Optical FBMC dual-mode index modulation communication system and signal detection method thereof

Abstract

The invention belongs to the field of wireless optical communication, and particularly relates to an optical FBMC (Fiber Baseboard Management Controller) dual-mode index modulation communication system and a signal detection method thereof, comprising the following steps: Step 1: transmitting two constellation modes by using all available subcarriers, and simultaneously retaining diversity gain of index modulation; 2, optimizing a dual-mode constellation through phase rotation and amplitude scaling, and generating an optimal constellation suitable for the optical FBMC system; 3, designing a deep learning assisted signal detection algorithm, and detecting the sent information; and Step 4, optimizing hyper-parameters of a signal detection algorithm by using an artificial mouse algorithm. Simulation and experiment results show that the system improves the bit error rate performance of the system through the constellation design, and realizes the bit error rate performance close to maximum likelihood detection on the premise of greatly reducing the detection complexity.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless optical communication, and in particular to an optical FBMC dual-mode index modulation communication system and a signal detection method thereof. BACKGROUND

[0002] In recent years, optical filter bank multi-carrier (FBMC) technology has attracted extensive attention in visible light communication, passive optical network and free space optical link due to its excellent spectral localization characteristics and the characteristics of not requiring cyclic prefix. Compared with the traditional optical orthogonal frequency division multiplexing system, optical FBMC can significantly reduce the sidelobe leakage, improve the spectral efficiency, and has stronger robustness in multipath or dispersion channel. However, the real value modulation introduced by optical FBMC restricts the actual information rate that can be carried in each symbol frame.

[0003] Index modulation (IM) transmits additional information bits through the activation index of subcarriers, time slots, antennas and other resources, thereby improving the spectral efficiency and error rate performance. Inspired by this, optical FBMC index modulation (FBMC-IM) carries additional bits through activated subcarrier indexes, which can not only improve the spectral efficiency without increasing the bandwidth, but also reduce the number of simultaneously activated subcarriers. However, in the optical FBMC-IM system, due to the existence of non-activated subcarriers, the spectral efficiency of the system still has some loss. Multi-mode index modulation fills all subcarriers with distinguishable constellations, so that each subcarrier can carry information, further improving the spectral efficiency of the system. However, for the multi-mode optical FBMC-IM system, in order to be able to correctly demodulate the constellation symbols in each subblock at the receiving end, the constellation symbol set used cannot have the same constellation symbol, so it is necessary to design a constellation mapping mode without intersection of symbols to improve the error performance of the system. In addition, when the above-mentioned system uses traditional maximum likelihood (ML) and approximate maximum likelihood detection algorithm, the detection complexity increases exponentially with the increase of the number of subcarriers or the order of constellation. Although deep learning has been widely used in wireless communication systems, especially in signal detection problems, manual parameter tuning is required under different deep learning model structures and data sets to achieve better performance. SUMMARY

[0004] In view of the problems in the background art, the present application discloses an optical FBMC dual-mode index modulation communication system and a signal detection method thereof, aiming to improve the spectral efficiency of the communication system, improve the error performance of the system, and realize approximately optimal low complexity signal detection.

[0005] To achieve the purpose of the present application, the present application realizes the following technical scheme: an optical FBMC dual-mode index modulation communication system and a signal detection method thereof, comprising the following steps: (1) Transmit two constellation modes using all available subcarriers while preserving the diversity gain of index modulation; (2) Optimize the dual-mode constellation through phase rotation and amplitude scaling to generate the best constellation suitable for optical FBMC systems; (3) Design a deep learning-assisted signal detection algorithm to reliably detect the transmitted information in a data-driven manner; (4) Use the artificial ground squirrel algorithm to optimize the hyperparameters of the signal detection algorithm, avoiding local optima and improving detection performance.

[0006] Further, in the step (1), for the optical FBMC dual-mode index modulation communication system, the number of subcarriers is , input bits are divided into subblocks by a bit splitter, each subblock contains bits and subcarriers, i.e. . Then, bits are further divided into two parts, index bits and information bits . Index bits are sent to an index selector for mapping subcarrier indices. Information bits are sent to two different constellation mappers for mapping through two different types of constellation modes and . After mapping, different FBMC subblock combinations are combined for Hermite symmetry operation, and the output signal is denoted as . Then, is processed by a polyphase network composed of point inverse Fourier transform and Martin-Mirabassi prototype filter, and the obtained bipolar signal can be represented as . Finally, the signal is superimposed with a direct current bias to drive the laser to generate an optical signal.

[0007] Further, in the step (2), for the optical FBMC dual-mode index modulation communication system, in order to reliably detect the constellation modes and at the receiving end, a phase rotation and amplitude scaling constellation design method is further proposed. At the same time, in order to ensure power consistency and avoid the influence of excessive or insufficient amplitude on system performance, each point in the constellation point set must satisfy the average signal power constraint: ; Further, in step (3), the deep learning-assisted signal detection method, named DMOFIMNet, consists of a preprocessing module, a ResNet module, and an output module connected in series. The preprocessing module stacks the real part, imaginary part, and energy of the signal together to generate a signal with dimension 1. signal matrix The data is one-hot encoded and then input into the ResNet module. To maintain the stability of the data feature distribution, a normalization layer is used to process the feature matrix, ensuring that the mean of each feature is 0 and the variance is 1. Furthermore, residual connections are introduced to smooth the training process and prevent overfitting as the model deepens. To utilize the class-discriminative local features from the convolutional layers for the final classification task, the signal output from the residual connections is processed by an output layer consisting of a one-dimensional convolutional layer and a Softmax activation function. The output of the Softmax activation function is the probability of the signal being transmitted; its dimension is 1024, encompassing all possible combinations of transmitted signals. Finally, the original transmitted information is recovered after decoding.

[0008] Furthermore, in step (4), the global optimization advantage of the artificial lemming algorithm is used to efficiently explore the hyperparameter space (such as learning rate, number of neurons in convolutional layers, etc.) and optimize the structure of the signal detection method DMOFIMNet to improve model training efficiency and detection performance.

[0009] The beneficial effects of this invention are as follows: This invention is an optical FBMC dual-mode index modulation communication system and its signal detection method. First, it combines dual-mode index modulation technology with filter bank multicarrier technology and introduces it into wireless optical communication to construct an optical FBMC dual-mode index modulation communication system, which improves spectral efficiency compared to optical filter bank multicarrier systems. Second, it proposes a dual-mode constellation design method through phase rotation and amplitude scaling to generate the optimal constellation suitable for the scheme. Based on this, it utilizes the advantages of the deep learning network ResNet in capturing long-term dependencies and global features between signals, and designs a signal detection method DMOFIMNet based on the ResNet network to replace the traditional ML detection method. At the same time, it uses the artificial lemming algorithm to optimize the hyperparameters of DMOFIMNet, avoiding the hyperparameters from getting trapped in local optima, thereby achieving a system bit error rate performance that approaches the optimal ML detection method while significantly reducing computational overhead. Attached Figure Description

[0010] Figure 1 This is a model of the optical FBMC dual-mode indexed modulation communication system of the present invention; Figure 2 The phase-rotating constellation design of this invention; Figure 3 This invention relates to a scale-adjustable constellation design. Figure 4 The structure of the DMOFIMNet signal detection method of the application; Figure 5 The ResNet hyperparameter optimization algorithm process of the artificial vole algorithm of the application; Figure 6 The change of the fitness value of the ResNet hyperparameter optimized by the artificial vole algorithm of the application with the number of iterations; Figure 7 The bit error rate comparison of the DMOFIMNet detection algorithm and the ML detection algorithm of the application; Figure 8 The calculation complexity comparison of the DMOFIMNet detection algorithm and the ML detection algorithm of the application. DETAILED DESCRIPTION

[0011] For the purpose, technical scheme and advantages of the embodiments of the application, the technical scheme in the embodiments of the application will be further described below with reference to the drawings. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0012] The embodiment of the application provides an optical FBMC double-mode index modulation communication system and a signal detection method thereof, and the system model is as shown in Figure 1 The specific steps are as follows: step 1: for the optical FBMC double-mode index modulation communication system, the number of subcarriers is , The input bits are divided into subblocks by a bit divider, each subblock contains bits and subcarriers, that is , . Then, bits are further divided into two parts, index bits and information bits . The index bits are sent to an index selector for mapping subcarrier indexes. The information bits are sent to two different constellation mappers, and are mapped by two different types of constellation modes and . After the mapping is completed, different FBMC subblocks are combined together, and then Hermite symmetry operation is performed, and the output signal is recorded as . Then, passes through The polyphase network processing composed of point-wise inverse Fourier transform and Martin-Mirabassi prototype filter, the obtained bipolar signal can be expressed as Finally, the transmitted signal is superimposed with a DC bias to drive the laser to generate optical signal, which is transmitted through Gamma-Gamma turbulence channel and converted into electrical signal by photodetector.

[0013] Step 2: For optical FBMC dual-mode index modulation communication system, the overall performance depends largely on the two distinguishable modes and adopted. In order to reliably detect the constellation mode at the receiving end, two dual-mode constellation design methods of phase rotation and amplitude scaling are further proposed. At the same time, in order to ensure power consistency and avoid affecting system performance due to excessive or insufficient amplitude, each point in the constellation point set needs to meet the average signal power constraint: (1); Take 4th order and as an example, the design with two amplitude layers, and , the constellation points are uniformly distributed at 90° apart on two circles with radii and , where , is the scale factor, which satisfies under the constraint of average normalized signal power. The Euclidean distance between adjacent two points is: (2); The constellation points of and are optimized to maximize the minimum Euclidean distance, and the Euclidean distance between adjacent two points is: (3); and The Euclidean distance between the nearest constellation points is: (4); The minimum Euclidean distance between the constellation points of the two modes is: (5); According to the bit error rate theory, the greater the minimum Euclidean distance, the lower the probability of misjudgment when the signal is disturbed by noise, so needs to be maximized.

[0014] Under the average power normalization constraint, Figure 2 The constellation design diagrams of fixed inner circle constellation points and only 15°, 30° and 45° rotation of outer circle constellation points are given, and the Euclidean distance between all constellation points in each rotation case is calculated. In the phase rotation constellation design, when the outer circle is rotated by 45°, the minimum Euclidean distance between the constellation points is the largest among different rotation angles. Therefore, in the scale stretching constellation design, the fixed phase rotation angle is 45°, and under the premise of meeting the power constraint, the minimum Euclidean distance is maximized by optimizing the scale factor, thereby improving the signal noise performance. Figure 3 The constellation design diagrams under different scale factors are given.

[0015] Step 3: The received electrical signal at the receiving end can be represented as , where represents convolution operation, represents photoelectric conversion efficiency, represents additive white Gaussian noise with mean 0 and variance , represents the time domain channel fading coefficient, which obeys the Gamma-Gamma distribution that can accurately describe the atmospheric turbulence channel, and its probability density function is: (6); After matched filtering, analysis filter bank (AFB) and demapping processing, the complex modulation signal is obtained.

[0016] Step 4: The complex modulation signal; considering that the energy of the signal reflects the attenuation of the signal in the transmission process and the energy level of the signal source itself, the signal is then processed by the deep learning assisted signal detection method DMOFIMNet. DMOFIMNet is composed of a preprocessing module, a residual module and an output module in series, and the specific structure is shown in Figure 4 . In order to make the detection algorithm effectively learn and extract signal features, the complex signal signal sequence is decomposed to obtain its real part and imaginary part ; considering that the energy of the signal reflects the attenuation of the signal in the transmission process and the energy level of the signal source itself, the energy of the signal is calculated. Finally, the , and are stacked together to generate a signal matrix with a dimension of , which is input to the ResNet module after One-hot encoding.​

[0017] Step 5: In the ResNet module, in order to maintain the stability of the data feature distribution, the feature matrix is processed using the normalization layer, so that the mean of each feature is 0 and the variance is 1. At the same time, the residual connection is introduced to make the training process more stable and avoid network overfitting when the model is deepened. The whole operation process can be represented as: (7); In the formula: , is the RELU activation function of the convolution layer; , denotes the weight of the convolution kernel; , denotes the bias of the convolution layer; denotes the convolution operation. In order to use the local features with class distinction in the convolution layer for the final classification task, the signal output by the residual connection is processed by an output layer composed of a one-dimensional convolution layer and a Softmax activation function. The output of the Softmax activation function is the probability of the possible transmission signal, and its dimension is 1024, containing all possible transmission signal combinations. Finally, the original transmission information is recovered after decoding. The whole process can be represented as: (8); In the formula: is the Softmax activation function of the convolution layer; denotes the weight of the convolution kernel; denotes the bias of the convolution layer; denotes the convolution operation.

[0018] Step 6: The learning rate, number of convolution layer neurons, and batch size of the ResNet model and other hyperparameters will significantly affect the classification accuracy and generalization ability on the validation set. Only the model trained under the optimal combination of hyperparameters can achieve the best classification effect. Therefore, the problem of finding the optimal combination of hyperparameters is regarded as a global optimization process in the hyperparameter space by applying the artificial ant algorithm (ALA). The ALA-ResNet algorithm for optimizing ResNet hyperparameters is shown in Figure 5 .

[0019] Step 7: Before entering the iterative optimization process, first set the size of the population to , and the maximum number of iterations to The hyperparameter ranges are defined, and the positions of all individuals are initialized within these ranges. Table 1 lists the specific hyperparameters involved and their ranges. Individual positions in the population are represented as index vectors, with values ​​selected from the hyperparameter set. For example, the learning rate set A = {0.0001, 0.001, 0.01} corresponds to the index space {0, 1, 2}, the hidden layer node count set B = {32, 64, 128, 512} corresponds to the index space {0, 1, 2, 3}, and the batch size combination C = {100, 1000} corresponds to the index space {0, 1}. Individual The position is ,in , , .

[0020] Table 1. Hyperparameter settings for ResNet model optimization using the artificial lemming algorithm. .

[0021] Step 8: To evaluate the performance of individuals in the population, a corresponding ResNet model is first constructed based on the number of neurons in the convolutional layer, the learning rate, and the batch size parameters for each individual. Then, using the cross-entropy loss function as the fitness function, the fitness value of each individual on the validation set is calculated sequentially. The smaller the fitness value, the better the performance of that hyperparameter combination in the current task. The formula for calculating the fitness function is as follows: (9); in, This indicates that the ResNet model is for the th The predicted output for each sample, For its corresponding real label, To verify the total number of samples in the set. This metric provides a quantitative basis for evaluating the performance of individuals and serves as an important standard for ranking and updating during algorithm iteration.

[0022] Step 9: In the ALA-ResNet algorithm, migration, burrowing, foraging, and enemy avoidance strategies are closely linked through an energy factor. In each iteration, to maintain a balance among the four search strategies, the energy factor gradually decreases during the iteration process. The formula for calculating the energy factor is as follows: (10).

[0023] Step 10: When the energy factor When the fitness value of the current iteration is less than the fitness value of the last iteration, the algorithm executes the migration strategy with a probability of 0.3 or the hole digging strategy with a probability of 0.7; otherwise, foraging or enemy avoidance is selected with a probability of 0.5. For each individual, the migration strategy randomly selects a new value from the two parameter sets, obtaining a new combination with a larger span in the parameter space, which helps the population quickly introduce diversity and jump out of the local. The hole digging strategy uses Lévy flight plus differential guidance towards the optimal direction to dig out better solutions around the global excellent area. The foraging strategy only changes the index of the neighbor position with a Hamming distance of 1, selecting parameters in the neighbors of the upper and lower indexes to strengthen the local convergence speed; the enemy avoidance strategy uses discrete Lévy flight to make the hyperparameter index jump 1 or more steps to avoid falling into the local optimum.

[0024] Step 11: During the iterative process of the ALA-ResNet algorithm, all individuals use these strategies to update the candidate solution and determine the optimal solution of the current iteration. By comparing the fitness values of the current iteration and the last iteration, a new optimal value is obtained and the global best hyperparameter index is updated. If the number of iterations of ALA-ResNet reaches the maximum number of iterations, the global best hyperparameters are decoded according to the global best hyperparameter index, and the ResNet model is trained using these best hyperparameters and used for signal detection tasks.

[0025] Step 12: In the initialization phase of the ALA-ResNet algorithm for hyperparameter optimization, the population size is set to 20 and the number of iterations is set to 50. Each individual in the population is composed of learning rate, number of neurons in convolutional layer, and batch size parameters, and the search range of each parameter is shown in Table 1. In the DMOFIMNet training phase, the learning rate, batch size, and number of neurons in the convolutional layer are optimized using the ALA-ResNet algorithm, and a large amount of online training is performed to ensure the accuracy and robustness of the model. During training, 10 , , training samples and their corresponding labels are generated under weak, medium, and strong turbulence (corresponding to atmospheric refractive index structure constants of 5 , respectively) using the Monte Carlo method, where 80% of the samples are used as the training set and 20% of the samples are used as the validation set; in addition, 10 4 training samples and their corresponding labels are generated as the test set. To fully train the DMOFIMNet, the training rounds are 600, and the difference between the original symbol and the estimated symbol is quantified using the cross-entropy loss function.

[0026] In order to better illustrate the technical effect of the present application, the effect of the present application is further described below in combination with simulation experiments. The hardware platform environment configuration of the simulation experiment of the present application is Intel(R) Core(TM) i7-14700KF @3.40GHz CPU and NVIDIA GeForce 4060Ti 8GB GPU. Figure 6 For the ALA-ResNet optimization process, the fitness value changes with the number of iterations. As can be seen from the figure, when the number of iterations is small, the fitness value shows a downward trend, indicating that the population is not mature and the algorithm has not converged to the global optimal solution. When the number of iterations exceeds 34 times, the fitness value reaches the minimum value, indicating that the population has matured and the algorithm has converged to the global optimal solution. Figure 7 The bit error rate performance of the ML detection algorithm and the DMOFIMNet detection algorithm under weak, medium and strong turbulence conditions is compared. As can be seen from the figure, the proposed DMOFIMNet detector can achieve performance close to the ML detection algorithm under weak, medium and strong atmospheric turbulence. In addition, thanks to its residual learning structure, the DMOFIMNet detection algorithm effectively alleviates the gradient vanishing problem and captures the deep feature correlation introduced by strong turbulence, so its performance under strong turbulence intensity is better than that of the ML detection algorithm. Figure 8 For comparison of the computational complexity of the DMOFIMNet detection algorithm and the ML detection algorithm, it can be seen from the figure that the computational complexity of the DMOFIMNet detection algorithm is reduced by 25% compared with the ML detection algorithm. This is because as the number of transmission frames increases, the search space of the ML detection algorithm increases exponentially, which in turn leads to an exponential increase in its computational complexity. While the computational complexity of the DMOFIMNet detection algorithm is mainly restricted by the number of network layers and the dimension of data, even in the case of a significant increase in the number of frames, its computational complexity growth trend is still relatively flat, thus exhibiting higher efficiency and robustness when processing large-scale symbol sequences.

Claims

1. An optical FBMC dual-mode index modulation communication system and a signal detection method thereof, characterized in that, Includes the following steps: Step 1: Utilize all available subcarriers to transmit both constellation modes while preserving the diversity gain of index modulation; Step 2: Optimize the dual-mode constellation by phase rotation and amplitude scaling to generate the optimal constellation for the optical FBMC system; Step 3: Design a deep learning-assisted signal detection algorithm to detect transmitted information; Step 4: Optimize the hyperparameters of the signal detection algorithm using the artificial lemming algorithm.

2. The optical FBMC dual-mode index modulation communication system and its signal detection method according to claim 1, characterized in that: In Step 1, for the number of subcarriers... Optical FBMC dual-mode indexed modulation communication system, The input bits are divided into 10 bits by a bit splitter. Each sub-block contains [number] sub-blocks. bits and Subcarriers, The bits are further divided into two parts, the index bits. and information bits ; Index bits The information bits are fed into the index selector for mapping subcarrier indices. They were sent to two different constellation mappers, through two different types of constellation modes. and Perform mapping; after mapping is complete, Combining different FBMC sub-blocks together to perform Hermitian symmetric operations, outputting... ; After by The bipolar signal is obtained by processing a polyphase network consisting of a point-inverse Fourier transform and a Martin-Mirabassi prototype filter; finally, the signal... After being superimposed with a DC bias, the laser is driven to generate an optical signal.

3. The optical FBMC dual-mode indexed modulation communication system and its signal detection method according to claim 1, characterized in that: In Step 2, a phase rotation and amplitude scaling constellation method is proposed for the optical FBMC dual-mode index modulation communication system.

4. The optical FBMC dual-mode index modulation communication system and its signal detection method according to claim 1, characterized in that: In Step 3, the deep learning-assisted signal detection method, named DMOFIMNet, consists of a preprocessing module, a ResNet module, and an output module connected in series.

5. The optical FBMC dual-mode index modulation communication system and its signal detection method according to claim 1, characterized in that: In Step 4, the artificial lemming algorithm is used for global optimization to explore the hyperparameter space and optimize the structure of DMOFIMNet.

Images