Self-organizing network quasi-neural channel coding and decoding method

Abstract

The invention discloses a pseudo-neural channel coding and decoding method for a self-organizing network. The method mainly solves the problems that in the prior art, multipath fading is serious, node control is difficult, and channel coding and decoding complexity is high. The scheme comprises the following steps: 1) constructing a multi-hop self-organizing pseudo-neural network by a front end node, three relay nodes and a control end node, and embedding a space-time coding module in the network; 2) inputting a pilot image to a network front-end node, performing signal forward transmission and gradient reverse update training on the quasi-neural network, and dynamically updating weight parameters of the node; 3) loading the updated node weight parameters to the pseudo-neural network to obtain a trained network; and 4) using the trained network to realize image end-to-end transmission. According to the invention, the training process of the quasi-neural network is adopted to replace the maximum likelihood decoding process of space-time block coding, so that the decoding error can be reduced and the reliability of image transmission can be improved while the design complexity of the decoding module of the receiving end can be effectively reduced.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and further relates to a coding method, specifically a self-organizing network pseudo-neural channel coding method, which can be used for relay-assisted multi-hop self-organizing networks to perform end-to-end transmission of pseudo-neural image information. Background Art

[0002] As 6G utilizes higher frequency bands such as terahertz, the wavelength used for signal transmission becomes shorter, and transmission will be more easily blocked by obstacles, especially in urban environments, where high-rise buildings and other obstacles will significantly weaken the strength of the transmitted signal. At the same time, the reflection, scattering and diffraction phenomena generated during signal transmission will also cause multipath fading, affecting the transmission quality of the signal and the reliability of communication. Therefore, extending the network coverage and improving the reliability of information transmission are urgent issues to be solved in the development of 6G and future networks. Compared with high-cost solutions such as large-scale MIMO systems and smart metasurface technology to extend network coverage, relay-assisted multi-hop transmission is an efficient and low-cost strategy to solve the 6G coverage problem. By deploying relay nodes between the transceiver and the transmitter, large obstacles in the propagation path are avoided, multipath fading is reduced, and the transmission stability of the signal is improved. With the increase in the number of nodes and antennas, it is necessary to efficiently organize the nodes and control the network to realize flexible networking of the multi-hop transmission system and dynamic update of the topology structure. In addition, increasing the number of antennas at each transmission node can further improve the spatial diversity gain of the system, thereby improving transmission reliability.

[0003] At present, there has been in-depth research on the way of using multi-hop transmission networks for information transmission. However, as the scale of the network expands, the interference problem in the cooperative communication network becomes increasingly prominent, and the difficulty of node control is also increasing, which affects the quality of information transmission. Therefore, applying diversity technology in multi-hop transmission networks to further improve transmission reliability is an important means to solve the above problems. Diversity technology mainly includes spatial diversity, frequency diversity and time diversity. Among them, space-time coding is a technology that simultaneously encodes in time and space dimensions to obtain diversity gain, which can achieve higher transmission rates and stronger anti-fading capabilities in multi-antenna systems. Common space-time coding technologies include space-time trellis codes and space-time block codes. Space-time trellis codes can obtain the maximum diversity gain and coding gain, but because Viterbi decoding is used at the receiving end, its decoding complexity increases exponentially with the increase of the number of antennas and the number of trellis code states, making it difficult to promote and use; space-time block codes can obtain the maximum diversity gain while constructing space-time codes through code orthogonal design and combining them with maximum likelihood decoding at the receiving end, so that the encoding and decoding algorithm has low complexity and good anti-interference performance.

[0004] In their paper “Cooperative Multi-hop Transmission in Wireless Networks”, Patrick Herhold and other authors proposed a cooperative relay scheme that revolutionized the concept of traditional relays and optimized multi-hop transmission in wireless networks through collaboration between multiple relay nodes. However, in the transmission process of this scheme, the source node starts to send the signal, and the relay node can adopt a variety of strategies to process the signal after receiving the signal. Although this cooperative relay scheme has improved transmission efficiency and reliability, it requires collaboration between multiple relay nodes to transmit the signal. This increases spatial diversity while also causing the signal to experience more reflections, scattering, and diffraction during propagation, generating more multipath signals, affecting the quality of signal transmission, and increasing the end-to-end delay of the system.

[0005] In their paper “Distributed Learning for MIMO Relay Networks”, Wang Rui et al. proposed a distributed learning-based relay beamforming algorithm DLRB to dynamically optimize the node weight parameters in the communication network to reduce information transmission errors. However, the DLRB algorithm assumes that there is no cooperative communication between relay nodes, that is, each relay node is only responsible for signal reception and forwarding without in-depth signal processing. Therefore, the algorithm cannot fully utilize the potential synergy between multiple relay nodes to further enhance the quality and reliability of signal transmission. In addition, due to the lack of direct communication and cooperation between relay nodes, these nodes may interfere with each other when forwarding signals at the same time. This mutual interference is not fully considered and suppressed, which may reduce the performance and transmission efficiency of the overall system.

[0006] Siavash M. Alamouti proposed a space-time coding scheme for improving the diversity gain of wireless communication systems in his paper “A Simple Transmit Diversity Technique for Wireless Communications”. While obtaining the maximum diversity gain, this scheme constructs space-time codes through code orthogonal design, which reduces the complexity of the coding algorithm. Although it can improve the anti-interference ability of signal transmission and the accuracy of decoding to a certain extent, when decoding at the receiving end, it is necessary to be based on the premise that the channel information is completely known. Therefore, accurate channel estimation is required before decoding to ensure the decoding accuracy. Accurate channel estimation is often difficult to achieve. In the presence of channel estimation errors, the decoding process of space-time block coding may be affected, resulting in an increase in the bit error rate. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a self-organizing network quasi-neural channel coding method to solve the problems of severe multipath fading, difficult node control, and high complexity of channel coding in the prior art.

[0008] The technical idea for implementing the present invention is: first, when end-to-end line-of-sight transmission cannot be met or is limited, a multi-hop transmission network is constructed by setting up relay nodes between the source node and the destination node to combat signal fading; second, in order to reduce transmission errors, a multi-user multi-hop communication network is used to construct a quasi-neural network, and the processing processes of traditional relay nodes such as signal amplification and forwarding and decoding and forwarding are converted into beamforming and nonlinear transformation on the quasi-neural node, and pilot forward transmission and gradient back propagation training are performed through the neural network to continuously update the transmission weights of the quasi-neural node to minimize the error of sending and receiving signals; finally, space-time block coding is introduced on the basis of the quasi-neural network, and multiple signal copies are sent through joint coding of space and time to improve the system diversity gain to combat multipath fading, further improve the transmission quality, and reduce the transmission error caused by multipath fading. In addition, since the training objective function of the quasi-neural network is similar to the maximum likelihood decoding algorithm of space-time block coding, the present invention adopts the training process of the quasi-neural network to replace the maximum likelihood decoding process of space-time block coding to resist the decoding error caused by the channel estimation error, which not only reduces the design complexity of the decoding module at the receiving end, but also reduces the decoding error and improves the transmission reliability.

[0009] The specific steps of the present invention to achieve the above-mentioned purpose include the following:

[0010] (1) constructing a multi-hop self-organizing quasi-neural network by a front-end node layer, a relay node layer and a control end node layer, wherein the front-end node layer includes a front-end node, the relay node layer includes three fixed relay nodes, and the control end node layer includes a control center node;

[0011] (2) Inputting the original pilot image into the multi-hop self-organizing quasi-neural network, setting the maximum number of iterations to T, the error threshold to η, and the learning rate to α, the network is trained;

[0012] (3) Initialize the relevant parameters of the multi-hop self-organizing quasi-neural network and initialize the current number of iterations t=1;

[0013] (4) The front-end node receives the original pilot image signal, and performs adaptive processing and channel splitting on the original pilot image signal according to the front-end node weight parameter, and then broadcasts the processed signal to the relay node layer through each channel;

[0014] (5) The relay node layer receives the pilot image signals of each channel from the front-end node, performs adaptive processing according to the relay node weight parameters, performs space-time block coding on the processed signals, and broadcasts them to the control center node through multiple time slots;

[0015] (6) The control center node receives the coded signals from each time slot, performs signal merging processing on all the coded signals, and performs adaptive processing and channel synthesis operations on the merged signals according to the weight parameters of the control center node, and restores the received coded signals to image signals, that is, obtains received image signals;

[0016] (7) Calculating the transmission error of the original pilot image signal and the received image signal according to the pixels at the corresponding positions, and determining whether the transmission error is less than a preset error threshold; if so, directly executing step (9); otherwise, continuing to execute step (8);

[0017] (8) Back-propagate the gradient error from the control center node to the relay node layer to the front-end node, update the weight parameters at each node, and determine whether the current number of iterations t is equal to T; if so, execute step (9); otherwise, add 1 to t and return to execute step (4);

[0018] (9) Load the current neural network weight parameters as the transmission weights of the multi-hop self-organizing quasi-neural network, and use the network to transmit the image to be transmitted.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] First, since the present invention adopts a pseudo-neural network to construct a collaborative link connecting each communication node, the processing processes of traditional relay nodes such as signal amplification and forwarding and decoding and forwarding are converted into beamforming and nonlinear transformation on the pseudo-neural node, and pilot forward transmission and gradient back propagation training are performed through the neural network to dynamically adjust the weight parameters of the transmission path. Compared with the traditional relay-assisted multi-hop network, the relay node of the present invention has the ability of collaborative communication and can minimize the error of sending and receiving signals.

[0021] Second, based on the DLRB algorithm, the present invention introduces space-time coding technology to provide diversity gain, fully utilizes the transmission potential of multi-user multi-hop communication networks, uses its extended spatial dimension to jointly encode time and space, and selects space-time block coding with low decoding complexity to establish correlation between the transmission content of different antennas, provide diversity and coding gain for the receiving end, and improve the reliability of information transmission.

[0022] Third, the present invention combines space-time block coding with a quasi-neural network architecture, which not only improves network coverage, but also introduces diversity gain to ensure signal transmission quality. In addition, traditional space-time block coding uses maximum likelihood decoding, and maximum likelihood decoding technology requires known and accurate channel information, so the accuracy of channel estimation will directly affect the decoding accuracy; since the loss function of the quasi-neural network is similar to the basic principle of maximum likelihood decoding, the present invention uses a quasi-neural network training process to replace the maximum likelihood decoding process, which reduces the decoding error while ensuring signal transmission quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart for realizing the present invention;

[0024] Figure 2 It is a schematic diagram of the application process of the present invention;

[0025] Figure 3 This is a flow chart of encoding and decoding training in the present invention;

[0026] Figure 4 The figure is a comparison curve diagram of the image transmission performance simulation using the method of the present invention and the existing method. DETAILED DESCRIPTION

[0027] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0028] Example 1: Reference Figure 1 The present invention proposes a self-organizing network pseudo-neural channel encoding and decoding method, which deploys multiple relay nodes between the transmitting end and the receiving end to form a relay-assisted multi-hop transmission network and uses a neural network to characterize it. The communication units in the multi-hop network are characterized as neuron nodes in the pseudo-neural network, and the topological structure of the pseudo-neural network is constructed according to its specific distribution. By inputting a pilot image, the constructed pseudo-network is trained by information forward transmission and error gradient back propagation until the error function is less than the set threshold or the set number of training rounds is completed. After the training is completed, the trained network parameters are loaded and image transmission is performed. It is implemented specifically according to the following steps:

[0029] Step 1. construct a multi-hop self-organizing quasi-neural network by a front-end node layer, a relay node layer and a control end node layer, wherein the front-end node layer contains a front-end node, the relay node layer contains three fixed relay nodes, and the control end node layer contains a control center node;

[0030] Step 2. Input the original pilot image into the multi-hop self-organizing quasi-neural network, set the maximum number of iterations to T, the error threshold to η, and the learning rate to α, and train the network. The input of the original pilot image into the multi-hop self-organizing quasi-neural network specifically converts the RGB color image to be transmitted into a binary bit stream, that is, the original pilot image signal, and then inputs the signal into the front-end node.

[0031] Step 3. Initialize the relevant parameters of the multi-hop self-organizing quasi-neural network, including the front-end node weight parameter w1, the relay node weight parameter w 21 ,w 22 ,w 23 , control the central node weight parameter w3; at the same time, initialize the current iteration number t = 1;

[0032] Step 4. The front-end node receives the original pilot image signal, and performs adaptive processing and channel splitting on the original pilot image signal according to the front-end node weight parameter w1, and then broadcasts the processed signal to the relay node layer through each channel. In this embodiment, the front-end node performs adaptive signal processing of beamforming and nonlinear transformation on the signal, and then splits the image information according to the R channel, G channel and B channel, and transmits the image information of each channel through three transmission paths respectively.

[0033] Step 5. The relay node layer receives the pilot image signals of each channel from the front-end node and calculates the pilot image signals according to the relay node weight parameter w 21 ,w 22 ,w 23 Adaptive processing is performed, space-time block coding is performed on the processed signal, and the signal is broadcast to the control center node through multiple time slots. In this embodiment, at the relay node, three relay nodes respectively perform the same adaptive signal processing of beamforming and nonlinear transformation as step 4 on the signals transmitted from different channels, and then the correlation between the transmitted signals is constructed through space-time block coding, and the encoded signal is relayed and forwarded by the three relay nodes respectively and sent to the control end node.

[0034] Step 6. The control center node receives the coded signals from each time slot, performs signal merging processing on all coded signals, and performs adaptive processing and channel synthesis operation on the merged signal according to the control center node weight parameter w3, and restores the received coded signal to an image signal, that is, obtains a received image signal. The specific implementation process of this embodiment is: the control center node performs signal merging on the binary bit streams received in each time slot, and performs adaptive signal processing of beamforming and nonlinear transformation on them to obtain the transformed binary bit stream, and converts it back to a single-channel image for channel synthesis to obtain the transmitted RGB image.

[0035] Step 7. Calculate the transmission error of the original pilot image signal and the received image signal according to the pixels at the corresponding positions, and determine whether the transmission error is less than the preset error threshold; if so, directly execute step 9, otherwise, continue to execute step 8.

[0036] The transmission error is specifically calculated according to the following formula:

[0037]

[0038] Among them, s is the original pilot image signal, To receive image signals.

[0039] Step 8. Back-propagate the gradient error from the control center node through the relay node layer to the front-end node, update the weight parameters at each node, and determine whether the current number of iterations t is equal to T; if so, execute step 9; otherwise, add 1 to t and return to step 4.

[0040] The back propagation of the gradient error from the control center node to the front-end node via the relay node layer is specifically implemented by using a neural network back propagation gradient descent method. The process first calculates the error term and the gradient of the error relative to the weight parameter of the control center node at the control center node, updates the weight parameter of the control center node according to the gradient, and transfers the error gradient to the relay node layer through the chain rule; repeats the same process, back propagates the error from the relay node layer to the front-end node, transfers the gradient layer by layer and updates the weight parameter.

[0041] Step 9. Load the current neural network weight parameters as the transmission weights of the multi-hop self-organizing quasi-neural network, and use the network to transmit the image to be transmitted.

[0042] Embodiment 2: The overall implementation steps of the encoding and decoding method proposed in this embodiment are the same as those in Embodiment 1. Figure 2 The overall inventive concept of the present invention is further described.

[0043] The present invention utilizes space-time block coding and adopts a multi-hop self-organizing network to construct a neural network, thereby realizing end-to-end image transmission. The transmission process is as follows: Figure 2 shown. Figure 2 The upper part is a schematic diagram of a multi-hop self-organizing network for image transmission. The quasi-neural network constructed using the multi-hop self-organizing network is shown in Figure 2As shown in the lower part. In the present invention, a communication node of a multi-hop self-organizing network is characterized as a neuron in a quasi-neural network, and each communication node contains a transmission antenna. The communication link between communication nodes is characterized as a weight connection between neurons, and the signal processing at the communication node is characterized as beamforming and nonlinear transformation in the quasi-neural network. The quasi-neural network constructed by the scheme includes a front-end node layer, a relay node layer and a control end node layer. The front-end node layer contains a front-end node, the relay node layer contains three fixed relay nodes, and the control end node layer contains a control center. Before transmission, a pilot image signal is sent to perform signal forward transmission and gradient reverse update training on the quasi-neural network, dynamically update the node weight parameters, and achieve minimum error transmission. This scheme will load the trained node weight parameters and perform end-to-end transmission of RGB color images to ensure the quality of image transmission.

[0044] Specifically, the present invention converts the RGB color image to be transmitted into a binary bit stream and inputs it into the front-end node, the front-end node performs adaptive signal processing of beamforming and nonlinear transformation on the signal, and then splits the image information according to the R channel, the G channel and the B channel, and transmits the image information of each channel through three transmission paths respectively; at the relay node, the three relay nodes respectively perform similar adaptive signal processing on the three signals, and then construct the correlation between the transmission signals through space-time group coding, and the encoded signals are relayed and forwarded by three relay nodes through four time slots respectively; the control center performs signal merging and adaptive signal processing on the binary bit streams received in the four time slots to obtain the transformed binary bit stream, and converts it back to a single-channel image for channel synthesis to obtain the transmitted RGB image. The present invention designs a "three-transmit and one-receive" space-time group coding scheme between the relay layer and the control center, which is suitable for a system in which three relay nodes send images through three transmitting antennas and a control center with one receiving antenna receives the images.

[0045] Embodiment 3: The overall implementation steps of the encoding and decoding method proposed in this embodiment are the same as those in Embodiment 1. Figure 3 The training process of the network constructed in the present invention is further described in detail.

[0046] Assume that the front-end node processes the image signal and splits it into three single-channel image information and sends it to the relay node. The relay node performs adaptive signal processing on the received signal to obtain the signals s1, s2, and s3. Before forwarding, the encoder maps them into the coded signal matrix S, as shown in the following formula (1):

[0047]

[0048] Where * represents the conjugate of the complex signal. The transmitted codeword matrix S is a mapping through the joint action of the space domain and the time domain. The row order of the matrix represents the antenna order, and the column order represents the time slot order. Collaborative diversity is achieved by three antennas sending different parts of the signal matrix S in four time slots in turn. Assume that in four time slots, the three antennas transmit signal matrices S1, S2, and S3 respectively. The first collaborative user sends the first row of the signal matrix, that is, The second collaborative user sends the second row of the signal matrix, i.e. The third collaborative user sends the third row of the signal matrix, i.e. Space-time coding requires that the signal vectors sent by the three antennas are mutually orthogonal:

[0049]

[0050] Here, H represents the transpose conjugate operation of the matrix.

[0051] The present invention combines the "three-transmit-one-receive" space-time block coding with the quasi-neural network to realize quasi-neural coding and decoding transmission, wherein a complete training process is completed by the coded forward transmission of the pilot image and the back propagation of the gradient information, and is implemented in the form of time division duplex. A complete round of training process is as follows: Figure 3 As shown in Figure 1, a complete round of training requires 8 time slots. Time slot 1 is defined to be consistent with time slots 3, 4, 5, and 6 and is T S , T S It represents the time taken to transmit a single-channel image bit stream in a single hop; the time of time slot 2 is T w (T w <T s ), T w Indicates the time taken by the relay node to send the pilot sequence for channel estimation; time slot 7 and time slot 8 are consistent and are T d (T d <T s ), T d Indicates the time taken for the error to propagate in the reverse direction within a single hop.

[0052] First, the pilot image coding is forward transmitted in time slots 1 to 6. Assume that w1, w 21 、w 22 、w 23 w2 represents the overall weight parameter matrix of the relay node layer, and f(·) represents the nonlinear activation function. r1 , H r2 , H r3 and H d1 , H d2 , H d3Respectively represent the channel state matrices from the front-end node to the relay and from the relay to the control center node, η r1 , η r2 , η r3 represents the channel noise from the front-end node to the relay, η d1 , η d2 , η d3 , η d4 They represent the sum of the channel noise in the four transmission time slots from the relay channel to the control center.

[0053] The pilot image coding forward transmission process is specifically as follows:

[0054] (1a) In time slot 1 (such as Figure 3 As shown in ①, the front-end node performs adaptive processing of the pilot image and transmits it forward to the three relay nodes in different channels. The RGB color image used for training is converted into a binary bit stream s for pilot. After the front-end node obtains the pilot signal s, it performs adaptive signal processing. First, beamforming is performed according to the weight parameter matrix w1 to obtain z=w1s, and then a nonlinear transformation is performed to obtain the signal x=f(w1s). The processed sequence x is split into R, G, and B single-channel images, that is, x=[x1,x2,x3], which are broadcasted in different channels and transmitted forward to the three relays. Relay 1, Relay 2, and Relay 3 are respectively transmitted through channel H r1 , H r2 , H r3 Receive the complex sequence x1, x2, x3 transmitted from the front-end node and obtain received signals r1, r2, r3.

[0055] (1b) In time slot 2 (such as Figure 3 As shown in ②, the three relay nodes send the newly generated pilot sequences to the control center to estimate the transmission channel. When the channel state information between the control center and the relay is unknown, the three relay nodes need to send new pilot sequences for channel estimation so that the control center can perform channel estimation. The channel estimation method can use LS channel estimation or MMSE channel estimation.

[0056] (1c) In time slots 3 to 6 (such as Figure 3 The three relay nodes (shown in ③-⑥) process the received signals respectively, cooperate to complete the space-time block coding, and transmit different signals to the control center in different time slots. Relay 1 performs beamforming on its received signal r1 to obtain Then perform nonlinear transformation to obtain Relay 2 performs beamforming on its received signal r2 to obtain Then perform nonlinear transformation to obtain Relay 3 performs beamforming on its received signal r3 to obtain Then perform nonlinear transformation to obtain Through the cooperation of three relay nodes, the processed signals a1, a2, and a3 are jointly space-time block coded to obtain the transmission signal matrix: The three relay nodes transmit different signals to the control center node in different time slots. Specifically, in time slot 3, relay node 1 sends signal a1, relay node 2 sends signal a2, and relay node 3 sends signal a3; in time slot 4, relay node 1 sends signal Relay node 2 sends a signal Relay node 3 does not send a signal, but only performs channel noise superposition; in time slot 5, relay node 1 sends a signal Relay node 2 does not send a signal, but only performs channel noise superposition. Relay node 3 sends a signal. In time slot 6, relay node 1 does not send a signal, but only performs channel noise superposition, and relay node 2 sends a signal Relay node 3 sends a signal The control center node receives the signal y at time slot 3 11 ,y 21 ,y 31 Splicing is y1 = [y 11 ,y 21 ,y 31 ], the signal y received in time slot 4 12 ,y 22 ,y 32 Splicing is y2 = [y 12 ,y 22 ,y 32 ], the signal y received in time slot 5 13 ,y 23 ,y 33 Splicing is y3 = [y 13 ,y 23 ,y 33 ], the signal y received in time slot 6 14 ,y 24 ,y 34 Splicing is y4 = [y 14 ,y 24 ,y 34 ].

[0057] (1d) The control center node performs signal merging. Assume that H d1 , H d2 , H d3 The channel gain h d1 、h d2 、h d3 In a certain period of time (t~t+3T s ) remains unchanged, that is n=1,2,3. Among them, |hdi | and θ di Respectively represent H d1 , H d2 , H d3 The channel gain is between t and t + 3T. s After the control center node receives the signals y1, y2, y3, and y4, it can be expressed as the superposition of signals a1, a2, and a3, that is, y1 = h d1 a1+h d2 a2+h d3 a3+η d1 , Among them, η d1 , η d2 , η d3 , η d4 Represents t to t+3T in sequence s The sum of the additive noise that affects the symbol code element at a certain moment, assuming that the noise is a Gaussian vector with zero mean and variance N0. At this time, there are signal components a1, a2, and a3 in y1, y2, y3, and y4 at the same time, so it is necessary to perform signal merging processing on y1, y2, y3, and y4 to suppress interference. Assuming that the channel state information estimate has been obtained in step (2), the merged signal is obtained through the signal combiner:

[0058]

[0059] After signal merging, There are only signal components of a1, a2, and a3 in Splice signals to form a signal matrix

[0060] (1e) The control center node processes the signal according to its own weight matrix and restores the original image. The control center node performs beamforming on the spliced ​​signal y to obtain the pilot estimation signal Binary bit stream Convert the channel into a single-channel image, and synthesize the single-channel image to restore the RGB image.

[0061] Gradient back propagation is then performed in time slots 7 and 8. Unlike forward transmission, which transmits a sequence containing image information, the back propagation only transmits the gradient information of the signal error, which is usually just a numerical vector or matrix with the same dimension as the weight matrix. Therefore, time slots 7 and 8 take time T d Less than the time T taken for forward transmission sThe gradient back propagation process is the key process of quasi-neural network training. It adopts the form of time division duplexing and utilizes channel reciprocity to achieve distributed update of the beamforming matrix of each node.

[0062] The gradient back propagation is implemented as follows:

[0063] (2a) The control center node completes the update of the weight parameter w3 and in time slot 7 (such as Figure 3 The error gradient is transmitted to the relay node (as shown in ⑦). The control center node calculates the error J and gradient information between the pilot estimation signal and the pilot signal And broadcast the signal to the relay node According to (α is the learning rate) Update the weight parameters.

[0064] (2b) Relay 1, Relay 2 and Relay 3 complete the weight parameter w respectively 21 、w 22 、w 23 Update, and in time slot 8 (such as Figure 3 The error gradient is transmitted to the front-end node (shown in Figure ⑧). The three relay nodes receive the information broadcasted in reverse by the control center node. According to the channel reciprocity, the signals received by the three relay nodes are According to the chain derivation rule, relay 1 calculates and Relay 1 broadcasts the signal to the source node Similarly, relay 2 calculates and Relay 2 broadcasts the signal to the source node Relay 3 calculates and Relay 3 broadcasts the signal to the source node The three relay nodes are respectively based on Update the beamforming matrix w 21 ,according to Update the beamforming matrix w 22 ,according to Update the beamforming matrix w 23 .

[0065] (2c) The front-end node completes the update of the weight parameter w1. Using channel reciprocity, the received signals are To splice The front-end node can obtain the error for the variable x by calculating the chain derivation rule * Gradient information And further calculated and According to Update beamforming matrix

[0066] The above training process is iterated continuously until the error function is less than the set threshold η or the set m rounds of training are completed, and then the training is terminated. The trained weight parameters are loaded, and the image to be transmitted is transmitted according to the forward transmission process.

[0067] The present invention constructs a quasi-neural network by constructing a multi-hop self-organizing network and combines it with space-time coding, embeds space-time group coding between the relay nodes of the quasi-neural network, provides diversity gain, and adopts two processes of forward transmission and error gradient back propagation to train the entire network, replacing the decoding process of traditional space-time coding to achieve minimum transmission error. Therefore, the quasi-neural network encoding and decoding transmission method of the present invention can not only use diversity gain to combat multipath fading, but also eliminate the decoding error caused by inaccurate channel estimation, so that the system transmission performance is further improved. The scheme deploys relay nodes on a long-distance transmission path with no line-of-sight transmission or limited line-of-sight transmission, adaptively processes the image components received from the transmitting end and forwards them to the receiving end to reduce the signal fading caused by long-distance transmission. In addition, the deployment of relay nodes can also be dynamically adjusted according to changes in terrain and environment to further optimize the transmission path, ensure that the image can be transmitted to the receiving end with a lower bit error rate in a complex environment, and improve the overall performance of the long-distance wireless communication system. By adopting space-time coding technology to establish the correlation between the channels of the transmitted image, make full use of time and space diversity, and solve the problem of low image transmission reliability in the form of channel coding. The present invention deploys space-time coding technology in the architecture of multi-hop transmission assisted by relay nodes, generates a virtual antenna array through the collaboration of multiple single-antenna relay nodes, fully utilizes multiple transmitting antennas participating in the transmission of image information to generate gain, and restores the original transmitted image through decoding at the receiving end, reduces the image transmission error, and improves the reliability of end-to-end image transmission. A neural network is used to control the communication nodes participating in collaborative coding and decoding during the multi-hop transmission process to solve the problem of difficult networking and node coordination in multi-user node collaborative communication. Based on the similarity between the neural network and the communication network structure, a neural network is constructed through a multi-hop self-organizing network based on space-time block coding, and the transmission weight of each node is updated by neural network training to adapt to the channel used to transmit image information between each node, thereby reducing the error of end-to-end image transmission. At the same time, through the continuous training of the neural network weight parameters, the image transmission and encoding and decoding methods are optimized, replacing the maximum likelihood decoding in space-time block coding to reduce the decoding complexity.

[0068] The effect of the present invention will be further described below in conjunction with experiments.

[0069] 1. Simulation conditions:

[0070] The simulation experiment of the present invention is carried out in the hardware environment of Intel Core i7 processor, 64-bit operating system and the software environment of Pycharm integrated development environment (IDE).

[0071] 2. Simulation content:

[0072] The simulation sets the learning rate to 0.01, the signal-to-noise ratio to 0-20dB, sets the same channel information and channel noise, and sets the training end condition to 2000 training rounds. After the training is completed, the same image is tested for transmission, the transmission error is calculated, and the performance curve of the transmission error versus the signal-to-noise ratio is plotted.

[0073] 3. Simulation results and analysis:

[0074] The simulation results are as follows Figure 4 As shown, it is a simulation comparison diagram of image transmission performance of the present invention, the traditional MIMO transmission system, and the quasi-neural network transmission system without considering space-time coding. As can be seen from the figure, the image transmission error of the present invention is lower than that of the traditional MIMO image transmission and the quasi-neural network image transmission without considering space-time coding, and the end-to-end image transmission reliability is improved.

[0075] The above experimental results prove the correctness and effectiveness of the method proposed in the present invention.

[0076] The present invention aims at the problems of limited coverage of high-frequency band transmission signals in 6G networks, significant multipath fading, and difficulty in node collaboration and control in large-scale networks, and proposes a self-organizing network quasi-neural channel coding method to achieve efficient image transmission. First, the method adopts a multi-user multi-hop communication network with relay assistance, and sets a relay station to ensure the transmission quality of the end-to-end long-distance transmission system, thereby effectively solving the problem of 6G high-frequency band transmission signal coverage and reducing multipath fading. Secondly, a quasi-neural network is used to construct a multi-user multi-hop communication network, a relay node cooperative communication link is constructed, and the node weight parameters are dynamically updated by forward transmission and error back propagation process to reduce the transmission error, so as to solve the problem of high difficulty of node collaboration and high transmission error in large-scale networks. Finally, in order to further reduce the transmission error caused by multipath fading, the present invention introduces space-time block coding on the basis of quasi-neural network, and adopts the training process of quasi-neural network instead of the maximum likelihood decoding process of space-time block coding to overcome the decoding error caused by channel estimation error and further improve the transmission quality of 6G communication system.

[0077] In summary, the present invention achieves high-quality signal transmission while improving the signal coverage of the communication system. It is suitable for application scenarios such as wireless communications in urban environments, long-distance transmission in harsh environments, and large-scale Internet of Things (IoT), and provides effective guarantee for future 6G high-frequency band transmission.

[0078] Parts of the present invention that are not described in detail belong to common knowledge among those skilled in the art.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Obviously, 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.

Claims

1. A self-organizing network pseudo-neural channel coding and decoding method, characterized in that: The steps include: (1) constructing a multi-hop self-organizing quasi-neural network by a front-end node layer, a relay node layer and a control end node layer, wherein the front-end node layer includes a front-end node, the relay node layer includes three fixed relay nodes, and the control end node layer includes a control center node; (2) Inputting the original pilot image into the multi-hop self-organizing quasi-neural network, setting the maximum number of iterations to T, the error threshold to η, and the learning rate to α, the network is trained; (3) Initialize the relevant parameters of the multi-hop self-organizing quasi-neural network and initialize the current number of iterations t=1; (4) The front-end node receives the original pilot image signal, and performs adaptive processing and channel splitting on the original pilot image signal according to the front-end node weight parameter, and then broadcasts the processed signal to the relay node layer through each channel; (5) The relay node layer receives the pilot image signals of each channel from the front-end node, performs adaptive processing according to the relay node weight parameters, performs space-time block coding on the processed signals, and broadcasts them to the control center node through multiple time slots; (6) The control center node receives the coded signals from each time slot, performs signal merging processing on all the coded signals, and performs adaptive processing and channel synthesis operations on the merged signals according to the weight parameters of the control center node, and restores the received coded signals to image signals, that is, obtains received image signals; (7) calculating the transmission error of the original pilot image signal and the received image signal according to the pixels at corresponding positions, and determining whether the transmission error is less than a preset error threshold; If it is less than, directly execute step (9); otherwise, continue to execute step (8); (8) Back-propagate the gradient error from the control center node to the front-end node via the relay node layer, update the weight parameters at each node, and determine whether the current number of iterations t is equal to T; If they are equal, execute step (9); otherwise, add 1 to t and return to execute step (4); (9) Load the current neural network weight parameters as the transmission weights of the multi-hop self-organizing quasi-neural network, and use the network to transmit the image to be transmitted.

2. The method according to claim 1, characterized in that: Step (2) inputs the original pilot image into the multi-hop self-organizing quasi-neural network, specifically converting the RGB color image to be transmitted into a binary bit stream, i.e., the original pilot image signal, and then inputting the signal into the front-end node.

3. The method according to claim 1, characterized in that: The relevant parameters described in step (3) include the front-end node weight parameter, the weight parameter of each relay node in the relay node layer, and the control center node weight parameter.

4. The method according to claim 1, characterized in that: The adaptive processing described in steps (4)-(6) refers to beamforming and nonlinear transformation processing of the signal.

5. The method according to claim 4, characterized in that: In step (4), the original pilot image signal is adaptively processed and channel split, and the front-end node performs adaptive signal processing on the signal, splits the image information according to the R channel, the G channel and the B channel, and transmits the image information of each channel through three transmission paths respectively.

6. The method according to claim 4, characterized in that: Step (5) performs adaptive processing according to the relay node weight parameters, performs space-time block coding on the processed signal, and broadcasts it to the control end node through multiple time slots. Specifically, at the relay node, three relay nodes respectively perform adaptive signal processing on the signals transmitted from different channels, and then construct the correlation between the transmitted signals through space-time block coding. The encoded signals are relayed and forwarded by the three relay nodes respectively and sent to the control end node.

7. The method according to claim 4, characterized in that: Step (6) performs signal merging processing on all coded signals, and performs adaptive processing and channel synthesis operations on the merged signals. Specifically, the binary bit streams received in each time slot of the control center are signal merged and adaptively processed to obtain a transformed binary bit stream, and the transformed binary bit stream is converted back to a single-channel image for channel synthesis to obtain an RGB image after transmission.

8. The method according to claim 1, characterized in that: The transmission error in step (7) is calculated according to the following formula: Among them, s is the original pilot image signal, To receive image signals.

9. The method according to claim 1, characterized in that: Step (8) is to back-propagate the gradient error from the control center node to the front-end node via the relay node layer, specifically by using a neural network back-propagation gradient descent method. The process first calculates the error term and the gradient of the error relative to the weight parameter of the control center node at the control center node, updates the weight parameter of the control center node according to the gradient, and transfers the error gradient to the relay node layer by the chain rule; repeats the same process, back-propagates the error from the relay node layer to the front-end node, transfers the gradient layer by layer, and updates the weight parameter.

Images