Modulation format and optical signal-to-noise ratio joint monitoring method based on amplitude analysis complex plane

Abstract

The invention discloses a modulation format and optical signal-to-noise ratio joint monitoring method based on an amplitude analysis complex plane, which relates to the field of communication and comprises the following steps of: performing constant modulus equalization processing on a signal to be identified to obtain a preprocessed digital signal; generating an amplitude analysis complex plane based on the preprocessed digital signal; and performing modulation format and optical signal-to-noise ratio identification on the amplitude analysis complex plane through the trained multi-target neural network. According to the invention, through the multi-target neural network, modulation format identification and optical signal-to-noise ratio monitoring can be carried out at the same time, and bottom-layer feature extraction is shared, so that redundant calculation is avoided, the complexity of the scheme is reduced, and the joint identification efficiency is improved. Through verification, the method can realize 100% correct recognition rate of five modulation formats when the optical signal-to-noise ratio is not lower than a corresponding threshold value of 20% FEC, and the maximum monitoring error of the optical signal-to-noise ratio of the five modulation formats is controlled within 1dB.

Description

Technical Field

[0001] The present invention relates to the field of communications, and in particular to a modulation format and optical signal-to-noise ratio joint monitoring method based on an amplitude-analyzed complex plane. Background Art

[0002] Elastic Optical Networks (EONs) can better address the inefficient spectrum resource utilization of traditional wavelength division multiplexing systems. Their flexible spectrum allocation and parameter adjustment capabilities further enable digital and dynamic networks. Furthermore, EON nodes and receivers must implement optical performance monitoring (OPM) to obtain information such as signal dispersion, modulation format, and optical signal-to-noise ratio (OSNR) to ensure efficient EON operation.

[0003] In recent years, researchers have broadly categorized research on OPM in EONs into two categories: single-parameter monitoring and joint multi-parameter monitoring. Single-parameter monitoring schemes, such as improved principal component analysis (PCA) algorithms, trajectory information, and particle swarm optimization (PSO), can identify modulation format (MFI). Training sequences and delayed tap-based sampling techniques are used for dispersion monitoring. Random forests and support vector machines (SVMs) are used for optical signal-to-noise ratio (OSNR) estimation. However, these schemes only monitor a single parameter; monitoring other parameters requires selecting new features or retraining them, which incurs higher computational complexity and time delays. To address this issue, joint multi-parameter monitoring has garnered increasing attention. Deep neural networks (DNNs) have been widely used in EON OPM due to their high monitoring accuracy and ability to autonomously learn features in dynamic, high-volume, and complex data environments. Some techniques have proposed using features such as constellation diagrams, spectra, intensity and differential phase images, and amplitude histograms as inputs to multi-task DNNs to achieve modulation format identification and OSNR estimation. However, with the continuous development of EONs, the requirements for convergence speed and computational complexity are becoming increasingly stringent. Therefore, it is necessary to study the joint multi-parameter OPM scheme suitable for high-order modulation formats (such as 128QAM) in the next-generation EON. Summary of the Invention

[0004] In view of the above-mentioned deficiencies in the prior art, the modulation format and optical signal-to-noise ratio joint monitoring method based on amplitude-resolved complex plane provided by the present invention solves the problems that the prior methods are difficult to take into account 128QAM signal recognition and have low recognition speed.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A method for jointly monitoring modulation format and optical signal-to-noise ratio based on an amplitude-resolved complex plane is provided, which comprises the following steps: Perform CMA equalization processing on the signal to be identified to obtain a pre-processed digital signal; generating an amplitude-analyzed complex plane based on the preprocessed digital signal; Perform preliminary feature extraction and normalization on the amplitude-resolved complex plane through the initial convolution module; Reduce the spatial dimension of the features output by the initial convolutional module through the maximum pooling layer; The features output by the maximum pooling layer are reused and fused at multiple levels through the feature extraction module; The transition layer compresses the dimension of the features output by the feature extraction module and reduces the spatial resolution of the features output by the feature extraction module; The spatial dimension of the feature map output by the transition layer is reduced to 1×1 through a global average pooling layer while keeping the channel dimension unchanged; The output layer jointly monitors the modulation format and optical signal-to-noise ratio based on the features output by the global average pooling layer, and outputs the detection results.

[0006] Furthermore, a specific method for generating an amplitude-analyzed complex plane based on the preprocessed digital signal includes the following steps: Normalize the power of the preprocessed digital signal and extract the amplitude information; The amplitude information is parsed to obtain a two-dimensional plane graph; The obtained two-dimensional plane is plotted as a heat map to obtain the amplitude-resolved complex plane.

[0007] Furthermore, the expression of power normalization is:

[0008] in Represents the mth level amplitude value of the preprocessed digital signal; represents the correlation probability of the mth level of the preprocessed digital signal; M represents the total number of levels of the modulation format.

[0009] Furthermore, the specific method of parsing the amplitude information to obtain a two-dimensional plane graph includes the following steps: According to the amplitude information, the pre-processed digital signal is subjected to discrete Fourier transform to obtain the frequency domain signal, which is expressed as follows:

[0010] in represents the frequency domain signal; Represents the preprocessed digital signal; is a natural constant; N is the number of symbols; k is the frequency domain index; j is the imaginary unit; is pi; n is the time domain index; For frequency domain signals Perform frequency domain processing, and the expression is:

[0011] in is the signal after frequency domain processing; The signal after frequency domain processing Perform inverse Fourier transform, the expression is:

[0012] in is the time domain signal obtained by inverse Fourier transform; The time domain signal obtained by inverse Fourier transform Together with the amplitude information, a two-dimensional graph is formed, that is, a two-dimensional plane graph is obtained.

[0013] Furthermore, the initial convolution module includes a Conv2D layer, a BatchNorm layer, and a ReLU activation layer connected in sequence.

[0014] Furthermore, the feature extraction module adopts a fully connected structure constructed with a multi-stage gated aggregation mechanism, including a first branch, a second branch and a feature fusion module, where: The first branch consists of the following connected in sequence: The first convolution submodule is used to adjust the channel dimension through 1×1 convolution and output the features after channel adjustment; The first activation submodule is used to perform SiLU activation on the output of the first convolution submodule and output the activated features; The second branch consists of the following connected in sequence: The multi-stage convolution submodule includes parallel depth-wise separable convolution groups, which extract features of different scales with different dilation rates and convolution kernel sizes, concatenate the features of different scales and output multi-scale features; The second convolution submodule is used to adjust the channel dimension through 1×1 convolution and output the features after channel adjustment; The second activation submodule is used to perform SiLU activation on the output of the second convolution submodule and output the activated features; The feature fusion module is used to add and fuse the outputs of the first branch and the second branch, perform 1×1 convolution on the fusion result to adjust the channel dimension, and output the final feature; The input of the first convolution submodule and the input of the multi-order convolution submodule are both inputs of the feature extraction module, and the output of the feature fusion module is the output of the feature extraction module.

[0015] Furthermore, the transition layer includes a batch normalization unit, a ReLU activation unit, a 1×1 convolution unit, and an average pooling operation unit connected in sequence.

[0016] Furthermore, in the output layer, the joint monitoring of modulation format and optical signal-to-noise ratio includes modulation format recognition and optical signal-to-noise ratio recognition, wherein modulation format recognition is classification recognition and is performed in the form of a classification task; optical signal-to-noise ratio recognition is regression recognition and is performed in the form of a regression task.

[0017] Furthermore, the signals to be identified include QPSK signals, 16QAM signals, 32QAM signals, 64QAM signals and 128QAM signals.

[0018] Furthermore, the initial convolution module, the maximum pooling layer, the feature extraction module, the transition layer, the global average pooling layer and the output layer constitute a multi-objective neural network. The training process of the multi-objective neural network includes the following steps: The amplitude-analyzed complex planes corresponding to QPSK, 16QAM, 32QAM, 64QAM, and 128QAM signals with known optical signal-to-noise ratios are obtained and used as training samples. The optical signal-to-noise ratio range of the QPSK signal is 7-26dB; the optical signal-to-noise ratio range of the 16QAM signal is 15-33dB; the optical signal-to-noise ratio range of the 32QAM signal is 18-37dB; the optical signal-to-noise ratio range of the 64QAM signal is 19-38dB; and the optical signal-to-noise ratio range of the 128QAM signal is 23-40dB. The actual optical signal-to-noise ratio value and modulation format corresponding to the training samples are used as labels. The multi-target neural network is trained through the training samples, and the parameters of the multi-target neural network are updated by backpropagation until the recognition accuracy of the multi-target neural network reaches the set value.

[0019] The beneficial effects of the present invention are: 1. This method can simultaneously achieve accurate recognition of the modulation format and precise estimation of the optical signal-to-noise ratio based on the characteristic differences of different modulation formats and different optical signal-to-noise ratio values ​​on the amplitude-resolved complex plane. It does not require multiple neural networks to realize modulation format recognition and optical signal-to-noise ratio monitoring respectively, which reduces the complexity and time required for the solution and improves the efficiency of the joint recognition of modulation format and optical signal-to-noise ratio.

[0020] 2. This method uses the amplitude-analyzed complex plane as input features, which enhances the signal discrimination and facilitates the subsequent multi-objective neural network training and recognition of modulation format and optical signal-to-noise ratio information. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the process of this method; Figure 2 Schematic diagram of coherent optical communication system simulation in the embodiment; Figure 3 for Figure 2 The architecture diagram of the digital signal processing module; Figure 4 Resolve the complex plane for the amplitude of the QPSK signal; Figure 5 The complex plane of the amplitude resolution of the 16QAM signal; Figure 6 The complex plane of the amplitude resolution of the 32QAM signal; Figure 7 The complex plane of the amplitude resolution of the 64QAM signal; Figure 8 The complex plane of the amplitude analysis of the 128QAM signal; Figure 9 Figure 2 is a diagram of a multi-objective neural network architecture; (a) is the multi-objective neural network architecture, and (b) is the network architecture of the feature extraction layer; Figure 10 This is a graph showing how the modulation format recognition accuracy changes with the optical signal-to-noise ratio; Figure 11 Schematic diagrams showing the relationship between the true optical signal-to-noise ratio (OSNR) and the estimated OSNR values ​​for the five signals in the embodiment; (a) shows the relationship between the true OSNR value and the estimated OSNR value for a QPSK signal; (b) shows the relationship between the true OSNR value and the estimated OSNR value for a 16QAM signal; (c) shows the relationship between the true OSNR value and the estimated OSNR value for a 32QAM signal; (d) shows the relationship between the true OSNR value and the estimated OSNR value for a 64QAM signal; and (e) shows the relationship between the true OSNR value and the estimated OSNR value for a 128QAM signal. Figure 12 A performance comparison diagram is shown in FIG. DETAILED DESCRIPTION

[0022] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0023] like Figure 1 and Figure 9 As shown, the modulation format and optical signal-to-noise ratio joint monitoring method based on the amplitude-resolved complex plane includes the following steps: S1, performing CMA equalization processing on the signal to be identified to obtain a pre-processed digital signal; S2, generates an amplitude-analyzed complex plane based on the preprocessed digital signal, with a size of 28×28×3; S3, perform preliminary feature extraction and normalization on the amplitude-resolved complex plane through the initial convolution module; the output feature size of the initial convolution module is 14×14×64; S4, using the Max Pooling layer to reduce the spatial dimension of the features output by the initial convolutional module; the output feature size of the Max Pooling layer is 7×7×64; S5, through the feature extraction module (Feature block1), the features output by the maximum pooling layer are reused and fused at multiple levels; the output feature size of the feature extraction module is 7×7×128; S6, compressing the dimension of the features output by the feature extraction module through the transition layer (Translayer1) and reducing the spatial resolution of the features output by the feature extraction module; the output feature size of the transition layer is 3×3×64; S7, through the global average pooling layer (Avg Pool Layer), the spatial dimension of the feature map output by the transition layer is reduced to 1×1, while keeping the channel dimension unchanged; the output feature size of the global average pooling layer is 1×1×64; S8. The output layer jointly monitors the modulation format and optical signal-to-noise ratio based on the features output by the global average pooling layer, and outputs the detection results.

[0024] The initial convolution module, maximum pooling layer, feature extraction module, transition layer, global average pooling layer and output layer constitute the multi-objective neural network.

[0025] In this embodiment, the specific method of generating the amplitude analytical complex plane based on the preprocessed digital signal in step S2 includes the following steps: S2-1, normalizing the power of the pre-processed digital signal to extract amplitude information; S2-2, analyzing the amplitude information to obtain a two-dimensional plane graph; S2-3. The obtained two-dimensional plane graph is plotted as a heat map to obtain an amplitude-analyzed complex plane.

[0026] The expression for power normalization is:

[0027] in Represents the mth level amplitude value of the preprocessed digital signal; represents the correlation probability of the mth level of the preprocessed digital signal; M represents the total number of levels of the modulation format.

[0028] In this embodiment, the specific method of parsing the amplitude information in step S2-2 to obtain the two-dimensional plane graph includes the following steps: S2-2-1. Perform discrete Fourier transform on the pre-processed digital signal according to the amplitude information to obtain the frequency domain signal, which is expressed as:

[0029] in represents the frequency domain signal; Represents the preprocessed digital signal; is a natural constant; N is the number of symbols; k is the frequency domain index; j is the imaginary unit; is the circumference of the circle; n is the time domain index; although k and n have different meanings, since the discrete Fourier transform is mapped from a time domain signal of length N to a frequency domain signal of length N, their value ranges are the same, both 0 to ; S2-2-2. Frequency domain signal Perform frequency domain processing, and the expression is:

[0030] in is the signal after frequency domain processing; S2-2-3. The signal after frequency domain processing Perform inverse Fourier transform, the expression is:

[0031] in is the time domain signal obtained by inverse Fourier transform; S2-2-4. The time domain signal obtained by inverse Fourier transform Together with the amplitude information, a two-dimensional graph is formed, that is, a two-dimensional plane graph is obtained.

[0032] This operation normalizes the signal amplitude information to a specific range, dividing the image into 80×80 squares. Each square contains the parsed signal amplitude information and the number of symbols within the square. However, once the number of symbols is determined, the symbol count within the interval varies depending on the modulation format and the signal's optical signal-to-noise ratio (OSNR). QPSK signals have a single-level amplitude distribution, with the total number of symbols concentrated within the corresponding square. 16QAM signals have a three-level amplitude distribution, with the total number of symbols concentrated within the corresponding square. 32QAM signals have a five-level amplitude distribution, with the total number of symbols concentrated within the corresponding square. 64QAM signals have a nine-level amplitude distribution, with the total number of symbols concentrated within the corresponding square. 128QAM signals have a 13-level amplitude distribution, with the total number of symbols concentrated within the corresponding square. While the total number of symbols is fixed, the amplitude distribution of each modulation format varies, and the number of grids corresponding to each distribution level is also different. Therefore, the number of symbols allocated to each grid is also different, laying the foundation for subsequent identification.

[0033] In this embodiment, in order to achieve preliminary feature extraction and normalization processing on the amplitude-resolved complex plane, the initial convolution module includes a Conv2D layer (Conv Layer), a BatchNorm layer, and a ReLU activation layer connected in sequence.

[0034] In this embodiment, the feature extraction module adopts a fully connected structure constructed with a multi-stage gated aggregation mechanism, including a first branch, a second branch and a feature fusion module, wherein: The first branch consists of the following connected in sequence: The first convolution submodule (Conv 1*1) is used to adjust the channel dimension through 1×1 convolution and output the features after channel adjustment; The first activation submodule (SiLU) is used to perform SiLU activation on the output of the first convolution submodule and output the activated features; The second branch consists of the following connected in sequence: The multi-level convolution submodule (DWConv Block) contains parallel depthwise separable convolution groups, which extract features of different scales with different dilation rates and convolution kernel sizes, concatenate the features of different scales and output multi-scale features; The second convolution submodule (Conv 1*1) is used to adjust the channel dimension through 1×1 convolution and output the features after channel adjustment; The second activation submodule (SiLU) is used to perform SiLU activation on the output of the second convolution submodule and output the activated features; The feature fusion module is used to add and fuse the outputs of the first branch and the second branch, perform 1×1 convolution on the fusion result to adjust the channel dimension, and output the final feature; The input of the first convolution submodule and the input of the multi-order convolution submodule are both inputs of the feature extraction module, and the output of the feature fusion module is the output of the feature extraction module.

[0035] In this embodiment, the transition layer includes a batch normalization unit, a ReLU activation unit, a 1×1 convolution unit, and an average pooling operation unit that are connected in sequence.

[0036] In this embodiment, joint monitoring of modulation format and optical signal-to-noise ratio (OSNR) includes modulation format identification and OSNR identification. Modulation format identification is performed as a classification task, while OSNR identification is performed as a regression task. Specifically, the output layer comprises a classifier (FC1) and a regressor (FC2) connected in parallel. The classifier is used to implement modulation format identification (classification task), while the regressor is used to implement OSNR identification (regression task). Signals to be identified include QPSK, 16QAM, 32QAM, 64QAM, and 128QAM signals.

[0037] In this embodiment, the training process of the multi-objective neural network includes the following steps: A1. Obtain the amplitude-analyzed complex planes corresponding to QPSK, 16QAM, 32QAM, 64QAM, and 128QAM signals under known optical signal-to-noise ratios (OSNRs) and use them as training samples. The OSNR range for QPSK signals is 7-26 dB; the OSNR range for 16QAM signals is 15-33 dB; the OSNR range for 32QAM signals is 18-37 dB; the OSNR range for 64QAM signals is 19-38 dB; and the OSNR range for 128QAM signals is 23-40 dB. A2. The actual optical signal-to-noise ratio value and modulation format corresponding to the training sample are used as labels. The multi-target neural network is trained through the training sample, and the parameters of the multi-target neural network are updated by backpropagation until the recognition accuracy of the multi-target neural network reaches the set value.

[0038] In the specific implementation process, in order to control the number of training samples, this embodiment selects the optical signal-to-noise ratio added to each signal at an interval of 1 dB. That is, the five signals contain optical signal-to-noise ratio values ​​of 20, 19, 20, 20, and 18 respectively. The total number of symbols corresponding to each feature map is 8000. A certain optical signal-to-noise ratio value of each modulation format contains 100 sets of feature maps as samples, so a total of 9700 sets of samples are obtained. The 9700 groups of samples were divided into training samples and validation samples in a ratio of 8:2. The classification task labels of the five modulation format signals were set to '0', '1', '2', '3' and '4', respectively. The regression task labels were set to the true optical signal-to-noise ratio values ​​corresponding to each modulation format signal. The multi-target neural network was trained and its parameters were updated using backpropagation until the recognition accuracy of the multi-target neural network reached the set value.

[0039] In order to verify the feasibility of this method, a numerical simulation of 28GBaud polarization multiplexed QPSK / 16QAM / 32QAM / 64QAM / 128QAM signal transmission was carried out. The simulation block diagram of the 28GBaud coherent optical communication system is shown in the figure. Figure 2As shown. At the transmitting end, the wavelength of the laser is 1550nm, the linewidth is 100kHz, I and Q are the in-phase component and the orthogonal component respectively, and the electrical signal drives the IQ modulator to generate a QPSK / 16QAM / 32QAM / 64QAM / 128QAM signal with a transmission rate of 28GBaud, and is converted into a polarization multiplexed signal through a polarization combiner to enter the transmission channel. The optical signal-to-noise ratio setting module can adjust the optical signal-to-noise ratio value of the signal. At the receiving end, the out-of-band noise of the optical signal is first filtered out by an optical bandpass filter, and then the optical signal is divided into two paths by a polarization beam splitter, and the two local oscillator optical signals separated from the local oscillator laser are mixed at 90° respectively. The mixed signal is converted into an electrical signal by a balanced detector, and the electrical signal is filtered based on a low-pass filter. Finally, the electrical signal is digitized by an analog-to-digital converter for subsequent digital signal processing. As shown Figure 3 As shown, Ex and Ey are respectively the mutually orthogonal x-polarization and y-polarization signals on the same wavelength received by the coherent optical receiver. The entire digital signal processing part includes three types of algorithms: modulation format non-related algorithm, this method and modulation format related algorithm, among which the modulation format non-related algorithm and the modulation format related algorithm are both existing algorithms and are not the creativity of this application.

[0040] In this embodiment, the amplitude analysis complex plane of QPSK / 16QAM / 32QAM / 64QAM / 128QAM signals are respectively as follows: Figure 4 、 Figure 5 、 Figure 6 、 Figure 7 and Figure 8 shown.

[0041] The modulation format recognition accuracy of this method under different optical signal-to-noise ratios is as follows: Figure 10 As shown in the figure, the vertical axis is the correct recognition rate (%), and the horizontal axis is the optical signal-to-noise ratio (dB). The optical signal-to-noise ratio ranges of QPSK, 16QAM, 32QAM, 64QAM, and 128QAM signals are 7~26dB, 15~33dB, 18~37dB, 19~38dB, and 23~40dB, respectively, with an interval of 1dB. For 28GBaud polarization-multiplexed QPSK / 16QAM / 32QAM / 64QAM / 128QAM signals, a 100% correct recognition rate was achieved over the entire optical signal-to-noise ratio range. The lowest optical signal-to-noise ratios of 7dB, 15dB, 18dB, 19dB, and 23dB are all lower than the optical signal-to-noise ratio threshold corresponding to 20% forward error correction (FEC) for each modulation format (corresponding to the 20% forward error correction (FEC) threshold). Figure 10 The vertical dashed line of the same color in the image has a bit error rate (BER) of 2.4×10 -2 ).

[0042] The relationship between the true optical signal-to-noise ratio values ​​and the estimated optical signal-to-noise ratio values ​​of the five signals proposed by this method is as follows: Figure 11 As shown in the figure, the reference line for the true OSNR value is depicted by a black diagonal line, the horizontal axis represents the true OSNR value, and the vertical axis represents the estimated OSNR value. It can be seen that when the OSNR value is low or high, the estimation error is slightly larger than that of the intermediate OSNR value, and the estimation error of the high-order modulation format is larger than that of the low-order modulation format. This is because at low OSNR, the large fluctuations in the features caused by noise increase the difficulty of OSNR estimation. At high OSNR, the influence of noise is greatly reduced, and the differentiation of the features of adjacent OSNR values ​​is significantly reduced, which increases the difficulty of OSNR estimation.

[0043] To further verify the complexity of our solution, we compared this method with four typical neural network models (DenseNet, ResNet, MobileNet, and VGG-like). The results are shown in Table 1. To ensure a fair comparison, all models maintained the same input image size (28×28) and number of symbols (8000). As shown in Table 1, this method contains only 0.04M parameters and 112.41M FLOPs. In comparison, DenseNet and ResNet have 38.39M and 23.52M parameters, respectively, and their computational complexity is significantly higher. Even the lightweight MobileNet and VGG-like models have significantly higher parameter and computational complexity than this method. Figure 12 The paper further demonstrates the performance of the five schemes in modulation format recognition accuracy and optical signal-to-noise ratio monitoring. It can be observed that while all models achieved high recognition rates (near or equal to 100%) in the modulation format recognition task, in the optical signal-to-noise ratio monitoring task, this method exhibited lower mean absolute error (MAE) across all modulation formats, with a particularly significant performance advantage under the more complex 128QAM. This demonstrates that while significantly reducing model complexity, this method still maintains high recognition accuracy and superior monitoring performance, making it particularly suitable for optical network scenarios with limited computing resources, demonstrating promising practical potential.

[0044] Table 1: Performance comparison

[0045] In summary, the present invention, through a specially designed multi-objective neural network, can simultaneously achieve accurate recognition of the modulation format and precise estimation of the optical signal-to-noise ratio based on the characteristic differences of different modulation formats and different optical signal-to-noise ratio values ​​on the amplitude-resolved complex plane. It does not require multiple neural networks to realize modulation format recognition and optical signal-to-noise ratio monitoring respectively, which reduces the complexity and time required for the solution and improves the efficiency of the joint recognition of modulation format and optical signal-to-noise ratio.

Claims

1. A method for joint monitoring of modulation format and optical signal-to-noise ratio based on amplitude-resolved complex plane, characterized in that: The following steps are involved: Perform CMA equalization processing on the signal to be identified to obtain a pre-processed digital signal; generating an amplitude-analyzed complex plane based on the preprocessed digital signal; Perform preliminary feature extraction and normalization on the amplitude-resolved complex plane through the initial convolution module; Reduce the spatial dimension of the features output by the initial convolutional module through the maximum pooling layer; The features output by the maximum pooling layer are reused and fused at multiple levels through the feature extraction module; The transition layer compresses the dimension of the features output by the feature extraction module and reduces the spatial resolution of the features output by the feature extraction module; The spatial dimension of the feature map output by the transition layer is reduced to 1×1 through a global average pooling layer while keeping the channel dimension unchanged; The output layer jointly monitors the modulation format and optical signal-to-noise ratio based on the features output by the global average pooling layer, and outputs the detection results.

2. The method according to claim 1, characterized in that The specific method for generating an amplitude-analyzed complex plane based on the preprocessed digital signal includes the following steps: Normalize the power of the preprocessed digital signal and extract the amplitude information; The amplitude information is parsed to obtain a two-dimensional plane graph; The obtained two-dimensional plane is plotted as a heat map to obtain the amplitude-resolved complex plane.

3. The method according to claim 2, characterized in that The expression for power normalization is: in Represents the mth level amplitude value of the preprocessed digital signal; represents the correlation probability of the mth level of the preprocessed digital signal; M represents the total number of levels of the modulation format.

4. The method according to claim 2, characterized in that The specific method of parsing the amplitude information to obtain a two-dimensional plane graph includes the following steps: According to the amplitude information, the pre-processed digital signal is subjected to discrete Fourier transform to obtain the frequency domain signal, which is expressed as follows: in represents the frequency domain signal; Represents the preprocessed digital signal; is a natural constant; N is the number of symbols; k is the frequency domain index; j is the imaginary unit; is pi; n is the time domain index; For frequency domain signals Perform frequency domain processing, and the expression is: in is the signal after frequency domain processing; The signal after frequency domain processing Perform inverse Fourier transform, the expression is: in is the time domain signal obtained by inverse Fourier transform; The time domain signal obtained by inverse Fourier transform Together with the amplitude information, a two-dimensional graph is formed, that is, a two-dimensional plane graph is obtained.

5. The method according to claim 1, wherein The initial convolutional module consists of a Conv2D layer, a BatchNorm layer, and a ReLU activation layer connected sequentially.

6. The method according to claim 1, characterized in that The feature extraction module adopts a fully connected structure constructed with a multi-stage gated aggregation mechanism, including the first branch, the second branch and the feature fusion module, where: The first branch consists of the following connected in sequence: The first convolution submodule is used to adjust the channel dimension through 1×1 convolution and output the features after channel adjustment; The first activation submodule is used to perform SiLU activation on the output of the first convolution submodule and output the activated features; The second branch consists of the following connected in sequence: The multi-stage convolution submodule includes parallel depth-wise separable convolution groups, which extract features of different scales with different dilation rates and convolution kernel sizes, concatenate the features of different scales and output multi-scale features; The second convolution submodule is used to adjust the channel dimension through 1×1 convolution and output the features after channel adjustment; The second activation submodule is used to perform SiLU activation on the output of the second convolution submodule and output the activated features; The feature fusion module is used to add and fuse the outputs of the first branch and the second branch, perform 1×1 convolution on the fusion result to adjust the channel dimension, and output the final feature; The input of the first convolution submodule and the input of the multi-order convolution submodule are both inputs of the feature extraction module, and the output of the feature fusion module is the output of the feature extraction module.

7. The method according to claim 1, characterized in that The transition layer consists of a batch normalization unit, a ReLU activation unit, a 1×1 convolution unit, and an average pooling operation unit connected sequentially.

8. The method according to claim 1, characterized in that In the output layer, the joint monitoring of modulation format and optical signal-to-noise ratio includes modulation format recognition and optical signal-to-noise ratio recognition, wherein modulation format recognition is classification recognition and is performed in the form of a classification task; optical signal-to-noise ratio recognition is regression recognition and is performed in the form of a regression task.

9. The method according to claim 1, characterized in that The signals to be identified include QPSK signals, 16QAM signals, 32QAM signals, 64QAM signals and 128QAM signals.

10. The method according to claim 1, characterized in that The initial convolution module, maximum pooling layer, feature extraction module, transition layer, global average pooling layer and output layer constitute a multi-objective neural network. The training process of the multi-objective neural network includes the following steps: The amplitude-analyzed complex planes corresponding to QPSK, 16QAM, 32QAM, 64QAM, and 128QAM signals with known optical signal-to-noise ratios are obtained and used as training samples. The optical signal-to-noise ratio range of the QPSK signal is 7-26dB; the optical signal-to-noise ratio range of the 16QAM signal is 15-33dB; the optical signal-to-noise ratio range of the 32QAM signal is 18-37dB; the optical signal-to-noise ratio range of the 64QAM signal is 19-38dB; and the optical signal-to-noise ratio range of the 128QAM signal is 23-40dB. The actual optical signal-to-noise ratio value and modulation format corresponding to the training samples are used as labels. The multi-target neural network is trained through the training samples, and the parameters of the multi-target neural network are updated by backpropagation until the recognition accuracy of the multi-target neural network reaches the set value.

Images