Modulation format recognition method for underwater visible light communication system based on reservoir calculation

Abstract

The present invention discloses a method for identifying the modulation format of an underwater visible light communication system based on reservoir calculation. The method utilizes the received signal at the receiving end of the system, first performs data preprocessing and feature enhancement on the received signal using a coordinate transformation algorithm and a folding algorithm, and then performs classification judgment for modulation format identification based on a reservoir calculation network. The method of the present invention can achieve low computational overhead, high real-time performance, high accuracy and strong robustness in the modulation format identification task of the underwater visible light communication system.

Description

Technical Field

[0001] The present invention relates to the field of electronic information technology, and in particular to a method for identifying a modulation format of an underwater visible light communication system based on reserve pool calculation. Background Art

[0002] Underwater Visible Light Communication (UVLC) is a new wireless communication technology that uses visible light as a carrier to transmit information through underwater channels. Compared with traditional underwater communication technologies, UVLC has advantages such as higher communication speed, stronger resistance to electromagnetic interference, and better communication stability. Therefore, it has been widely used in underwater resource development, marine science, and other special scenarios [1,2]. In the foreseeable future, UVLC should have stronger dynamic performance to meet the needs of various frequency bands and improve spectrum utilization. As one of the key technologies that directly affects the data transmission quality and rate of UVLC, modulation format recognition (MFR) has the ability to autonomously detect the modulation format of the received signal without any prior information from the transmitter, allowing the receiver to dynamically adjust the demodulation mode, enhancing the dynamic performance and frequency utilization of the communication system.

[0003] However, in the task of modulation format recognition, due to the lack of prior information from the transmitter, it is difficult to directly classify the modulation format. In the long-term development process, researchers have proposed various methods to improve the performance of modulation format recognition tasks [3]. These methods can be roughly divided into two categories: probability-based (LB) methods and feature-based (FB) methods. In the probability-based method, researchers use methods such as probability and hypothesis testing parameters to solve the modulation format recognition problem. Through appropriate assumptions and appropriate thresholds, the error probability is minimized, thereby obtaining the optimal solution in the Bayesian sense [4]. Although this method can provide the optimal solution, its computational complexity is high and it is difficult to implement in practical applications. In contrast, the feature-based method extracts significant features from the received data and then uses these features to identify the modulation format. This method is easy to implement and can obtain near-optimal modulation format recognition accuracy through appropriate feature extraction [5]. Of course, how to design an efficient feature extraction algorithm is also an important issue.

[0004] Over the past few years, with the rapid development of artificial intelligence algorithms, neural network-based algorithms have achieved remarkable success in many scientific fields. Neural network-based algorithms have also been applied to modulation format recognition tasks. Researchers have achieved good performance in various modulation format recognition tasks through a variety of deep learning algorithms, such as convolutional neural networks, recurrent neural networks, and graph neural networks [6,7]. However, due to the big data-driven nature of neural networks, in actual application scenarios, users often find it difficult to obtain sufficient massive data for adequate neural network training. In addition, the huge computing resources and computing time overhead of neural network training are often unacceptable in systems with high real-time requirements. Therefore, how to design new neural network structures, simplify network frameworks, reduce training overhead, and improve algorithm performance is a topic worthy of in-depth exploration.

[0005] [References]

[0006] 1.S.Arnon,"Underwater optical wireless communication network,"Opt.Eng49(1),015001(2010).

[0007] 2. H. Kaushal and G. Kaddoum, "Underwater Optical Wireless Communication," IEEE Access 4, 1518–1547 (2016).

[0008] 3.OADobre, A.Abdi, Y.Bar-Ness, and W.Su, "Survey of automatic modulation classification techniques: classical approaches and new trends," IETCommun.1(2),137(2007).

[0009] 4. W. Wei and JMMendel, "Maximum-likelihood classification for digitalamplitude-phase modulations," IEEE TRANSACTIONS ON COMMUNICATIONS 48(2), (2000).

[0010] 5. AKJain, RPWDuin, and J.Mao, "Statistical pattern recognition: review," IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE 22(1), (2000).

[0011] 6.FNKhan, K.Zhong, WHAl-Arashi, C.Yu, C.Lu, and APTLau, "ModulationFormat Identification in Coherent Receivers Using Deep Machine Learning," IEEEPHOTONICS TECHNOLOGY LETTERS28(17), (2016).

[0012] 7. D. Wang, M. Zhang, Z. Li, J. Li, M. Fu, Y. Cui, and X. Chen, "Modulation FormatRecognition and OSNR Estimation Using CNN-Based Deep Learning," IEEEPhoton.Technol.Lett.29(19),1667–1670(2017). Summary of the Invention

[0013] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide a modulation format identification method for underwater visible light communication systems based on reservoir computing (RC) that combines coordinate transformation and folding algorithms. The method of the present invention can achieve low computational overhead, high real-time performance, high accuracy and strong robustness in the modulation format identification task of underwater visible light communication systems.

[0014] In the present invention, digital signal processing is first used to obtain the corresponding IQ (I-phase, Q-quadrature) signal from the receiving end of underwater visible light communication. Coordinate transformation and folding algorithms are then used for efficient feature extraction, improving the accuracy of subsequent modulation format classification. Finally, a reservoir calculation is used for efficient and simple classification, resulting in the corresponding modulation format classification result. The present invention is described in detail below.

[0015] A method for identifying a modulation format of an underwater visible light communication system based on a reservoir calculation comprises the following steps:

[0016] Step 1): In the underwater visible light communication system, the data sent by the LED at the transmitting end passes through the underwater channel and is received at the receiving end, and a complex valued received signal is obtained after digital signal processing;

[0017] Step 2): Decompose the received complex-valued signal in rectangular coordinates and polar coordinates respectively through a coordinate transformation algorithm to obtain the IQ component in the rectangular coordinate system and the polar angle and diameter component in the polar coordinate system, and then aggregate the data in different coordinate systems;

[0018] Step 3): Through the folding algorithm, the feature maps corresponding to the data in the rectangular coordinate system and the polar coordinate system are folded along the symmetry axis using symmetry to reduce the range of the feature map, thereby removing redundant information and highlighting significant features;

[0019] Step 4): The data of the folded feature map obtained in step 3) is sent to the reserve pool to calculate the RC network. The ridge regression method is used to generate the output weight of the optimal solution in the least squares sense. The node states of the intermediate layer are linearly combined, and the softmax function is used to obtain the probability that the input data belongs to each modulation format. The category corresponding to the maximum probability is selected, which is the final classification decision of the modulation format.

[0020] In the present invention, in step 3), for the two-dimensional signal in the polar coordinate system, taking into account the symmetry of the characteristic graph distribution, the folding algorithm is performed according to different angles; the angle in the current polar coordinate system is recorded as Range arrive Where n is the current folding order, the order is 0 when not folded, i represents the i-th data point, then using symmetry, the symmetry axis of the current signal distribution is The symmetry axis is always the average value of the current angle range, and the folding operation is then performed along this symmetry axis:

[0021]

[0022] Taking into account the symmetry, each folding operation to the left or right of the symmetry axis is equivalent. The equivalent folding operation in the other direction is expressed as:

[0023]

[0024] In the present invention, in step 3), the folding order is 3-4 times; in step 4), the number of nodes in the middle layer of the reserve pool calculation RC network is 400-800.

[0025] In the present invention, in step 4), a nonlinear activation function is used in the reserve pool calculation RC network, and the nonlinear function is tanh, sigmoid or ReLU.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. Applying the reservoir computing network algorithm to the modulation format recognition task of underwater visible light communication systems not only achieves higher accuracy than traditional classification algorithms, but also significantly reduces the training overhead and time cost compared to deep learning-based neural network methods, thereby achieving better accuracy, real-time performance, and robustness.

[0028] 2. By introducing the coordinate transformation algorithm, the different characteristics of PSK and QAM signals are highlighted, thereby efficiently enhancing the features of the input data, highlighting local significant features, reducing feature redundancy, and combining global features with local features.

[0029] 3. By introducing the folding algorithm, the original feature map undergoes folding operations of different orders, which effectively reduces the redundancy of the input data and combines the global features with the local features, which not only reduces the computational complexity and time overhead of the algorithm, but also improves the accuracy of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a schematic diagram of the modulation format recognition technology of the underwater visible light communication system based on reservoir calculation according to the present invention.

[0031] Figure 2 This is a system diagram of the modulation format recognition technology of the underwater visible light communication system based on reservoir calculation of the present invention.

[0032] Figure 3 The coordinate transformation of the present invention is to transform the six modulation format data into rectangular coordinate system and polar coordinate system.

[0033] Figure 4 These are the transformation results of the folding algorithm of the present invention performing folding operations of different orders on six modulation format data.

[0034] Figure 5 This is a study of the debugging format recognition accuracy of the present invention as the emission voltage of the LED at the transmitting end changes.

[0035] Figure 6 This paper studies the changes in the accuracy of the algorithm of the present invention with hyperparameters such as nonlinear activation function, size of the intermediate layer of the reservoir pool, and leakage coefficient. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0037] This paper addresses the modulation format recognition task for underwater visible light communications, focusing on designing lightweight and efficient algorithms to mitigate the significant computational overhead and time consumption of common neural network algorithms, thereby meeting the requirements of high-performance real-time systems. Furthermore, given the complexity of underwater channels, the paper addresses the design of targeted feature extraction algorithms that preprocess and enhance the input data, ensuring that subsequent algorithms remain lightweight and highly accurate.

[0038] In the present invention, for the modulation format recognition task in the underwater visible light communication system, that is, using the received signal at the receiving end of the system, data preprocessing and feature enhancement are first performed, and then the classification decision of the modulation format recognition is performed. In actual scenarios, due to the complexity of the underwater channel, how to effectively extract the features of the received signal and complete the modulation format classification decision with as little time overhead and computing resources as possible has very wide application value. The present invention proposes a modulation format recognition technical solution based on feature extraction algorithm and reserve pool calculation (RC). The specific process includes: for the received signal at the receiving end of underwater visible light communication, using digital signal processing technology to obtain the corresponding IQ signal (I: in-phase, Q: quadrature), then using coordinate transformation algorithm and folding algorithm to perform efficient feature extraction to improve the accuracy of subsequent modulation format classification, and finally using reserve pool calculation to perform efficient and simple classification to obtain the corresponding modulation format classification result. We used the technical solution for modulation format identification proposed in this patent to conduct experiments on six modulation formats: OOK, 4QAM, 8QAM-DIA, 8QAM-CIR, 16APSK and 16QAM in an underwater visible light communication system. The experimental results verified the effectiveness and efficiency of the technical solution proposed in this patent.

[0039] 1. Coordinate transformation algorithm

[0040] In the underwater visible light communication system, the complex signal received by the receiver is recorded as (where N represents the length of each group of received signals). To obtain the corresponding IQ signals, consider:

[0041] Inphase=Imag[y(i)]

[0042] Quadrature=Real[y(i)] (1)

[0043] By separating the real and imaginary parts of a complex signal, we transform the received one-dimensional complex signal into a two-dimensional IQ signal. In traditional modulation format recognition algorithms, subsequent classification algorithms directly use the IQ signal as input and classify it to complete modulation format recognition. This patent considers further feature extraction from traditional IQ signals to make them easier to classify and determine modulation formats.

[0044] The traditional IQ signal can be considered as the components of the complex signal in a two-dimensional rectangular coordinate system. This decomposition method is suitable for modulation formats such as QAM signals that are evenly distributed on the x-axis and y-axis of the two-dimensional rectangular coordinate system. However, for modulation formats such as PSK with circular symmetry, the decomposition method of the two-dimensional rectangular coordinate system is not suitable. Therefore, the present invention considers introducing the decomposition of the complex signal in a two-dimensional polar coordinate system.

[0045]

[0046] θ=Atan2(Imag[y(i)],Real[y(i)]) (2)

[0047] Among them, ρ represents the radius length in polar coordinates, θ represents the angle in polar coordinates, and θ can be expressed as 0° to 360° or -180° to 180°, with no essential difference in the final result.

[0048] In polar coordinates, the circular symmetry of PSK is fully characterized, making its characteristics more prominent. This circular symmetry also reveals richer characteristics of orthogonal symmetrical signals such as QAM (just as orthogonal decomposition characterizes PSK signals). Through orthogonal decomposition in a rectangular coordinate system and circular decomposition in a polar coordinate system, the characteristics of the received complex signal are more fully characterized, which facilitates the accuracy of subsequent modulation format identification and classification.

[0049] 2. Folding algorithm

[0050] Traditional modulation format recognition algorithms consider the signal distribution within the complete constellation diagram (for example, for IQ signals in a rectangular coordinate system, the I min ≤I≤I max , Q min ≤Q≤Q max A complete constellation diagram within the initial range). The problem with this method is that the range of the feature map represented by the initial constellation map is too large, and key features such as the distance and angle from most constellation points to the origin are not prominent, which increases the difficulty of the subsequent classification algorithm to learn features and thus perform modulation format recognition. To this end, this patent considers introducing a folding algorithm to further narrow the range of the feature map by folding the initial feature map, making the key features more prominent while eliminating some redundant and repeated features. At the same time, the features reflected in the feature maps after different folding times (called folding orders) are all emphasized, and take into account the two purposes of eliminating feature redundancy and retaining prominent features to varying degrees.

[0051] Specifically, for the two-dimensional signal in the polar coordinate system, considering the symmetry of the feature map distribution, we perform the folding algorithm according to different angles. Range arrive (where n is the current folding order, the order is 0 when not folded, and i represents the i-th data point), then using symmetry, the symmetry axis of the current signal distribution is The folding operation is then performed along this axis of symmetry:

[0052]

[0053] Of course, considering the symmetry, each folding operation to the left or right of the symmetry axis is equivalent. The equivalent folding operation in the other direction can be expressed as:

[0054]

[0055] In the present invention, the initial polar coordinate angle The range is expressed as -180° to +180°. When the original feature map is folded, it is folded in accordance with the symmetry axis. Fold four times. In principle, higher-order folding operations can be performed, but in practical applications, the folded feature map will be too small, and some useful information will be lost, thus affecting the accuracy of subsequent modulation format recognition. At the same time, folding operations with fewer orders cannot fully remove redundant information in the original feature map, nor can they highlight local feature information. Therefore, in practice, it is generally more appropriate to set the folding order to 3 or 4. In different application scenarios, the specific folding order should be determined based on the trade-off between removing redundant information in the original feature map and highlighting the local information of the folded feature map.

[0056] 3. Reserve pool calculation

[0057] In recent years, with the widespread application of deep learning technology, modulation format recognition technology based on deep learning has emerged. It uses the signal from the receiving end to directly feed the data into the neural network for end-to-end training, and finally directly outputs the classification decision. However, due to the complexity of the transmission channel, in order to fully learn the characteristics of the received signal under various modulation formats, the neural network often needs to be designed with a deep and complex architecture, which leads to huge computational costs and time overhead. Running on hardware platforms with limited computing resources often meets the requirements of real-time adaptation. To this end, the present invention proposes a simple and computationally efficient classification algorithm based on reservoir calculation for modulation format recognition.

[0058] Reservoir computing (RC) is a special recurrent neural network architecture. Its network structure consists of three parts: input layer, intermediate layer, and output layer. The input layer contains multiple nodes, each corresponding to an input data, the intermediate layer is a type of recurrent neural network, and the final output layer is an adder with weights. Accordingly, the reservoir computing algorithm includes three important weights: the weight of the input layer (denoted as W in ), the connection weights between the middle layer nodes (denoted as W res ) and the output weight between the intermediate layer and the output layer (denoted as W out ).

[0059] The biggest feature of the classic reservoir computing network structure is the weight W of its input layer in and the weight W of the middle layer res are randomly generated and remain unchanged during the training process. i , i=1,2,…,K}, there are N nodes in the middle layer, and the state of each node is recorded as u(t). Then the state update method of the middle layer node is:

[0060] u(t)=f(W in ·x(t)+W res ·u(t-1)) (5)

[0061] (where f represents a nonlinear activation function, which enhances the nonlinear fitting ability of the network)

[0062] In practice, in order to enhance the dynamic performance of the reserve pool calculation, a leakage constant can be introduced to dynamically update the intermediate layer state:

[0063] u(t)=(1-α)·u(t-1)+αf(W in ·x(t)+W res ·u(t-1)) (6)

[0064] Here, α is the leakage constant, and its change can affect the dynamic performance of the reservoir calculation.

[0065] Finally, the input data enters the network through the input layer, and then goes through the dimensionality increase operation of the middle layer. Theoretically, it has linear separability in high-dimensional space. The output layer trains an adder for linear classification to obtain the final classification result. Specifically, the output {y i , i=1,2,…,L} is obtained by linear weighted combination of the intermediate layer node states u(t):

[0066] y(t)=W out ·u(t) (7)

[0067] The output layer weight W outThe training goal is to minimize the minimum mean square error between the predicted results and the actual categories:

[0068]

[0069] Among them, ∈ is the regularization coefficient, which is used to prevent overfitting of this optimization problem.

[0070] There are many ways to train the final output weights. For example, methods based on SVM and MLP can all be effectively trained. However, in order to make the structure of the reserve pool calculation more simple and efficient, the solution of the present invention adopts the ridge regression method with the optimal solution in the sense of least squares. Its closed-form solution can be expressed as:

[0071]

[0072] Therefore, for the entire reservoir computing network architecture, the weight W of the input layer in and the weight W of the middle layer res They are all randomly generated and do not require training. The only thing that needs to be trained is the output layer weight W out , and it can be directly obtained through the closed-form solution of ridge regression, which fully guarantees the portability and real-time performance of the modulation format recognition algorithm.

[0073] In addition, the present invention also conducts an in-depth study on the design of the storage pool computing network structure. in and the weight W of the middle layer res , generally generated using a random normal distribution or a random uniform distribution. In practice, good random weights can be obtained after multiple attempts. Regarding nonlinear activation functions that affect the network's nonlinear fitting ability, experiments have shown that using nonlinear activation functions significantly improves the performance of the reserve pool calculation algorithm compared to not using nonlinear activation functions. Among the many nonlinear activation functions, we selected three nonlinear activation functions: tanh, sigmoid, and ReLU for comparative analysis:

[0074]

[0075] ReLU(x)=max(0,x) (12)

[0076] Experimental results show that the ReLU activation function performs slightly worse than the other two, due to its truncation of negative values. The tanh and sigmoid nonlinear activation functions have similar performance because they share similar shapes, with rapid changes in the center and saturation on both sides, resulting in similar nonlinear fitting capabilities in the algorithm.

[0077] The size of the middle layer has the greatest impact on the complexity of the reserve pool computing network. Experiments show that as the size of the middle layer changes, the algorithm performance will increase to a certain extent. However, this slight performance increase will slow down and eventually reach saturation. At the same time, as the size of the middle layer expands, the algorithm's computational time increases exponentially. Considering the weight of algorithm accuracy and time consumption, the most appropriate size of the reserve pool middle layer is around 500. At this time, the optimal accuracy rate reaches more than 90% under various transmission voltage conditions.

[0078] To make the objectives, technical solutions, and advantages of the present invention clearer, the following describes the implementation scheme of the present invention in detail with reference to the accompanying drawings and experimental results. The modulation recognition task of underwater visible light communication based on the above module is as follows:

[0079] Step 101: The underwater visible light communication system based on QAM-CAP / APSK-CAP modulation first converts the original data into six modulation formats. After upsampling, it is divided into two IQ signals for multi-carrier transmission. Each group of subcarriers has 1024 subcarriers, and each subcarrier transmits 128 data items. The data is sent by the LED at the transmitting end, passes through the 1.2m underwater channel, and is received at the receiving end. After digital signal processing, the complex valued received signal is obtained.

[0080] Step 102: Decompose the received complex signal into rectangular coordinates and polar coordinates using a coordinate transformation algorithm to obtain the IQ component in the rectangular coordinate system and the polar angle and diameter component in the polar coordinate system. The data in different coordinate systems are then aggregated.

[0081] Step 103: Using a folding algorithm, fold the feature maps corresponding to the data in different coordinate systems to remove redundant information and highlight significant features. In the experiment, the optimal folding order is about 3 or 4;

[0082] Step 104: Flatten each set of 128 data points in two paths to obtain input data of length 256. Each set of input data is sequentially input into the reservoir computing network. The input weights between the input layer and the middle layer are randomly generated. Assuming that the number of nodes in the middle layer is 500, the input weights are a random matrix of 256×500.

[0083] Step 105: By randomly generating 500×500 intermediate layer connection weights, selecting a suitable activation function (such as tanh), and combining the input data after the transformation of the input weights, the state of the internal nodes at each time step is updated.

[0084] Step 106: Generate the output weights of the optimal solution in the least squares sense through the ridge regression method, linearly combine the node states of the intermediate layer, and use the softmax function to obtain the probability that the input data belongs to each modulation format. Select the category corresponding to the maximum probability, which is the final classification decision of the modulation format.

[0085] According to the experimental results, Figure 3 and Figure 4 The transformation results of the coordinate transformation and folding algorithms for the complex-valued information received in the underwater visible light communication system are given. These two feature enhancement algorithms significantly highlight local features while reducing feature redundancy, thereby improving the efficiency and accuracy of the algorithm.

[0086] Under the appropriate selection of reserve pool calculation hyperparameters, Figure 5 This is a study of the debugging format recognition accuracy of the present invention as the LED emission voltage at the transmitter changes. When the LED emission voltage changes from 0.1V to 1.3V, the best accuracy is almost above 90%, and the highest is close to 100%. At the same time, Figure 5 The illustration shows how the algorithm's accuracy varies with the size of the reservoir's middle layer when the transmit voltage is 0.3V. In our experiments, we chose a reservoir size of approximately 500, achieving an accuracy exceeding 98%. The algorithm also completes calculations in a few seconds, significantly reducing the computational overhead and time consumption of the modulation format recognition algorithm.

[0087] at last, Figure 6 This study investigates how the algorithm's modulation format recognition accuracy varies with hyperparameters such as the nonlinear activation function, the size of the reservoir intermediate layer, and the leakage coefficient. Different hyperparameters can affect the algorithm's accuracy by as much as 10%, so choosing the right hyperparameters can further enhance the algorithm's performance in practical modulation format recognition tasks.

[0088] In the present invention, parameters such as the generation method of random weights, the selection of nonlinear activation functions, the scale of intermediate layer nodes, and the number of folding operations can be adjusted according to actual applications to obtain optimal performance.

[0089] In summary, the present invention proposes combining reservoir computing and feature enhancement algorithms to perform modulation format recognition tasks in underwater visible light communication systems. The use of a reservoir computing network significantly reduces the algorithm's computational overhead and training time, improving its real-time performance. Furthermore, by introducing coordinate transformation and folding algorithms, efficient data enhancement can be performed on the original input data, highlighting local salient features, reducing feature redundancy, and combining global and local features, further enhancing the performance and robustness of the present invention's algorithm.

[0090] The division of the various implementation steps in the present invention is only for the purpose of clearly explaining the principles. Some steps can be combined or split during implementation. As long as they contain similar implementation principles and logical relationships, they are within the scope of protection of this patent.

[0091] Those skilled in the art will appreciate that the aforementioned embodiments are specific examples of the present invention, and that in actual applications, various modifications may be made to the embodiments and details without departing from the spirit and scope of the present invention. For example, the order of folding operations in the folding algorithm, the method for generating random weights in the reservoir calculation network structure, the size of the intermediate layers, and the selection of nonlinear activation functions can all be adjusted according to actual needs to achieve optimal results.

Claims

1. A method for identifying modulation formats of underwater visible light communication systems based on reservoir calculation, characterized in that: The following steps are involved: Step 1): In the underwater visible light communication system, the data sent by the LED at the transmitting end passes through the underwater channel and is received at the receiving end, and a complex signal is obtained after digital signal processing; Step 2): Decompose the received complex-valued signal into rectangular coordinates and polar coordinates using a coordinate transformation algorithm to obtain the IQ component in the rectangular coordinate system and the polar angle and diameter component in the polar coordinate system. Then, the data in the rectangular coordinate system and the polar coordinate system are concatenated to achieve aggregation of data features in different coordinate systems. Step 3): Through the folding algorithm, the feature maps corresponding to the data in the rectangular coordinate system and the polar coordinate system are folded along the symmetry axis using symmetry to reduce the range of the feature map, thereby removing redundant information and highlighting significant features; Step 4): The data of the folded feature map obtained in step 3) is sent to the reserve pool to calculate the RC network. The ridge regression method is used to generate the output weights of the optimal solution in the least squares sense. The node states of the intermediate layer are linearly combined, and the softmax function is used to obtain the probability that the input data belongs to each modulation format. The category corresponding to the maximum probability is selected, which is the final classification decision of the modulation format of the underwater visible light communication system.

2. The method for identifying the modulation format of an underwater visible light communication system according to claim 1, wherein: In step 3), for the two-dimensional signal in the polar coordinate system, considering the symmetry of the feature map distribution, the folding algorithm is performed at different angles; the angle in the current polar coordinate system is recorded as Range arrive Where n is the current folding order, the order is 0 when not folded, i represents the i-th data point, then using symmetry, the symmetry axis of the current signal distribution is The symmetry axis is always the average value of the current angle range, and the folding operation is then performed along this symmetry axis: Taking into account the symmetry, each folding operation to the left or right of the symmetry axis is equivalent. The equivalent folding operation in the other direction is expressed as:

3. The method for identifying the modulation format of an underwater visible light communication system according to claim 1, wherein: In step 3), the folding order is 3-4 times. In step 4), the number of nodes in the middle layer of the reserve pool calculation RC network is 400-800.

4. The method for identifying the modulation format of an underwater visible light communication system according to claim 1, wherein: In step 4), a nonlinear activation function is used in the reserve pool calculation RC network, and the nonlinear function is tanh, sigmoid or ReLU.

5. The method for identifying the modulation format of an underwater visible light communication system according to claim 1, wherein: The modulation formats of underwater visible light communication systems include OOK, 4QAM, 8QAM-DIA, 8QAM-CIR, 16APSK and 16QAM.

Images