source node processing weight parameter u and the i-th relay node processing weight parameter v are set i , the destination node processing weight parameter w is a random complex number, the learning rate lr = 0.05, and the error threshold λ = 0.01;

[0065] (4.2) Set the IP address of USRP in the host computer of each node, start the sending session and receiving session of each USRP, and the host computer obtains Beijing time through the Internet. Each node initializes the internal timestamp of USRP to 0 at the same specified time and sets the time slot length T = 1s.

[0066] Step 5: Process the training symbol p generated by the source node to obtain the baseband signal y of the source node and send it.

[0067] (5.1) The source node sets the seed of the random number generator to seed, generates a random 01 sequence, modulates it through the QPSK modulator, and obtains the complex training symbol p;

[0068] (5.2) The source node calls the source signal processing submodule to regulate and perform nonlinear processing on the training symbol p to obtain the initial signal: Where, s = up;

[0069] (5.3) The source node calls the sending submodule to set the USRP hardware parameters, including the sending channel 0, the sending antenna TX1, and the carrier frequency f 1 =1.8GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, transmission time 1+4eT, control USRP to transmit at frequency f at time 1+4eT 1 The baseband signal y is sent.

[0070] Step 6: The i-th relay node receives the baseband signal y from the source node through the wireless channel and obtains the relay signal r i , and process it to obtain the baseband signal c of the relay node i .

[0071] (6.1) The i-th relay node calls the receiving submodule to set the USRP hardware parameters, including receiving channel 1, receiving antenna RX1, carrier frequency f 1 =1.8GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, sampling time t s =100ms, receiving time 1+4eT, control USRP to receive at frequency f at time 1+4eT 1 Receive the source node baseband signal y and get the relay signal r i :

[0072] r i =H i y+N i ,

[0073] Among them, H i represents the channel coefficient from the source node to the i-th relay node, N i represents the noise from the source node to the i-th relay node;

[0074] (6.2) The i-th relay node calls the relay signal processing submodule to process the relay signal r iControl and perform nonlinear processing to obtain a nonlinear transformation signal: Among them, a i =v i H r i , H represents the conjugate transpose operation;

[0075] (6.3) The i-th relay node calls the air superposition preprocessing submodule to select the air calculation operator type for the nonlinear transformation signal b i Perform superposition preprocessing of air calculation to obtain the baseband signal of the relay node: Among them, k i is a real number between 0 and 1,

[0076] (6.4) The i-th relay node calls the sending submodule to set the USRP hardware parameters, including the sending channel 0, the sending antenna TX1, and the carrier frequency f 2 =2.0GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, transmission time 2+4eT, control USRP to relay node baseband signal c i At time 2+4eT with frequency f 2 Send it;

[0077] Step 7: The destination node receives the baseband signal c from the relay node via the wireless channel. i The fused signal d is obtained and processed to obtain the predicted symbol x.

[0078] (7.1) The destination node calls the receiving submodule to set the USRP hardware parameters, including receiving channel 1, receiving antenna RX1, carrier frequency f 2 =2.0GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, sampling time t s =100ms, receiving time 2+4eT, control USRP to receive at frequency f at time 2+4eT 2 Receive the relay node baseband signal c after superimposing the wireless channel i , and get the fused signal d:

[0079]

[0080] Among them, H i+n represents the channel coefficient from the i-th relay node to the destination node, N i+n represents the noise from the i-th relay node to the destination node;

[0081] (7.2) The destination node calls the air superposition post-processing submodule to select the air calculation operator type corresponding to the relay node, and performs air calculation superposition post-processing on the fusion signal d to obtain the destination signal:

[0082] (7.3) The destination node calls the destination signal processing submodule to regulate the destination signal z to obtain the predicted signal x = w H z, where H represents the conjugate transpose operation.

[0083] Step 8, calculate the error J between the predicted symbol and the training symbol, broadcast it to the source node and the relay node, and judge the error J.

[0084] (8.1) The destination node sets the seed of the random number generator, generates a random 01 sequence, modulates it through the QPSK modulator to obtain a complex training symbol p, and then calculates the error J between the predicted signal x and the training symbol p:

[0085] J=|xp| 2 ;

[0086] (8.2) The destination node calls the sending submodule to set the USRP hardware parameters, including the sending channel 0, receiving antenna RX1, and carrier frequency f 3 =2.2GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, sampling time t s =100ms, sending time 2.5+4eT, control USRP to send at frequency f at time 1+4eT 3 Broadcast the error signal J;

[0087] (8.3) The source node and relay node call the receiving submodule to set the USRP hardware parameters, including receiving channel 1, receiving antenna RX1, carrier frequency f 3 =2.2GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, sampling time t s =100ms, receiving time 2.5+4eT, control USRP to receive at frequency f at time 2.5+4eT 3 The error signal J is received.

[0088] (8.4) Compare the error signal J with the set error threshold λ:

[0089] If J>λ, then execute step 9.

[0090] Otherwise, go to step 12.

[0091] Step 9: Update the destination node processing weight parameter w, calculate and send the error gradient of the destination signal z.

[0092] (9.1) The destination node calculates and calls the destination parameter update submodule to update the processing weight parameter w to obtain the updated processing weight parameter: in, * indicates the conjugation operation;

[0093] (9.2) The destination node calls the destination parameter update submodule to calculate the error gradient of the fusion signal z

[0094]

[0095] (9.3) The destination node calls the sending submodule to set the USRP hardware parameters, including the sending channel 0, the sending antenna TX1, and the carrier frequency f 2 =2.0GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, transmission time 3+4eT, control USRP to fuse the error gradient of signal z At time 3+4eT with frequency f 2 to send.

[0096] Step 10: The i-th relay node uses the target gradient signal Update the processing weight parameter v of the i-th relay node i , calculate and send the relay signal r i The error gradient.

[0097] (10.1) The i-th relay node calls the receiving submodule to set the USRP hardware parameters, including receiving channel 1, receiving antenna RX1, carrier frequency f 2 =2.0GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, sampling time t s =100ms, receiving time 3+4eT, control USRP to receive at frequency f at time 3+4eT 2 Receive the error gradient of the fusion signal sent by the destination node Get the target gradient signal

[0098]

[0099] Among them, H i+n represents the channel coefficient from the i-th relay node to the destination node, η i+n represents the noise from the destination node to the i-th relay node, and T represents the transposition operation;

[0100] (10.2) The i-th relay node calls the relay parameter update submodule to process the weight parameter v i Update and obtain the updated processing weight parameters: in, * indicates a conjugate operation, and H indicates a conjugate transpose operation;

[0101] (10.3) The i-th relay node calls the relay parameter update submodule to calculate the relay signal r i The error gradient

[0102]

[0103] (10.4) The i-th relay node calls the sending submodule to set the USRP hardware parameters, including the sending channel 0, the sending antenna TX1, and the carrier frequency f 1 =1.8GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, transmission time 4+4eT, control USRP to relay signal r i The error gradient At time 4+4eT with frequency f 1 to send.

[0104] Step 11: The source node updates the source node processing weight parameter u using the relay gradient signal φ.

[0105] (11.1) The source node calls the receiving submodule to set the USRP hardware parameters, including receiving channel 1, receiving antenna RX1, carrier frequency f 1 =1.8GHz, IQ sampling rate f s =1M, antenna gain gain = 20dBi, sampling time t s =100ms, receiving time 4+4eT, control USRP to receive at frequency f at time 4+4eT 1 The error gradient of the relay signal sent by the receiving relay node after being superimposed in the air Get the relay gradient signal φ:

[0106]

[0107] Among them, H i represents the channel coefficient from the source node to the i-th relay node, η i represents the noise from the i-th relay node to the source node, and T represents the transposition operation;

[0108] (11.2) The source node calls the source parameter update submodule to update the processing weight parameter u and obtains the updated processing weight parameter: in, * indicates the conjugation operation;

[0109] (11.3) Update the training round e=e+1 and return to step 5.

[0110] Step 12: The source node, relay node, and destination node stop the reverse propagation process.

[0111] The source node stops the receiving process, the two relay nodes stop the reverse receiving and sending processes, and the destination node stops the sending process;

[0112] Step 13, according to the current processing weight parameters of each node, the source node sends the information to be transmitted, the relay node receives and forwards it to the destination node, and the destination node receives and obtains the information to be transmitted.

[0113] (13.1) The source node, relay node, and destination node reinitialize the USRP timestamp to 0;

[0114] (13.2) The source node regulates and processes the information to be transmitted according to the current processing weight parameter u and sends it;

[0115] (13.3) The i-th relay node processes the weight parameter v according to the current i Control and process the received signal sent by the source node and send it;

[0116] (13.4) The destination node regulates and processes the signal sent by the received relay node according to the current weight parameter w, and recovers the information to be transmitted.

[0117] At this point, the quasi-neural network transmission based on air computing is completed.

[0118] The effect of the present invention can be further illustrated by the following simulation experiment.

[0119] 1. Simulation Experiment Conditions

[0120] Four USRPs and their host computers were used to build a quasi-neural network transmission platform based on air computing with a source node, two relay nodes, and a destination node. The quasi-neural network module, air computing module, and USRP control module were built and deployed in each node. In addition, two USRPs and their host computers were used as interference nodes. The interference nodes used a partial band interference model, and set the interference signal bandwidth to 20%, 40%, and 60% of the transmission signal bandwidth to interfere with the transmission channel. The information to be transmitted is a 66KB picture. It is transmitted 1000 times. If the bit error rate of the restored picture is below 0.05%, it is considered a successful transmission.

[0121] Set the communication frequency f between the source and relay nodes 1 =1.8GHz, the communication frequency between the relay node and the destination node f2 =2.0GHz, quasi-neural network learning rate lr=0.05, error threshold λ=0.01, time slot length T=1s.

[0122] 2. Simulation content and results

[0123] Simulation 1: Under the above simulation experimental conditions, the method of the present invention and the traditional multi-path transmission method are used to transmit the information to be transmitted, and the success transmission probability of the two methods is statistically analyzed. The results are as follows: Figure 3 The horizontal axis is the interference degree, and the vertical axis is the probability of successful image transmission.

[0124] from Figure 3 It can be seen that under interference levels of 20%, 40% and 60%, the probabilities of successful image transmission of the present invention are 93%, 82% and 66% respectively, and the probabilities of successful image transmission of the multi-path transmission method are 87%, 72% and 57% respectively. The probabilities of successful image transmission when the method of the present invention performs information transmission are higher than those of the multi-path transmission method, indicating that the present invention can reduce the bit error rate in the end-to-end signal transmission process of the self-organizing network.

[0125] Simulation 2: Under the above simulation experimental conditions, the method of the present invention and the traditional multi-path transmission method are used to transmit the information to be transmitted, and the time slot consumption of the two methods is counted. The results are as follows: Figure 4 The horizontal axis is the number of service transmissions, and the vertical axis is the number of time slots consumed.

[0126] from Figure 4 It can be seen that compared with the traditional method, the present invention reduces the time slot consumption in the end-to-end transmission process. When the service is transmitted 1000 times, fewer time slots can be spent to complete the service transmission, indicating that the present invention can improve the information transmission rate.

[0127] The above description is only a specific example of the present invention and does not constitute any limitation to the invention. It is obvious that for professionals in this field, after understanding the content and principles of the present invention, they may make various modifications and changes in form and details without departing from the principles and structures of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

[0128] It should be noted that the step numbers in the specification and claims of the present invention are only for a clear description of the implementation scheme of the present invention to facilitate understanding, and the order of the numbers is not limited.

Claims

1. A quasi-neural network transmission method based on air computing, characterized in that: The steps include: (1) Use the universal software radio peripheral USRP and the host computer to build an air computing quasi-neural network transmission platform including a source node S, n relay nodes R, and a destination node D, and synchronize the time of each USRP through a clock synchronizer; (2) Construct the quasi-neural network module, air computing module and USRP control module: 2a) establishing a quasi-neural network module including a source signal processing submodule, a relay signal processing submodule, a destination signal processing submodule, a source parameter updating submodule, a relay parameter updating submodule and a destination parameter updating submodule, for regulating the signal and updating the processing weight parameters; 2b) establishing an air computing module including an air superposition pre-processing submodule and an air superposition post-processing submodule, which is used to transfer the position where the multi-path signals are fused from the receiving end to the air channel; 2c) establishing a USRP control module including a sending submodule, a receiving submodule and a synchronization submodule for sending and receiving signals at a specified time; (3) Deploy all submodules in the quasi-neural network module, the air computing module, and the USRP control module to each host computer; (4) Initialize the pseudo-neural network parameters of each node, including the source node processing weight parameter u and the i-th relay node processing weight parameter v i , the destination node processes the weight parameter w, the learning rate lr, the error threshold λ, the training round e=0, and at the same time initializes the USRP timestamp to 0, sets the time slot length T, and the random number generator seed seed, i=1,...,n; (5) Use the training symbol forward transmission and error gradient back propagation to update the processing weight parameters of each node: (5a) Process the training symbol p generated by the source node to obtain the baseband signal y of the source node and send it; (5b) The i-th relay node receives the baseband signal y from the source node through the wireless channel and obtains the relay signal r i And process it to get the baseband signal c of the relay node i and send; (5c) The destination node receives the baseband signal c from the relay node after the wireless channel is superimposed i , get the fused signal d, and process it to get the predicted symbol x; (5d) Calculate the error J between the predicted symbol x and the training symbol p, broadcast it to the source node and the relay node, and compare the error signal J with the set error threshold λ: If J>λ, then execute step (5e), Otherwise, execute step (6); (5e) Update the destination node processing weight parameter w, calculate and send the error gradient of the destination signal z (5f) The i-th relay node receives the error gradient of the target signal and obtains the target gradient signal Using the target gradient signal Update the processing weight parameter v of the i-th relay node i , and the relay signal r i The error gradient Send it; (5g) The source node receives the relay signal error gradient, obtains the relay gradient signal φ, uses the relay gradient signal φ to update the source node processing weight parameter u, updates the training round e=e+1, and returns to step (5a); (6) The source node, relay node, and destination node stop the reverse propagation process; (7) According to the current processing weight parameters of each node, the source node sends the information to be transmitted, the relay node receives and forwards it to the destination node, and the destination node receives and restores the information to be transmitted.

2. The method according to claim 1, characterized in that Step (1) Use the universal software radio peripheral USRP and the host computer to build an air computing quasi-neural network transmission platform including a source node S, n relay nodes R, and a destination node D, and implement it as follows: Use n+2 USRPs and their host computers, select one USRP and its host computer as the source node, n USRPs and their host computers as relay nodes, and one USRP and its host computer as the destination node, and connect each USRP to the corresponding host computer via a Gigabit Ethernet cable or a PCIe high-speed cable; Install the USRP driver in the host computer, and install the ANT antenna on the TX1 interface of channel 0 and the RX1 interface of channel 1 of each USRP; The clock synchronizer inputs the synchronization signals PPS and REF to each USRP through SMA cables.

3. The method according to claim 1, characterized in that Step 2a) uses the quasi-neural network module to regulate the signal and update the processing weight parameters, which is performed using each submodule in the quasi-neural network module, wherein: The source signal processing submodule, whose input is the training symbol p, is used to adjust p according to the processing weight parameter u to obtain the initial signal: s = up, and perform nonlinear processing on the signal s through the nonlinear transformation function σ, and output the baseband signal to be sent by the source node: The relay signal processing submodule, whose input is the relay signal r i , used to process the weight parameter v i R i Control and obtain the transfer signal: a i =v i H r i , and transforms the signal a through a nonlinear transformation function σ i Perform nonlinear processing and output nonlinear transformation signal: Where H represents the conjugate transpose operation; The target signal processing submodule, whose input is the target signal z, is used to adjust z according to the processing weight parameter w to obtain the predicted signal x=w H z, where H represents the conjugate transpose operation; The source parameter update submodule, whose input is the error J, the processing weight parameter u, the training symbol p, the initial signal s, the baseband signal y to be sent by the source node, the relay gradient signal φ and the pseudo-neural network learning rate lr, is used to calculate the error gradient of the conjugate of the processing weight parameter u: And update the source node processing weight parameter according to the error gradient conjugate of the processing weight parameter u: in, * indicates the conjugation operation; The relay parameter update submodule has the error J and the processing weight parameter v as input. i , relay signal r i , transfer signal a i , nonlinear transformation signal b i , target gradient signal and the quasi-neural network learning rate lr, which is used to calculate the processing weight parameter v i Conjugate error gradient: Calculate the relay signal r i The error gradient is: And according to the processing weight parameter v i The conjugate error gradient updates the weight parameter v i ′: in The target parameter update submodule, whose input is the error J, the processing weight parameter w, the target signal z, the prediction signal x, the training symbol p and the pseudo-neural network learning rate lr, is used to calculate the error gradient of the conjugate of the processing weight parameter w: Calculate the error gradient of the target signal z: Update the processing weight parameter according to the error gradient conjugated with the processing weight parameter w: in, 4. The method according to claim 1, characterized in that: Step 2b) uses the air computing module to transfer the location where the multi-path signals are merged from the receiving end to the air channel, which is performed by using each submodule in the air computing module, wherein: The air superposition preprocessing submodule, whose input is the nonlinear transformation signal b i , used to preprocess the nonlinear transformation signal to obtain the baseband signal to be sent by the relay node in The aerial superposition post-processing submodule, whose input is the fusion signal d, is used to post-process the fusion signal to obtain the target signal 5. The method according to claim 1, characterized in that In step 2c), the USRP control module is used to send and receive signals at a specified time, which is performed by using each submodule in the USRP control module, wherein: The sending submodule is used to open the USRP sending session, configure the USRP's sending channel, sending antenna, carrier frequency, IQ rate and antenna gain, and write the signal to the USRP for sending; The receiving submodule is used to open the USRP receiving session, configure the USRP receiving channel, receiving antenna, carrier frequency, IQ rate, sampling time and antenna gain, and return after receiving the signal; The synchronization submodule is used to set the USRP internal timestamp to zero at the same time according to the synchronization signal and system time, and to set the trigger time of the USRP sending and receiving actions.

6. The method according to claim 1, characterized in that: Step (5a) processes the training symbol p generated by the source node to obtain the baseband signal y of the source node and sends it; the implementation is as follows: (5a1) The source node sets the seed of the random number generator to seed, generates a random 01 sequence, and modulates it through a QPSK modulator to obtain a complex training symbol p; (5a2) The source node adjusts p according to the processing weight parameter u to obtain the initial signal: s = up; (5a3) The source node performs nonlinear processing on the initial signal s through the nonlinear transformation function σ to obtain the baseband signal of the source node: (5a4) The source node sets the carrier frequency f1 and the transmission time 1+4eT, and controls the USRP to transmit the source node baseband signal y.

7. The method according to claim 1, characterized in that: In step (5b), the i-th relay node receives the source node baseband signal r through the wireless channel i Processing is performed to obtain the baseband signal c of the relay node i And send it, implemented as follows: (5b1) The i-th relay node processes the weight parameter v i For the relay signal r i Control and obtain the transfer signal a i =v i H r i , where r i =H i y+N i , H i represents the channel coefficient from the source node to the i-th relay node, N i represents the noise from the source node to the i-th relay node, and H represents the conjugate transpose operation; (5b2) The i-th relay node uses the nonlinear transformation function σ to transfer the signal a i Perform nonlinear processing to obtain a nonlinear transformation signal: (5b3) The i-th relay node selects the air computing operator type to perform the nonlinear transformation on the signal b i Perform superposition preprocessing of air calculation to obtain the baseband signal of the relay node: Among them, k i is a real number between 0 and 1, (5b4) The i-th relay node sets the carrier frequency f2 and the transmission time 2+4eT, and controls the USRP to transmit the relay node baseband signal c i to send.

8. The method according to claim 1, characterized in that: Step (5c) The destination node receives the baseband signal of the relay node through the wireless channel to obtain the fused signal d, and processes it to obtain the predicted symbol x, which is implemented as follows: (5c1) The destination node sets the carrier frequency f2 and the receiving time 2+4eT, and controls the USRP to receive the fused signal sent by n relay nodes after being superimposed in the air: Among them, H i+n represents the channel coefficient from the i-th relay node to the destination node, N i+n represents the noise from the i-th relay node to the destination node; (5c2) The destination node selects the air computing operator type corresponding to the relay node, performs air computing superposition and post-processing on the fused signal d to obtain the destination signal: Where n is the number of relay nodes; (5c3) The destination node adjusts the destination signal z according to the processing weight parameter w to obtain the predicted signal x=w H z, where H represents the conjugate transpose operation.

9. The method according to claim 1, characterized in that: In the step (5d), the error J between the predicted symbol and the training symbol is calculated by firstly setting the seed of the random number generator at the destination node to generate a random 01 sequence, modulating it through a QPSK modulator to obtain a complex training symbol p, and then calculating the error between the predicted signal x and the training symbol p: J = |xp| 2 ; In the step (5d), it is broadcast to the source node and the relay node, and the destination node sets the carrier frequency f3 and the sending time 2.5+4eT, controls the USRP to send the error J, and the source node and the relay node set the carrier frequency f3 and the receiving time 2.5+4eT, and controls the USRP to receive and obtain the error J.

10. The method according to claim 1, characterized in that: In step (5e), ​​the destination node processing weight parameter w is updated, and the error gradient of the destination signal z is calculated and sent. The implementation is as follows: (5e1) The destination node calculates the error gradient of the conjugate weight parameter w: Where * represents the conjugation operation; (5e2) The destination node processes the error gradient conjugated with the weight parameter w Update the processing weight parameter w to obtain the updated processing weight parameter: (5e3) The destination node calculates the error gradient of the fused signal z: (5e4) The destination node sets the carrier frequency f2 and the transmission time 3+4eT, and controls the USRP to fuse the error gradient of the signal z to send.

11. The method according to claim 1, characterized in that: In step (5f), the i-th relay node uses the target gradient signal Update the processing weight parameter v of the i-th relay node i , and the relay signal r i The error gradient Sending is implemented as follows: (5f1) The i-th relay node calculates the processing weight parameter v i Conjugate error gradient: in, H i+n represents the channel coefficient from the i-th relay node to the destination node, η i+n represents the noise from the destination node to the i-th relay node, * represents the conjugate operation, and T represents the transpose operation; (5f2) The i-th relay node processes the weight parameter v i The conjugate error gradient is used to process the weight parameter v i Update to get the updated processing weight parameters: (5f3) The i-th relay node calculates the relay signal r i The error gradient is: (5f4) The i-th relay node sets the carrier frequency f1 and the transmission time 4+4eT, and controls the USRP to relay the signal r i The error gradient is sent.

12. The method according to claim 1, characterized in that: In step (5h), the relay gradient signal φ is used to update the source node processing weight parameter u, which is implemented as follows: (5h1) The source node calculates the error gradient of the weight parameter u: in, H i represents the channel coefficient from the source node to the i-th relay node, η i represents the noise from the i-th relay node to the source node, * represents the conjugate operation, and T represents the transpose operation; (5h2) The source node updates the processing weight parameter u according to the error gradient of the processing weight parameter u to obtain the updated processing weight parameter:

Images