A method for combating multipath fading in OFDM systems

Abstract

The present invention relates to the field of wireless communication technology, specifically, a method for OFDM systems to combat multipath fading, wherein data bits are subjected to four-dimensional spherical code modulation to generate a transmission signal; the transmission signal passes through a multipath fading channel to obtain a corresponding reception signal; the reception signal is input into a pre-trained neural network receiver to obtain a detected signal. The neural network receiver is obtained by training a deep complex fully connected neural network, and the deep complex fully connected neural network is formed by connecting complex fully connected layers of different sizes in series. This method overcomes the problems in the prior art that the receiver needs an equalizer to combat multipath fading and that index modulation is difficult to agree on a pilot, and provides an improved pilot-assisted spherical code index modulation scheme in conjunction with a complex neural network receiver model, which can achieve higher performance in combating multipath fading.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communications, in particular to a method for an OFDM system to combat multipath fading. Background Art

[0002] OFDM technology is a highly efficient multi-carrier wireless data transmission technique. Its key concept is to divide a channel into several orthogonal sub-channels, converting high-speed data signals into parallel, low-speed sub-data streams, which are then modulated and transmitted on each sub-channel. This multi-carrier modulation technique effectively combats frequency-selective fading and overcomes inter-symbol interference (ISI), and has long been a research hotspot in wireless communications. It is widely used in 802.11a / g Wi-Fi, 802.16e WiMAX, and 4G LTE standards, and OFDM continues to dominate 5G standards.

[0003] In OFDM-IM, bit information is transmitted not only through signal constellation points but also through the selection of activated orthogonal subcarrier locations. Specifically, after all subcarriers are grouped, a subset of subcarriers in each subcarrier group is activated, while the remaining subcarriers remain idle. Constellation symbols are then modulated at the locations of the activated subcarriers. Therefore, OFDM-IM not only effectively reduces inter-symbol interference (ISI) but also dynamically adjusts the structure based on index bits. Utilizing idle subcarriers improves system energy efficiency and reliability. While OFDM-IM offers these advantages, its complexity makes signal detection difficult.

[0004] The paper "J. Qu, J. Zou, J. Sun, M. Li and C. Xu, "Index Modulation Based on Four-dimensional Spherical Code," 2020 IEEE 3rd International Conference on Information Communication and Signal Processing (ICICSP), 2020, pp. 331–336, doi:10.1109 / ICICSP50920.2020.9232083," proposes an OFDM-IM system based on four-dimensional spherical codes. Compared with traditional OFDM-IM systems, this system uses high-dimensional spherical coded modulation, which increases the distance between constellation points, reduces the likelihood of signal detection errors, and improves OFDM-IM system performance.

[0005] The paper "J.Zou, J.Qu, Y.Guo, G.Liu and F.Shu,"Index Modulation Based on Four-Dimensional Spherical Code and Its DNN-Based Receiver Design," in IEEE Transactions on Vehicular Technology, vol.70, no.12, pp.13401-13405, Dec.2021, doi:10.1109 / TVT.2021.3120262." proposed an algorithm to find a better constellation mapping to minimize the bit error rate based on the four-dimensional spherical code OFDM-IM system with stronger transmission performance. A mapping relationship table with better performance was obtained, and the trained neural network receiver achieved the same performance as the maximum likelihood receiver.

[0006] Since their inception, deep neural networks have been successfully applied to various fields, including computer vision and natural language processing. Many years ago, they also began to rapidly develop in wireless communications, particularly in channel estimation and signal detection. Deep neural network receivers can achieve superior performance compared to traditional receivers when adequate training data is available. Summary of the Invention

[0007] The technical problem to be solved by this method is: to overcome the problems in the prior art that the receiver needs an equalizer to combat multipath fading and that index modulation is difficult to agree on a pilot, and to provide an improved method of combating multipath fading by combining pilot-assisted spherical code index modulation with a complex neural network receiver.

[0008] This method adopts the following technical solutions:

[0009] A method for combating multipath fading in an OFDM system comprises the following steps:

[0010] Step 1: pre-processing the bit information to be sent;

[0011] Step 2: input the pre-processed data into a four-dimensional spherical code index modulation encoder to obtain a spherical code transmission signal;

[0012] Step 3: After adding the pilot signal to the transmission signal, it is input into the IFFT, parallel-to-serial conversion, and cyclic prefix addition modules of the OFDM system to generate the transmission signal;

[0013] Step 4: Input the transmitted signal into the three multipath fading channel models of EPA, EVA, and ETU defined in the 3GPP protocol to obtain the received signal;

[0014] Step 5: Process the received signal by removing the cyclic prefix, performing serial-to-parallel conversion, and performing FFT.

[0015] Step 6. Treat the processed received signal as a complex data matrix, separate the real and imaginary parts of the data matrix, and reconstruct them in series to obtain a real data matrix of twice the size. Input the data matrix into the pilot extraction neural network model to extract the pilot, use the pilot to perform LS channel estimation, and then input the estimated signal into the deep complex fully connected neural network receiver for signal detection.

[0016] In the above technical solution, in step six, the pilot extraction neural network model includes an input layer, a real fully connected layer of 300 nodes, a batch normalization layer, a real fully connected layer of 150 nodes, a batch normalization layer, a real fully connected layer of 80 nodes, a batch normalization layer, and an output layer of 64 nodes. The deep complex fully connected neural network receiver model includes an input layer, a real fully connected layer with 300 nodes, a batch normalization layer, a real-to-complex layer, a complex fully connected layer with 200 nodes, a complex batch normalization layer, a complex fully connected layer with 100 nodes, a complex batch normalization layer, a complex fully connected layer with 50 nodes, a complex-to-real layer, and an output layer with 48 nodes; the structure of the complex fully connected layer includes four real fully connected layers, which process the real and imaginary parts of the input complex numbers respectively, and finally splice them into a complex output; the training of the deep neural network receiver includes: generating random bit information to construct a label data set, modulating the bit information through spherical code index, adding a channel model, etc. to obtain a sample data set, and putting the sample data set into a deep pilot extraction neural network and a deep receiver neural network for training to obtain a pilot extraction network model and a receiver network model.

[0017] The beneficial effect of this method is that the proposed improved spherical code index modulation scheme achieves better performance against multipath fading compared with the traditional method and spherical code index modulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a block diagram of an OFDM spherical code index modulation system according to an embodiment of the present invention.

[0019] Figure 2 This is a diagram of how information bits in each OFDM sub-block activate corresponding subcarriers according to index bits in an embodiment of the present invention.

[0020] Figure 3 This is a mapping matrix diagram of spherical codes corresponding to information bits in each OFDM sub-block in an embodiment of the present invention.

[0021] Figure 4 This is a block diagram of an improved OFDM spherical code index modulation system according to an embodiment of the present invention.

[0022] Figure 52 is a diagram of two pilot insertion schemes of the spherical code index modulation system according to an embodiment of the present invention.

[0023] Figure 6 These are parameters corresponding to the multipath fading channel in the embodiment of the present invention, including tap delay and tap power map.

[0024] Figure 7 2 is a structural diagram of a deep neural network in an embodiment of the present invention.

[0025] Figure 8 2 is a structural diagram of a deep complex neural network of a receiver in an embodiment of the present invention.

[0026] Figure 9 This is a comparison chart of the channel estimation performance between a spherical code index modulation system without pilots and a common OFDM system in an embodiment of the present invention.

[0027] Figure 10 3 is a performance comparison diagram of a spherical code index modulation system without a pilot in an embodiment of the present invention and a spherical code index modulation system using two newly proposed pilot schemes. DETAILED DESCRIPTION

[0028] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0029] like Figure 1 As shown, a method for combating multipath fading in an OFDM system specifically includes the following steps:

[0030] Step 1: pre-processing the bit information to be sent;

[0031] Step 2: input the pre-processed data into a four-dimensional spherical code index modulation encoder to obtain a spherical code transmission signal;

[0032] Step 3: After adding the pilot signal to the transmission signal, it is input into the IFFT, parallel-to-serial conversion, and cyclic prefix addition modules of the OFDM system to generate the transmission signal;

[0033] Step 4: Input the transmitted signal into the three multipath fading channel models of EPA, EVA, and ETU defined in the 3GPP protocol to obtain the received signal;

[0034] Step 5: Process the received signal by removing the cyclic prefix, performing serial-to-parallel conversion, and performing FFT.

[0035] Step 6: Treat the processed received signal as a complex data matrix, separate the real and imaginary parts of the data matrix, and reconstruct them in series to obtain a real data matrix of twice the size; input the data matrix into a deep complex fully connected neural network for training to obtain a neural network receiver.

[0036] Specifically, the OFDM system is used as the basis, and each OFDM symbol transmits 64 subcarriers, where every 8 subcarriers are used as an OFDM subblock, and 4 bits of information are transmitted according to Figure 3 After spherical code index modulation, Figure 2 The activation mode is mapped to 2 of the 8 subcarriers for transmission, and the remaining subcarriers are used Figure 4 The pilot is inserted in two pilot insertion methods. In this embodiment, the pilot value used is 0+1*j. Finally, the eight OFDM sub-blocks are combined into an OFDM symbol. Similar to the common OFDM system, IFFT, CP addition, and parallel-to-serial conversion are performed to obtain the transmit signal.

[0037] Figure 4 This is a block diagram of a spherical code index modulation system using a pilot scheme. At the transmitting end of the OFDM system, the bit stream signal is first prepared. The signal is modulated using a spherical code index and a pilot signal is added. At this point, the signal is in the frequency domain. The next step is to undergo the OFDM system's serial-to-parallel conversion and the inverse Fourier transform (IFFT). To eliminate inter-subcarrier interference, a cyclic prefix (CP) is added before the OFDM symbol. The length of the CP should take into account the actual channel and must be greater than or equal to the channel's maximum delay spread. After OFDM modulation, the signal is sent into the wireless channel for transmission. At the receiving end of the OFDM system, the received signal is processed by first removing the cyclic prefix and then performing a fast Fourier transform (FFT) to convert the parallel data into serial data for subsequent channel estimation and signal detection modules. Before channel estimation, a pilot extraction neural network is used to detect the pilot position in the signal. The pilot position is used as one of the key information for channel estimation. Finally, the signal after channel estimation is sent to the receiver neural network to recover the bit stream information of the transmitting end.

[0038] The parameters corresponding to the three multipath fading channel models EPA, EVA and ETU defined in the 3GPP protocol are as follows: Figure 6 As shown, the channel model can be generated by RayleighChannel in the comm communication toolbox in MATLAB, or it can be obtained by programming using other software such as Python using known channel tap delays and channel tap power distributions.

[0039] After passing through the multipath fading channel, Gaussian white noise is added to obtain the received signal at the receiver.

[0040] The signal received at the receiving end is in the form of a complex matrix, which undergoes serial-to-parallel conversion and CP removal. The complex signal matrix is ​​split into real and imaginary parts, which are then concatenated into a real matrix of twice the size and fed into the deep neural network.

[0041] The deep pilot extraction neural network model structure in this method is as follows Figure 7 As shown, the fully connected neural network based on different sizes of series connection includes an input layer, a real fully connected layer of 300 nodes, a batch normalization layer, a real fully connected layer of 150 nodes, a batch normalization layer, a real fully connected layer of 80 nodes, a batch normalization layer, and an output layer of 64 nodes;

[0042] In this method, the sample of the deep pilot extraction neural network model is the processed received signal, and the label is the pilot position matrix of the same size of 64, where the value of the pilot position in the pilot position matrix is ​​1, and the value of the data position and the idle position is 0;

[0043] In this method, the output of the trained deep pilot extraction neural network model is the pilot position matrix, which is combined with the received frequency domain signal for least squares channel estimation to obtain the input of the receiver neural network.

[0044] The least squares method is the LS algorithm, whose criterion is to minimize the weighted error between the estimated value and the true value. Assume that the frequency domain received signal at the pilot position is:

[0045] y=[Y(0),…,Y(N p -1)] T ,

[0046] The noise is:

[0047] w=[W(0),…,W(N p -1)] T ,

[0048] The corresponding transmitted pilot signal is:

[0049] X=diag{x}=diag{[X(0),...,X(N p -1)] T},

[0050] Among them, N p is the number of pilots, then the LS estimation result The cost function is:

[0051]

[0052] right Ask about The derivative of , we can get:

[0053]

[0054] Therefore:

[0055]

[0056] It can be seen that the LS algorithm ignores the channel and noise factors in the calculation process, and its estimation performance is easily affected by inter-carrier interference and noise. However, due to its small computational complexity and low implementation difficulty, it is suitable for the data processing stage of high-precision channel estimation and channel detection.

[0057] In this method, the first fully connected layer of the receiver neural network is set to real number form. It is necessary to divide the complex data matrix into two parts, the real part and the imaginary part, and splice them into a real number matrix of twice the size before inputting it into the neural network.

[0058] The receiver neural network structure of this method is as follows Figure 8 As shown in the figure, the deep complex fully connected neural network based on complex fully connected layers of different sizes in series includes an input layer, a real fully connected layer of 300 nodes, a batch normalization layer, a real-to-complex conversion layer, a complex fully connected layer of 200 nodes, a complex batch normalization layer, a complex fully connected layer of 100 nodes, a complex batch normalization layer, a complex fully connected layer of 50 nodes, a complex-to-real conversion layer, and an output layer of 48 nodes;

[0059] Furthermore, the structure of the complex fully connected layer includes four real fully connected layers, which process the real and imaginary parts of the input complex number respectively, and finally splice them into a complex output. The principle of a single complex neuron is expressed by the following formula:

[0060]

[0061] in, It is the output and input of a single complex neuron. When inputting, the complex neuron needs to be Convert to real neurons, and convert real neurons to complex neurons when outputting

[0062] In this method, the middle layer between the deep neural network and the receiver neural network is also called the hidden layer, and the ReLU activation function is used. It has faster learning speed, better training performance, and wider application in the deep learning architecture. Its mathematical representation is:

[0063] f(x)=max(0,x)

[0064] In this method, the output layers of the deep pilot extraction neural network and the receiver neural network use the Sigmoid activation function. The Sigmoid function is a commonly used output layer activation function and is suitable for scenarios such as binary classification and logistic regression modeling of neural networks. Its mathematical representation is:

[0065]

[0066] In this method, both the deep pilot extraction neural network and the receiver neural network model use the adaptive moment estimation optimization algorithm, also known as the Adam algorithm. During the training process, the Adam algorithm has its own learning rate and momentum, so the parameter updates are independent, which also improves the stability and speed of training. In actual use, the performance is good. In the present invention, the learning rate of the Adam algorithm is set to 0.001.

[0067] The loss function used by the neural network model in this method is the mean square error function, and its mathematical expression is:

[0068]

[0069] Where n is the number of signals in a single batch of samples, Y i and Denotes the training label and neural network output, respectively. During the training process, the Adam algorithm dynamically adjusts the network parameters and learning rate to minimize the loss function, gradually approaching the global optimum after multiple iterations.

[0070] The training of this method needs to be divided into two steps. The first step is to train the deep pilot extraction neural network to extract the pilot. The second step is to use the extracted pilot to perform LS channel estimation on the received signal and send it to the receiver neural network for training, thus completing the training of the neural network of the entire system.

[0071] Based on the problem that simple neural network receivers cannot effectively recover signals after complex multipath fading, this method uses spherical code index modulation to preliminarily solve the above problem. On this basis, two new pilot insertion methods are invented, and the neural network is used to obtain the pilot position for channel estimation, which further improves the performance of the spherical code index modulation system.

[0072] This method first uses spherical code index modulation combined with a neural network receiver to compare the traditional channel estimation method. The traditional channel estimation method uses pilots to perform LS channel estimation and then uses a traditional receiver to detect the signal. The number of pilots in the above traditional channel estimation method is 16, and each OFDM symbol transmits the same information as the method used in this method.

[0073] from Figure 9As can be seen from the figure, the three multipath channels have varying degrees of complexity. Under the same transceiver system, the ETU channel is the most complex, resulting in the highest received bit error rate (BER), followed by the EVA channel, and the EPA channel with the lowest BER. The scheme using spherical code index modulation in the figure significantly outperforms the standard OFDM channel estimation scheme across the entire SNR range in all three multipath channels. When the channel signal-to-noise ratio (SNR) is 15dB, the spherical code index modulation scheme achieves a BER performance gain of approximately 9dB in the ETU channel, approximately 5.7dB in the EVA channel, and approximately 3dB in the EPA channel. It is worth noting that the ETU channel is more complex than the EVA and EPA channels, and the system performance gain is also the most significant in the ETU channel, while the system performance gain is relatively minimal in the EPA channel.

[0074] This fully demonstrates the potential of using spherical code modulation schemes with larger Euclidean distances to combat multipath fading. Simulation results show that the present invention, after comparing spherical code index modulation schemes with commonly used channel estimation schemes, has found superior performance. It can achieve the performance of conventional OFDM systems using channel estimation without channel estimation, and often achieves even greater performance gains in more complex channels.

[0075] Figure 10 The performance comparison between the spherical code index modulation system without pilot and the spherical code index modulation system with two newly proposed pilot schemes is shown in Figure 2. Figure 10 It can be seen that when the spherical code index modulation scheme adopts the full pilot scheme proposed in the present invention, under an environment where the channel signal-to-noise ratio is 15dB, the performance is close to that of the spherical code index modulation scheme without a pilot in the ETU channel. In the EVA channel, a bit error rate performance gain of about 3.7dB can be obtained, and in the EPA channel, a bit error rate performance gain of about 7.8dB can be obtained. It is worth noting that the ETU channel is more complex than the EVA and EPA channels. With the help of the channel estimation provided by the pilot, the system cannot compensate for the impact of the signal transmitted in the idle subcarrier on the subcarrier transmitting data. Compared with the EPA channel, the performance gain can be greatly improved.

[0076] from Figure 10A comparison of the two pilot schemes shows that, in most cases, the scheme using pilots has better resistance to multipath fading than the scheme without pilots. In the ETU channel environment, the bit error rate of the spherical code index modulation scheme using half pilots is the worst compared to the schemes using full pilots and no pilots. Analysis shows that this is because the addition of pilots. Although the pilots transmitted on idle subcarriers have an effect on improving the bit error rate after channel estimation, the number of pilots is half that of the full pilot scheme, which cannot compensate for the inter-carrier interference caused by the subcarriers transmitting data in the multipath channel environment. In the EVA channel environment, the full pilot scheme is better than the scheme without pilots, and both schemes are still better than the half pilot scheme. Finally, in the EPA channel environment, since the channel environment is less complex than that of ETU and EVA, the half pilot scheme can show the best effect. From the simulation results, it can be seen that the spherical code modulation scheme does not have consistent advantages in resisting different multipath channels after using the two different pilot schemes proposed by the present invention. From the results, it can be seen that when the channel environment is good, the half-pilot scheme is suitable to obtain the best performance improvement. However, in the case of a poor channel environment, since the use of more pilots does not bring performance improvement, it is suitable to adopt a scheme without pilots, which can reduce the complexity of the receiving end and alleviate its computational workload.

[0077] On the other hand, one advantage of this method is that data-driven neural networks can often be regarded as a black box. When combined with the communications field, professional knowledge in the communications field is often not used. Although this method uses a neural network as a receiver, it uses more knowledge in the communications field, which also weakens the role of the neural network in combating multipath fading, providing new ideas for further research in the future.

[0078] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for combating multipath fading in an OFDM system, characterized in that: The following steps are involved: Step 1: pre-processing the bit information to be sent; Step 2: input the preprocessed data into a four-dimensional spherical code index modulation encoder to obtain a spherical code transmission signal; Step 3: After inserting the pilot signal into the transmission signal, the signal is input into the IFFT, parallel-to-serial conversion, and cyclic prefix adding module of the OFDM system to generate the transmission signal; Step 4: Input the transmitted signal into three multipath fading channel models of EPA, EVA and ETU defined in the 3GPP protocol to obtain a received signal; Step 5: Process the received signal, remove the cyclic prefix, perform serial-to-parallel conversion, and perform FFT; Step six: regard the processed received signal as a complex data matrix, separate the real and imaginary parts of the data matrix, and reconstruct them in series to obtain a real data matrix of twice the size. Input the data matrix into the pilot extraction neural network model to extract the pilot, use the pilot to perform LS channel estimation, and then input the estimated signal into the deep complex fully connected neural network receiver for signal detection.

2. The method for combating multipath fading in an OFDM system according to claim 1, characterized in that: The specific process of step three is: using the OFDM system as a basis, each OFDM symbol transmits 64 subcarriers, where every 8 subcarriers are used as an OFDM sub-block, and 4 bits of information are modulated according to the spherical code index and mapped to 2 of the 8 subcarriers for transmission. The remaining subcarriers are inserted with pilots using two pilot insertion methods, where the pilot value used is 0+1*j. Finally, the 8 OFDM sub-blocks are combined into an OFDM symbol, and IFFT, CP addition, and parallel-to-serial conversion operations are performed to obtain the transmitted signal.

3. The method for combating multipath fading in an OFDM system according to claim 2, characterized in that: The pilot extraction neural network model in step six includes an input layer, a real fully connected layer of 300 nodes, a batch normalization layer, a real fully connected layer of 150 nodes, a batch normalization layer, a real fully connected layer of 80 nodes, a batch normalization layer, and an output layer of 64 nodes.

4. The method for combating multipath fading in an OFDM system according to claim 2, characterized in that: The deep complex fully connected deep neural network receiver model in step six includes an input layer, a real fully connected layer of 300 nodes, a batch normalization layer, a real to complex layer, a complex fully connected layer of 200 nodes, a complex batch normalization layer, a complex fully connected layer of 100 nodes, a complex batch normalization layer, a complex fully connected layer of 50 nodes, a complex to real layer, and an output layer.

5. The method for combating multipath fading in an OFDM system according to claim 4, characterized in that: The structure of the complex fully connected layer includes four real fully connected layers, which process the real part and imaginary part of the input complex number respectively, and finally splice them into complex output. The principle of a single complex neuron is expressed by the following formula: in, It is the output and input of a single complex neuron. When inputting, the complex neuron needs to be Convert it into a real neuron, and convert the real neuron into a complex neuron when outputting it 6. The method for combating multipath fading in an OFDM system according to claim 5, characterized in that: The training of the deep neural network receiver includes: generating random bit information to construct a label data set, modulating the bit information through spherical code exponents, adding channel model processing to obtain a sample data set, putting the sample data set into a deep neural network for training, and obtaining a deep neural network receiver model.

7. The method for combating multipath fading in an OFDM system according to claim 6, characterized in that: In step six, the output of the trained deep neural network is a pilot position matrix, which is combined with the received frequency domain signal for least squares channel estimation to obtain the input of the receiver neural network.

8. The method for combating multipath fading in an OFDM system according to claim 7, characterized in that: The least squares criterion is to minimize the weighted error between the estimated value and the true value. Assume that the frequency domain received signal at the pilot position is: and=[Y(0),…,Y(N) p -1)] T The noise is: w=[W(0),…,W(N p -1)] T , The corresponding transmitted pilot signal is: X=diag{x}=diag{[X(0),...,X(N p -1)] T }, Among them, N p is the number of pilots, then the LS estimation result The cost function is: right Ask about The derivative of , we can get: Therefore, there are:

Images