End-to-end communication system with variable rate mechanism

Abstract

The invention belongs to the technical field of communication, and particularly relates to an end-to-end communication system with a variable rate mechanism. According to the invention, the transmission rate mode supported by the existing E2E system is expanded, and effective utilization of channel resources by the E2E system can be realized: firstly, through reasonable network structure division and expansion, the E2E system can work in multiple rate modes; secondly, a variable rate network training mechanism is adopted, so that the E2E system can achieve the performance close to the optimal performance in various rate modes through less parameter training; and finally, the trained network is matched with a signal-to-noise ratio estimation technology, so that the E2E system can evaluate the surrounding environment state and carry out rate adjustment, and the rate self-adaption of the network aiming at the environment condition is realized, so that certain BER performance is kept, and meanwhile, communication is carried out with the efficiency as high as possible.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to an end-to-end communication system with a variable rate mechanism. Background Art

[0002] In traditional communication system design, the entire communication process is decomposed into multiple independent functional modules, each of which performs a specific function, such as source / channel coding, modulation, channel estimation, equalization, etc. Although this method can obtain a communication system that meets application requirements, it cannot guarantee that the individually optimized processing blocks can achieve the best link performance when connected together. On the contrary, the end-to-end (E2E) communication technology that has emerged in recent years is a communication system design method that uses neural networks to build autoencoders and uses deep learning technology to jointly optimize the transmitter and receiver. In theory, it can obtain the communication waveform with the best global performance (T.O'Shea and J.Hoydis, "An introduction to deep learning for the physical layer," IEEE Trans.Cognit.Commun.Netw., vol.3, no.4, pp.563–575, Dec.2017.).

[0003] E2E communication is a data-driven communication mode based on deep learning technology. Each part of the E2E system is composed of neural networks. Training these neural networks under the supervision of E2E loss can make the system close to the global optimum, thereby solving the problem that traditional communication systems are difficult to achieve optimal performance in some scenarios where it is difficult to model the channel. In actual environments, channel conditions (including noise and channel fading) are always changing. When the channel conditions are poor, the probability of transmission bit errors is high, the channel capacity is small, and error-free transmission can only be performed at a lower rate. When the channel conditions are good, the channel capacity is large, and the channel can support error-free transmission of information at a higher rate. Summary of the invention

[0004] In view of the problem that the existing E2E system cannot make full use of limited channel resources, the present invention proposes an E2E communication system with a variable rate mechanism. In order to make full use of limited channel resources, the E2E system needs to support multiple rate modes and be able to adaptively adjust the information transmission rate according to channel conditions. Therefore, based on the structure of the E2E system, the present invention invents a new E2E communication waveform system that can support multiple transmission rates, and can achieve near-optimal performance in various rate modes through fewer parameter training. Ultimately, the trained E2E system can make full use of channel resources in a changing channel.

[0005] The technical solution of the present invention is:

[0006] An end-to-end communication system with a variable rate mechanism, wherein the transmitter and the receiver of the system are constructed by neural networks, the transmitter is divided into a relationship extraction part and a transmission rate control part, and the receiver is divided into a receiving rate control part and a relationship interpretation part, respectively. and To represent the mapping they represent; Among them, the transmission rate control part and the receiving rate control part expand the transmission rate mode that the network can support by changing the modulation order. The communication process is a joint action of the four-part network. The transmitter converts the transmission data into a transmission signal suitable for transmission in the channel according to the channel characteristics, and the receiver recovers the data, thereby realizing communication based on the variable rate mechanism; Correspondingly, the network constituting the transmitter and the receiver adopts a variable rate training mechanism for training, specifically:

[0007] Initialization settings have a total of T sets of rate transmission schemes, the training group size is b, the number of training rounds is E, the Adam optimizer parameters are α, β1, β2, and the initial weights of the relationship extraction part and the relationship interpretation part are The initial weights of the transmission rate control part and the reception rate control part are

[0008] First, the network is trained based on the first set of rate transmission schemes. For each training round, the network parameters are optimized by the Adam optimizer using the binary cross entropy loss function. To update:

[0009]

[0010] After the first set of schemes is trained, for the remaining rate transmission schemes, the rate obtained by the first set of schemes training is fixed in each training round. The weight of the part is calculated by using the binary cross entropy loss function. To update:

[0011]

[0012] After the training is completed, the bit error rate performance corresponding to different rate transmission schemes under different signal-to-noise ratios is obtained through testing. Then, based on the signal-to-noise ratio estimation technology, the communication rate is switched according to the communication environment of the current communication system to achieve the optimal communication rate selection while meeting the bit error performance.

[0013] The beneficial effects of the present invention are as follows: the present invention expands the transmission rate modes that can be supported by the existing E2E system, and can realize the effective use of channel resources by the E2E system: first, through reasonable network structure division and expansion, the E2E system can work in multiple rate modes; second, by adopting a variable rate network training mechanism, the E2E system can achieve near-optimal performance in various rate modes through fewer parameter training; finally, the trained network is coordinated with the signal-to-noise ratio estimation technology, so that the E2E system can evaluate the surrounding environmental conditions and adjust the rate, thereby realizing the rate adaptation of the network according to the environmental conditions, so as to communicate with the highest possible efficiency while maintaining a certain BER performance. DETAILED DESCRIPTION

[0014] Figure 1 1 is a diagram comparing an E2E network architecture provided by an embodiment of the present invention and a traditional communication system.

[0015] Figure 2 It is a schematic diagram of the principle of implementing a rate-varying mechanism in an E2E network provided by an embodiment of the present invention.

[0016] Figure 3 It is a schematic diagram of BER performance of different waveform schemes of an E2E system for completing variable rate network training provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0019] The end-to-end communication system architecture based on the autoencoder adopted by the present invention and its comparison with the traditional communication system are shown in Figure 1 The transmitter and receiver of the system of the present invention are both constructed by neural networks, using Θ t and θ r Denote their parameters. First, the launch network f t (·) Transmit data s=[s1,…,s L ],s i ∈{0,1},i∈{1,2,…,L} is transformed into a transmission signal suitable for transmission in the channel according to the channel characteristics. Where L represents the input data length, N represents the transmission signal length, and L / N≤1 can be regarded as the code rate of the system. The output of the transmitter passes through the channel to obtain the received signal Send it to the demodulation network f r (·) Perform data recovery and finally obtain the recovered data

[0020] Since Convolutional Neural Network (CNN) has excellent feature extraction capabilities, requires few parameters, and can easily change the dimension of input data, the present invention uses convolutional neural network to design transmitters and receivers. In order to ensure that the physical constraints on the transmitted signal x are met, the hardware of the transmitter will limit x. Therefore, a normalization layer is added at the end of the transmitter to impose specific constraints on x, limiting its average power to The last layer of the demodulation network is activated with a Sigmoid layer, and the binary cross entropy loss function is used as the loss function of the E2E system. Based on this, stochastic gradient descent and back propagation are performed to optimize the entire network.

[0021] The present invention Figure 1 Based on the E2E architecture shown in FIG, the transmitting network and the demodulation network are divided into four parts: the symbol relationship extraction part (transmitting network), the rate control part (transmitting network), the rate control part (demodulation network) and the symbol relationship interpretation part (demodulation network), respectively. and To represent the mapping they represent, this partitioning relationship is shown in Figure 2 Among them, the rate control part mainly expands the transmission rate mode that the network can support by changing the modulation order (MO), which is achieved by setting different step sizes to change the length of the output signal in the convolutional neural network.

[0022] Figure 2 The upper middle part represents the transmitter, the lower middle part represents the receiver, and three cases where MO is 2, 4, and 16 are shown. Based on this segmentation relationship, the present invention further proposes a training mechanism for a variable rate network. As shown in the figure, during training, the system first performs integrated training in the first waveform scheme, and obtains four trained parts. In the subsequent training of the waveform scheme, the trained inter-symbol relationship extraction and interpretation parts are fixed, and only the network of the rate control part is updated. Therefore, under different MOs, that is, different transmission rates, the inter-symbol relationship extraction part of the transmitter and the inter-symbol relationship interpretation part of the receiver are fixed, and their respective rate control parts are equivalent to a replaceable "control head", and there are as many sets of rate control parts as there are rate modes that can be replaced when changing the system rate.

[0023] The process of the variable rate network training mechanism is as follows:

[0024] 1. Initialize T sets of rate transmission schemes, training group size b, number of training rounds E, Adam optimizer parameters α, β1, β2, initial weights of the symbol-to-symbol relationship extraction part (transmitter) and the symbol-to-symbol relationship interpretation part (receiver) Initial weights of the rate control part of the T-set scheme

[0025] 2. For the first set of rate transmission schemes, for each training round, the network parameters are optimized through the Adam optimizer using the binary cross entropy loss function. To update:

[0026]

[0027] 3. After the first set of schemes are trained, for the remaining rate transmission schemes, the rate obtained by the first set of schemes training is fixed in each training round. The weight of the part is calculated by using the binary cross entropy loss function. To update:

[0028]

[0029] After the network is trained according to this variable rate training mechanism, the bit error rate performance corresponding to different rate transmission schemes under different signal-to-noise ratios can be tested. Based on the bit error rate performance obtained by the network, combined with the existing signal-to-noise ratio estimation technology, it is possible to evaluate the quality of the current environmental conditions, thereby switching to network parameters suitable for the current environment and realizing rate adaptation of the E2E system. Specifically, if Figure 2 Take the three waveform systems in as an example, from left to right they are respectively recorded as waveforms one, two and three, and it is assumed that the entire link has an allowable bit error rate. At this time, the entire rate adaptation process can be carried out as follows: the system starts working with waveform two. When the estimated signal-to-noise ratio shows that the network performance at this time is better than the allowable bit error rate, the system can be adjusted to waveform three to continue working; when the estimated signal-to-noise ratio shows that the network performance has deteriorated to below the allowable bit error rate, it is necessary to downgrade to waveform two; if waveform two still cannot meet the performance requirements, it can continue to downgrade to waveform one; when waveform one also cannot meet the requirements, the entire system needs to be retrained according to the current channel environment.

[0030] In the present invention, the communication model of the E2E system is:

[0031]

[0032] Where h is the channel coefficient vector, n is the channel noise vector, represents the Hadamard product, i.e., the element-by-element multiplication between vectors. In theory, the training mechanism proposed in the present invention is equivalent to solving the following optimization problem in advance, i.e., the integrated training under the waveform scheme:

[0033]

[0034] The constraints in the first row represent the information transmission and recovery process, and the constraints in the second row represent the channel and noise distribution preset during training. And the signal power P0. Get the optimal solution Then solve the following optimization problems in turn, that is, separate training of the rate control part under other waveform schemes:

[0035]

[0036] Finally, we get all the optimal solutions After the training is completed, the receiver can control the transmission rate of the system according to the channel conditions. If the error performance is the top priority, and the transmission rate is increased as much as possible while ensuring system performance, the system can first apply the Tth set of solutions; when the error performance cannot meet the requirements, it will start to downgrade to the T-1th set; if it still does not meet the requirements, continue to downgrade. If the system is in a completely good working state, you can try to upshift to increase the transmission rate. In theory, when the transmitter and receiver can be strictly divided into four parts, that is, the rate control part does not have the function of extracting the relationship between symbols, It is the optimal solution, which is not completely true in practice. However, the present invention has shown through experiments that The system performance obtained when is the network weight is close to global training:

[0037]

[0038] The obtained performance and simulation results are shown in Figure 3 ,The experiment is carried out under AWGN channel, and the E b / N0 is 10dB. As can be seen from the figure, the results obtained by non-global training are almost the same as those of global training. At least it is a suboptimal solution or even close to the optimal solution, that is, the four-part division of the E2E system is feasible. In addition, it can be found that the non-global trained system performs even better than the global training at low signal-to-noise ratio (i.e., non-training area, that is, distribution outside the distribution included in the training set). This means that the adaptation result of the E2E system for tasks outside the training distribution during global training may not be as good as the result obtained by partial initialization using a pre-trained network. This result proves the feasibility of the E2E communication system with a variable rate mechanism proposed in the present invention and the advantages of the variable rate network training mechanism adopted.

[0039] Example

[0040] In this example, the parameters of the transmitting network and the demodulating network are shown in Table 1, where the rate control part is composed of the last convolutional layer of the transmitting network and the first convolutional layer of the demodulating network. Table 1 gives the parameters when MO is 2.

[0041] Table 1 Transmission and demodulation network parameters of the embodiment

[0042]

[0043] The specific structure of the rate control part under different modulation orders is shown in Table 2.

[0044] Table 2 Neural network structures of three rate control parts

[0045]

[0046] The specific implementation steps of this example are as follows:

[0047] Step 1: Build an E2E communication system according to Table 1 and Table 2.

[0048] Step 2: For the E2E system, perform variable rate network training to enable the network to work in multiple rate modes. The training method flow is as follows:

[0049] 1. Initialize T sets of rate transmission schemes, training group size b, number of training rounds E, Adam optimizer parameters α, β1, β2, initial weights of the symbol-to-symbol relationship extraction part (transmitter network) and the symbol-to-symbol relationship interpretation part (demodulation network) Initial weights of the rate control part of the T-set scheme

[0050] 2. For the first rate transmission scheme, for each training round, a set of 0-1 information bits s∈{0,1} is randomly generated b×1×L As the transmission data, the transmission data s is transmitted through the transmission network to obtain the transmission signal The two parts of the second dimension represent the real part and the imaginary part of the baseband signal respectively. Let the transmitted signal x pass through the channel to obtain the received signal y, and then let the received signal y pass through the demodulation network to obtain the recovered information data The binary cross entropy loss function is used to calculate s and The E2E loss between the two networks is calculated and the network parameters are optimized by the Adam optimizer. To update:

[0051]

[0052] 3. After the first set of schemes are trained, for the remaining rate transmission schemes, the rate obtained by the first set of schemes training is fixed in each training round. Repeat step 2 for the weight of part , and use the binary cross entropy loss function to To update:

[0053] Step 3: Use the trained network to test and record the bit error rate performance of each rate scheme of the system under different signal-to-noise ratios as the basis for rate adjustment in actual network applications.

[0054] Step 4: Deploy the widely trained transmission network and demodulation network to the actual scenario. Combine the network with the existing signal-to-noise ratio estimation technology to achieve rate adaptation of the network according to environmental conditions. The specific steps are as follows:

[0055] 1. First, apply the T-th set of rate transmission schemes and estimate the signal-to-noise ratio in the current environment. Then, evaluate the bit error rate performance of the current system based on the estimation results and the results of the network test.

[0056] 2. If the bit error rate performance at this time cannot meet the system requirements, start to downgrade to the T-1 solution and repeat operation 1. If it still does not meet the requirements, continue to downgrade;

[0057] 3. If the bit error rate performance at this time is evaluated to be better than the allowed bit error rate, it is considered that the system is in a good working state, and an attempt is made to upgrade to increase the transmission rate.

[0058] The data to be transmitted is divided into frames according to the predetermined frame length and handed over to the transmitting network. The transmitted signal is received by the demodulation network after passing through the channel. The signal-to-noise ratio estimation technology is used to monitor the BER performance of the system. According to the estimated performance, the system can change the waveform scheme in time to enable the system to communicate with the highest possible efficiency while maintaining the BER level, thereby achieving effective utilization of channel resources.

[0059] In this embodiment, the BER performance of the system is observed under an AWGN channel to verify the performance of the proposed E2E network with a variable rate mechanism, and the length L of the input data is 128. Figure 3 The BER performance test comparison under different transmission rate schemes is described after the E2E system in this embodiment completes the training using the proposed variable rate network training mechanism. b / N0 is 10dB. As can be seen from the figure, the results of dividing the network into four parts and performing non-global training are almost the same as the results of global training of the entire network (baseline). At least it is a suboptimal solution or even close to the optimal solution. The four-part segmentation of the E2E system performed by the present invention is feasible. In addition, the performance of the system trained by non-global training is low when the signal-to-noise ratio is low (i.e., non-training E b / N0 area) is even better than the performance of global training. This means that the adaptation results of the E2E system for tasks outside the training distribution during global training may not be as good as the results obtained by using a pre-trained network for partial initialization. This further reflects the advantages of the variable rate E2E communication system proposed by the present invention.

Claims

1. An end-to-end communication system with a variable rate mechanism, characterized in that: The transmitter and receiver of the system are constructed by neural networks. The transmitter is divided into a relationship extraction part and a transmission rate control part, and the receiver is divided into a receiving rate control part and a relationship interpretation part. and To represent the mapping they represent; Among them, the transmission rate control part and the receiving rate control part expand the transmission rate mode that the network can support by changing the modulation order. The communication process is a joint action of the four-part network. The transmitter converts the transmission data into a transmission signal suitable for transmission in the channel according to the channel characteristics, and the receiver recovers the data, thereby realizing communication based on the variable rate mechanism; Correspondingly, the network constituting the transmitter and the receiver adopts a variable rate training mechanism for training, specifically: Initialization settings have a total of T sets of rate transmission schemes, the training group size is b, the number of training rounds is E, the Adam optimizer parameters are α, β1, β2, and the initial weights of the relationship extraction part and the relationship interpretation part are The initial weights of the transmission rate control part and the reception rate control part are First, the network is trained based on the first set of rate transmission schemes. For each training round, the network parameters are optimized by the Adam optimizer using the binary cross entropy loss function. To update: After the first set of schemes is trained, for the remaining rate transmission schemes, the rate obtained by the first set of schemes training is fixed in each training round. The weight of the part is calculated by using the binary cross entropy loss function. To update: After the training is completed, the bit error rate performance corresponding to different rate transmission schemes under different signal-to-noise ratios is obtained through testing. Then, based on the signal-to-noise ratio estimation technology, the communication rate is switched according to the communication environment of the current communication system to achieve the optimal communication rate selection while meeting the bit error performance.

Images