Optical network multi-optical-path transmission quality prediction method based on deep learning

Abstract

The invention provides an optical network multi-optical-path transmission quality prediction method based on deep learning, and relates to the field of optical network transmission. The method comprises the following steps: acquiring and processing optical network state data to obtain a training set; constructing an MAAResU-Net model based on a residual neural network, an adaptive mask attention mechanism and a U-shaped symmetric network, and training the constructed MAAResU-Net model by using the training set; in the training process, after optical network state data features are deeply extracted through a residual neural network, unopened optical paths are adaptively filtered by applying an adaptive mask attention mechanism, the importance of different features and the dependency relationship between links and channels are calculated, and the optical network state data features are obtained. After a U-shaped symmetric network is adopted to perform up-sampling and residual fusion to recover a feature space, an interval regression strategy is adopted to perform QoT prediction of each light path; and predicting the QoT index value of the light path by using the trained MAAResU-Net model. Therefore, the multi-optical-path QoT high-precision prediction of the optical network can be realized.

Description

Technical Field

[0001] The present invention relates to the field of optical network transmission, and particularly to a method for predicting the transmission quality of multiple optical paths in an optical network based on deep learning. Background Art

[0002] As the backbone support technology of modern communication networks, optical networks are facing huge challenges brought about by the rapid development of mobile networks. Current Wavelength Division Multiplexing (WDM) optical networks and Elastic Optical Networks (EONs) use fixed margins to cope with signal degradation, which results in a waste of bandwidth resources. Accurate estimation of Transmission Quality (QoT) can reduce network margins and improve spectrum utilization. It is expected that in the future 6G era, the number of terminal connections will reach hundreds of billions, and the monthly average traffic will reach trillions of GB. This not only poses extremely high requirements for the transmission capacity of optical networks but also makes traditional optical network technologies (such as Intensity Modulation / Direct Detection, IMDD) difficult to meet the needs of future traffic growth due to low spectral efficiency. To improve spectral efficiency and transmission performance, coherent optical communication technology has gradually become the mainstream. Coherent optical communication technology uses coherent detection and Digital Signal Processing (DSP) technology to support high-order modulation formats and compensation for fiber dispersion. However, coherent optical communication also brings new challenges. The signal is subject to cumulative damages such as non-linear effects (such as Four-Wave Mixing, FWM, and Cross-Phase Modulation, XPM), Amplified Spontaneous Emission (ASE) noise, and interference between high-power wavelengths during transmission, which will lead to the deterioration of Transmission Quality (QoT). The fixed receiving sensitivity of the receiver and digital signal processing algorithms require the signal quality at the receiving end to meet certain signal quality thresholds (such as Forward Error Correction, FEC, signal-to-noise ratio, bit error rate, etc.) to achieve error-free reception. Therefore, it is necessary to predict QoT before optical path deployment to ensure the transmission quality of optical network signals.

[0003] In recent years, machine learning technology has been increasingly applied in optical communication networks and has become an effective tool for solving problems in the field of optical communication. Machine learning can extract useful information and features from complex network data and improve the performance and efficiency of optical communication systems. In the prediction of optical network QoT, machine learning technology has received extensive attention due to its strong fitting ability and prediction speed after training. By collecting optical path data information from the physical layer of the optical network, learning how to obtain the corresponding QoT value from the physical layer state information of the optical path through the established optical path information, and being able to apply the model to predict the QoT values of other unestablished optical paths, it provides a reliable basis for optical path planning and resource allocation, and promotes the optical communication network towards a new era of intelligence.

[0004] The existing optical network transmission quality (QoT) prediction technologies are mainly divided into traditional methods based on mathematical models and intelligent prediction methods based on machine learning. The traditional QoT prediction methods based on mathematical models mainly perform physical modeling on the signal transmission process, and obtain the predicted value of the received signal by analyzing the damage caused by the physical model to the signal. It can be divided into two categories. One is the split-step Fourier method with high accuracy, but its computational complexity is extremely high and it cannot be extended to large network topologies. The other is the approximate formula model (such as the Gaussian noise model), whose calculation speed has been greatly improved. However, in order to compensate for the model inaccuracy caused by violating assumptions and parameter uncertainties, a very high link margin is introduced, resulting in low network resource utilization.

[0005] The intelligent prediction methods based on machine learning are mainly divided into two ideas. One is to use the classification algorithm of machine learning to construct a QoT classification model to judge whether the optical path of the requested connection meets the threshold conditions for establishing a connection. The other is to use the regression algorithm of machine learning to construct a QoT regression model to predict the QoT value (such as OSNR, BER, Q-factor, etc.) according to the relevant characteristics of the optical path of the requested connection.

[0006] From the perspective of the applicable scope, the intelligent prediction methods based on machine learning are mainly divided into three categories: single optical path QoT prediction, multi-channel QoT prediction, and multi-optical path QoT (i.e., QoT of multiple links and multiple channels) prediction. Single optical path QoT prediction mainly focuses on the transmission quality of newly deployed end-to-end optical paths. It can not only use the classification model to predict whether the optical path can establish a connection, but also apply the regression model to obtain the specific QoT index value. However, its application scenario is relatively limited. Single optical path QoT prediction does not consider the potential impact of the existence of other channels in the link on the QoT of the optical path to be predicted, and it is difficult to reflect the real transmission environment. Moreover, there are situations in the network where multiple channels need to be predicted simultaneously. Multi-channel QoT prediction includes both classification and regression models, but it mainly predicts the QoT of multiple channels under a fixed route and cannot estimate the QoT indexes of multiple links and multiple channels at the same time. And due to the influence of fiber nonlinear effects, newly deployed optical paths will affect previously deployed optical paths, which may cause the QoT of the deployed optical paths to deteriorate, resulting in unreasonable resource allocation and even unreliable transmission. Therefore, it is necessary to perform multi-optical path QoT prediction on both newly deployed optical paths and already deployed optical paths in the network. There is relatively little research on existing multi-optical path QoT prediction models. There is only a multi-optical path QoT classification model, which can judge whether all newly deployed optical paths can successfully establish a connection under the current network state, but it cannot obtain the specific QoT value, cannot obtain the specific impact information of newly deployed optical paths on already deployed optical paths, and the currently designed model does not consider the characteristics of the optical network itself (there are some or a large number of unopened optical paths in the network, and the mutual influence between different links and channels), resulting in poor model feature extraction ability and serious redundancy of model parameters.

[0007] In summary, the current multi-optical path QoT prediction is a classifier that can only predict the availability of optical paths, but cannot give the QoT values of multi-optical paths in an optical network, making it impossible to guide network planning, such as the selection of modulation formats and transmission routes. Summary of the Invention

[0008] To solve the technical problem in the prior art that only the availability of optical paths can be predicted and the QoT values of multi-optical paths in an optical network cannot be predicted, an embodiment of the present invention provides a method for predicting the transmission quality of multi-optical paths in an optical network based on deep learning. The technical solution is as follows:

[0009] On the one hand, a method for predicting the transmission quality of multi-optical paths in an optical network based on deep learning is provided. This method is implemented by an optical network multi-optical path transmission quality prediction device, and the method includes:

[0010] Obtain the optical network status data for processing to obtain a training set;

[0011] Construct a MAAResU-Net model based on a residual neural network, an adaptive mask attention mechanism, and a U-shaped symmetric network, and use the training set to train the constructed MAAResU-Net model; where MAAResU-Net represents an adaptive mask attention symmetric residual network. During the training process, after deeply extracting the features of the optical network status data through the residual neural network, the adaptive mask attention mechanism is applied to adaptively filter the unopened optical paths, calculate the importance of different features, and the dependence relationship between links and channels. Then, the U-shaped symmetric network is used for upsampling and residual fusion to restore the feature space, and an interval regression strategy is adopted to predict the QoT of each optical path;

[0012] Use the trained MAAResU-Net model to predict the QoT index value of the optical path.

[0013] Further, the obtaining the optical network status data for processing to obtain a training set includes:

[0014] Obtain the optical network status data and combine them into a four-dimensional shape Xori; where the obtained optical network status data includes: S data samples, each sample contains N physical characteristics of L×F optical paths corresponding to all L links and their F channels in the entire optical network, Xori ∈ R S×N×L×F , S is the number of samples, N represents the number of collected physical characteristics, L represents the number of optical links, and F represents the number of channels per link;

[0015] Divide the QoT index values of each optical path in the optical network into K discrete intervals according to the range, and generate corresponding K category labels;

[0016] Select N from the N physical characteristics of the obtained optical network status datasel items constitute the optical network state feature matrix Among them, B represents the size of each batch of data;

[0017] Preprocess the data in the optical network state feature matrix X to obtain a training set; among them, the preprocessing includes: filtering abnormal data and normalization processing.

[0018] Furthermore, before training the constructed MAAResU-Net model using the training set, the method further includes:

[0019] Generate a mask matrix M according to whether the optical path of channel f of link l is open:

[0020]

[0021] Furthermore, the adaptive mask attention symmetric residual network model includes: an encoder, a bottleneck layer, an adaptive attention mechanism module, a decoder, and a fully connected layer;

[0022] Among them, the encoder uses the ResNet structure to deeply extract multi-scale features of the input data; among them, ResNet is a residual neural network, and the input data of the encoder during the training process is the training set;

[0023] The bottleneck layer is used to fuse the multi-scale features extracted by the encoder and output the feature map X bott , providing deep feature information for the decoder upsampling;

[0024] The adaptive attention mechanism module is used to combine the physical meaning of the optical network, apply the mask attention mechanism to perform a mask operation on the feature map X bott to adaptively filter the unopened optical paths, and apply the channel and spatial dual attention mechanisms to calculate the importance of different features and the dependence relationship between the link and the channel, so that the MAAResU-Net model focuses on the effective optical paths and features and outputs the feature map X sa ;

[0025] The decoder is used to upsample the input feature map X sa to restore the spatial resolution and output the feature map X dec ;

[0026] The fully connected layer is used to adopt an interval regression strategy to map the features at each position of the feature map X dec output by the decoder to the indices of K discrete intervals, and calculate the median of the obtained discrete intervals as the predicted value of the corresponding optical path QoT index; among them, the output interval index represents the predicted interval of the corresponding optical path QoT index value.

[0027] Furthermore, the basic structure of the ResNet is a residual block Res-Block, and each residual block is followed by a max pooling layer;

[0028] Among them, each residual block contains two convolutional layers, and each convolutional layer is followed by batch normalization and a ReLU activation function.

[0029] Furthermore, the adaptive attention mechanism module is used to match the dimension of the mask matrix M with the feature map X bott through a broadcast mechanism for effective masking operations. Specifically, the mask matrix M is downsampled using convolutional or pooling operations to obtain a dimension-matched matrix M';

[0030] The feature map X is masked using the matrix M' bott to obtain the feature X masked :

[0031] X masked = X bott × M'

[0032] The channel attention mechanism is applied to each layer of the feature map X masked to learn the global information of each channel and calculate the importance of each channel, and each channel is weighted to enhance the attention of the MAAResU-Net model to important channels, obtaining the feature map X ca :

[0033] X ca = X masked × M c (X mased )

[0034] M c (X mased ) = sigmod(MLP(AvgPool(X mased )) + MLP(MaxPool(X mased )))

[0035] Among them, M c () represents channel attention, sigmod() represents the sigmod activation function, MLP() represents passing through a multi-layer perceptron, AvgPool() represents the average pooling layer, and MaxPool() represents the max pooling layer;

[0036] Spatial attention is applied to weight each position in the feature map X ca to focus on more important spatial regions, obtaining the feature map X sa :

[0037] X sa = X ca × M s (Xca )

[0038] M s (X ca ) = sigmod(f 3×3 ([AvgPool(X ca );MaxPool(X ca )]))

[0039] Among them, M s () represents spatial attention, and f 3×3 () represents performing a convolution operation using a 3×3 convolution kernel.

[0040] Furthermore, the decoder adopts the upsampling structure of U-Net to upsample the input feature map X sa to restore the spatial information of the feature map. The feature map with restored spatial information is concatenated with the feature map transmitted from the corresponding Res-Block in the encoder, so that the MAAResU-Net model integrates the rich detailed information of the encoder feature map. Among them, U-Net is a U-shaped symmetric network.

[0041] Furthermore, the basic structure of U-Net is Unet-Block, and the number of Unet-Blocks is the same as the number of residual blocks Res-Blocks in the encoder.

[0042] On the other hand, a multi-optical path transmission quality prediction device for an optical network is provided. The multi-optical path transmission quality prediction device for an optical network includes: a processor; a memory, and computer-readable instructions are stored on the memory. When the computer-readable instructions are executed by the processor, any one of the methods in the above-mentioned optical network multi-optical path transmission quality prediction method based on deep learning is implemented.

[0043] On the other hand, a computer-readable storage medium is provided. At least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement any one of the methods in the above-mentioned optical network multi-optical path transmission quality prediction method based on deep learning.

[0044] The beneficial effects brought by the technical solutions provided in the embodiments of the present invention at least include:

[0045] 1) Deeply extract multi-scale features of the optical network state data through a residual neural network, and perform feature fusion through a bottleneck layer to reduce the computational complexity;

[0046] 2) The adaptive mask attention mechanism combines the physical meaning of the optical network to mask the unopened optical paths, thereby achieving adaptive filtering of unopened optical paths, eliminating the interference of invalid unopened optical paths, and applying the channel and spatial dual attention mechanisms to calculate the importance of different features and the dependencies between links and channels, enhancing the learning of key features, and enabling the MAAResU-Net model to focus on effective optical paths and features;

[0047] 3) The U-Net structure is used for upsampling and residual fusion to gradually restore the feature space, enabling the MAAResU-Net model to fuse the rich detailed information of the encoder feature map, enhancing the global information learning, and contributing to improving the prediction accuracy of the model.

[0048] 4) The interval regression strategy is adopted to predict the QoT of each optical path, which can effectively avoid the influence of outliers on the model, improve the robustness of the model, quickly achieve high-precision prediction of the multi-optical path QoT of the optical network, and thus contribute to guiding the selection of modulation formats, reducing the network margin, and achieving reliable transmission of the optical network. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0050] Figure 1 is a flowchart of a method for predicting the transmission quality of multi-optical paths in an optical network based on deep learning provided by an embodiment of the present invention;

[0051] Figure 2 is a schematic structural diagram of the MAAResU-Net model provided by an embodiment of the present invention;

[0052] Figure 3 is a schematic structural diagram of the residual block Res-Block provided by an embodiment of the present invention;

[0053] Figure 4 is a schematic structural diagram of the Unet-Block structure provided by an embodiment of the present invention;

[0054] Figure 5 is a schematic structural diagram of a device for predicting the transmission quality of multi-optical paths in an optical network provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] The following describes the technical solutions in the present invention with reference to the drawings.

[0056] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as an "example" in the present invention should not be construed as being more preferred or advantageous than other embodiments or design solutions. Rather, the use of the word "example" is intended to present concepts in a specific manner. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two.

[0057] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same. "(of)", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same.

[0058] In the embodiments of the present invention, sometimes a subscript such as W 1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meanings they express are the same.

[0059] To make the technical problems to be solved, technical solutions and advantages of the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0060] The embodiments of the present invention provide a method for predicting the transmission quality of multiple optical paths in an optical network based on deep learning. This method can be implemented by an optical network multiple optical path transmission quality prediction device, and this optical network multiple optical path transmission quality prediction device can be a terminal or a server. As Figure 1 shown in the flowchart of the method for predicting the transmission quality of multiple optical paths in an optical network based on deep learning, the processing flow of this method can include the following steps:

[0061] S1. Obtain the optical network state data for processing to obtain a training set; specifically, it can include the following steps:

[0062] S11. Obtain the optical network state data and combine it into a four-dimensional shape Xori; among them, the obtained optical network state data includes: S data samples, each sample contains N physical characteristics of L×F optical paths corresponding to all L links and their F channels in the entire optical network, Xori ∈ R S×N×L×F , S is the number of samples, N represents the number of physical characteristics collected, L represents the number of optical links, and F represents the number of channels per link;

[0063] In this embodiment, the optical network state data can be obtained through an experimental or simulation platform. Among them, the obtained physical characteristics include: modulation format, link length, number of spans, etc.

[0064] S12. Divide the QoT index values of each optical path in the optical network into K discrete intervals according to the range, and generate the corresponding K category labels.

[0065] In this embodiment, QoT is a key indicator to measure the quality and reliability of optical signals during transmission. Its commonly used measurement standards include bit error rate (BER), signal-to-noise ratio (SNR), optical signal-to-noise ratio (OSNR), and Q-factor, etc.

[0066] In this embodiment, divide the QoT index values of each optical path in the optical network into K discrete intervals according to the range, and generate the corresponding K category labels. Taking OSNR as an example, the size range of each discrete interval is:

[0067]

[0068] where OSNR max and OSNR min are the maximum and minimum values of the optical signal-to-noise ratio respectively.

[0069] S13. Select N items from the N physical features of the obtained optical network status data to form the optical network status feature matrix sel where B represents the size of each batch of data. where B represents the size of each batch of data;

[0070] S14. Preprocess the data in the optical network status feature matrix to obtain a training set; where the preprocessing includes: filtering abnormal data and normalization processing.

[0071] In this embodiment, range filtering is performed on the optical network status data with abnormal QoT index values in the optical network status feature matrix according to the network situation or the settings of the simulation platform, and abnormal samples are deleted.

[0072] In this embodiment, each feature in the optical network status feature matrix X after the filtering operation is mapped to the interval (0, 1) through the normalization formula; where the normalization formula is:

[0073]

[0074] where x m is the normalized data, x is the data to be normalized, x min is the minimum value of the data to be normalized, and x max is the maximum value of the data to be normalized.

[0075] S2. Construct the MAAResU-Net model based on the residual neural network, the adaptive mask attention mechanism, and the U-shaped symmetric network, and use the training set to train the constructed MAAResU-Net model; where MAAResU-Net represents the Adaptive Mask Attention Symmetric Residual Network. During the training process, after deeply extracting the features of the optical network state data through the residual neural network, the adaptive mask attention mechanism is applied to adaptively filter the unopened optical paths, calculate the importance of different features, and the dependencies between links and channels. Then, after upsampling and residual fusion to restore the feature space using the U-shaped symmetric network, the interval regression strategy is adopted to predict the QoT of each optical path.

[0076] In this embodiment, before using the training set to train the constructed MAAResU-Net model, the method further includes:

[0077] Generate a mask matrix M according to whether the optical path of channel f of link l is opened:

[0078]

[0079] In this embodiment, the generated mask matrix M is used to implement subsequent mask operations.

[0080] In this embodiment, as Figure 2 shown, the Adaptive Mask Attention Symmetric Residual Network model includes: an encoder, a bottleneck layer (Bottleneck), an Adaptive Mask Attention mechanism (Mask-Aware Attention) module, a decoder, and a fully connected layer;

[0081] Among them, the encoder uses the ResNet structure to deeply extract multi-scale features of the input data and outputs the feature map X enc ; where ResNet is the residual neural network, and the input data of the encoder during the training process is the training set;

[0082] The bottleneck layer is used to fuse the multi-scale features extracted by the encoder and outputs the feature map X bott , providing deep feature information for the decoder upsampling;

[0083] The adaptive attention mechanism module is used to combine the physical meaning of the optical network, apply the mask attention mechanism to perform a mask operation on the feature map X bott to adaptively filter the unopened optical paths, and apply the channel and spatial dual attention mechanisms to calculate the importance of different features and the dependencies between links and channels, so that the MAAResU-Net model focuses on the effective optical paths and features and outputs the feature map X sa ;

[0084] The decoder is used to upsample the input feature map Xsa Perform upsampling to restore the spatial resolution and output the feature map X dec ;

[0085] A fully connected layer, which is used to adopt an interval regression strategy to map the features at each position of the feature map X output by the decoder dec to the indices of K discrete intervals, and according to the obtained indices of the discrete intervals, calculate the median of the discrete intervals as the predicted value of the corresponding optical path QoT index; wherein, the output interval index represents the predicted interval of the corresponding optical path QoT index value.

[0086] To better understand the MAAResU-Net model, its results are described in detail as follows:

[0087] 1) Encoder

[0088] In this embodiment, the basic structure of the ResNet is a residual block Res-Block, as Figure 3 shown, and a max pooling layer (MaxPooling) is connected after each residual block;

[0089] Among them, each residual block contains two convolutional layers, and a batch normalization (BatchNormalization, BN) and a ReLU activation function are connected after each convolutional layer.

[0090] In this embodiment, according to the number of links included in the network topology to be predicted, different numbers of Res-Blocks can be selected for deep feature extraction. In this embodiment, taking the application of two Res-Blocks as an example, a max pooling layer is connected after each residual block to gradually reduce the spatial resolution of the feature map, and the calculation process and the output dimensions of the feature data of each layer are shown in Table 1.

[0091] Table 1 Input and output dimensions of each layer of the encoder

[0092] Computation layer Input dimension Output dimension Input layer <![CDATA[(B,N sel ,L,F)]]> <![CDATA[(B,N sel ,L,F)]]> Res-Block <![CDATA[(B,N sel ,L,F)]]> (B, 64, L, F) MaxPooling(2×2) (B, 64, L, F) (B, 64, L / 2, F / 2) Res-Block (B, 64, L / 2, F / 2) (B, 128, L / 2, F / 2) MaxPooling(2×2) (B, 128, L / 2, F / 2) (B, 128, L / 4, F / 4)

[0093] 2) Bottleneck layer

[0094] In this embodiment, the main function of the bottleneck layer is to fuse the multi-scale features extracted by the encoder and provide deep feature information for the decoder upsampling. It receives the feature map X enc from the encoder and outputs the feature map X bott .

[0095] In this embodiment, the basic structure of the bottleneck layer is a 3×3 convolution and a ReLU activation function, and the input and output dimension transformation is shown in Table 2:

[0096] Table 2 Input and output dimensions of the bottleneck layer

[0097] Computation layer Input dimension Output dimension Bottleneck (B, 128, L / 4, F / 4) (B, 256, L / 4, F / 4)

[0098] 3) Adaptive attention mechanism module

[0099] In this embodiment, the adaptive attention mechanism module uses the generated mask matrix M to filter the unopened optical paths to reduce the interference of invalid information; and applies the channel and spatial dual attention mechanisms to calculate the importance of different features and the dependence relationship between the link and the channel, enhancing the learning of key features and enabling the MAAResU-Net model to focus on the effective optical paths and features.

[0100] In this embodiment, the adaptive attention mechanism module first makes the mask matrix M match the dimension of the feature map X through the broadcast mechanism bott for effective masking operation. The specific method is to perform downsampling on M using convolution or pooling operations to obtain a matrix M' with matching dimensions;

[0101] Use the matrix M' to mask the feature map X bott to obtain the feature X masked :

[0102] X masked = X bott × M'

[0103] After this adaptive attention mechanism module is placed in the bottleneck layer, on the one hand, the bottleneck layer has extracted enough global features. After adding the mask attention mechanism, it ensures that after the MAAResU-Net model learns higher-level abstract features, it can effectively filter the invalid information of unopened optical paths and only focus on the features of effective channels, thereby increasing the learning ability of the model for global information and effective information in the optical network field;

[0104] Next, apply the channel attention mechanism to learn the global information of each layer (each channel) of the feature map X masked to calculate the importance of each channel, and weight each channel to enhance the attention of the MAAResU-Net model to important channels, obtaining the feature map X ca :

[0105] X ca = X masked × M c (X mased )

[0106] M c (X mased ) = sigmod(MLP(AvgPool(X mased ) + MLP(MaxPool(X mased )))

[0107] where Mc ( ) represents channel attention, sigmod( ) represents the sigmod activation function, MLP( ) represents passing through a multi-layer perceptron, AvgPool( ) represents the average pooling layer, and MaxPool( ) represents the max pooling layer;

[0108] Apply spatial attention to the feature map X ca at each position in it to focus on more important spatial regions, obtaining the feature map X sa :

[0109] X sa = X ca × M s (X ca )

[0110] M s (X ca ) = sigmod(f 3×3 ([AvgPool(X ca ) ; MaxPool(X ca )]))

[0111] Among them, M s () represents spatial attention, and f 3×3 () represents performing a convolution operation using a 3×3 convolution kernel.

[0112] 4) Decoder

[0113] In this embodiment, the decoder is responsible for restoring the spatial resolution and generating the prediction result. It adopts the upsampling structure of the U-shaped symmetric network (U-Net) to upsample the input feature map, gradually restoring the spatial information of the feature map, and concatenating (Concat) the feature map passed by the corresponding encoder layer to the current layer, enabling the model to fuse high-resolution information and enhancing global information learning. Its input is the feature map X sa after the attention mechanism, and finally outputs the feature map X dec .

[0114] The decoder first adopts the upsampling structure of U-Net to upsample the input feature map X sa to restore the spatial information of the feature map. The feature map with restored spatial information is concatenated with the feature map passed by the corresponding Res-Block in the encoder, enabling the MAAResU-Net model to fuse the rich detailed information of the encoder feature map. In this way, while obtaining multi-scale features, information loss is also avoided, which helps to improve the prediction accuracy of the model.

[0115] In this embodiment, the basic structure of U-Net is Unet-Block, and the number of Unet-Blocks used should be the same as the number of Res-Blocks in the encoder, as Figure 4 shown.

[0116] In this embodiment, the dimension transformation of the feature maps of each layer of Unet-Block is shown in Table 3:

[0117] Table 3 Input and Output Dimensions of Unet-Block

[0118] Computation layer Input dimension Output dimension Upsampling(2×2) (B, 256, L / 4, F / 4) (B, 128, L / 2, F / 2) Feature transfer and concatenation (B, 128, L / 2, F / 2) (B, 256, L / 2, F / 2) 3×3 convolution (B, 256, L / 2, F / 2) (B, 128, L / 2, F / 2) 3×3 convolution (B, 128, L / 2, F / 2) (B, 128, L / 2, F / 2) Batch Normalization (B, 128, L / 2, F / 2) (B, 128, L / 2, F / 2)

[0119] In this embodiment, after applying 2 Unet-Blocks with the same number as the Res-Blocks, the dimension transformation of the overall decoder part is shown in Table 4:

[0120] Table 4 Input and Output Dimensions of the Decoder

[0121] Computation layer Input dimension Output dimension Unet-Block (B, 256, L / 4, F / 4) (B, 128, L / 2, F / 2) Unet-Block (B, 128, L / 2, F / 2) (B, 64, L, F)

[0122] 5) Fully Connected Layer

[0123] Currently, network operators attach great importance to the stability of services and prefer to retain a large margin to avoid optical path degradation or interruption. Conventional QoT regression prediction methods are very sensitive to outliers when predicting multi-channel and multi-optical path networks, resulting in a large mean absolute error (MAE) and being unable to be applied to actual networks. The present invention proposes a scheme of interval regression strategy to reduce the influence of outliers on the model.

[0124] In this embodiment, the core idea of the interval regression strategy adopted by the fully connected layer is to map continuous regression values into multiple discrete intervals and predict the indices of these discrete intervals through the output X dec of the model. The number of interval partitions affects the prediction accuracy of the QoT index. The more the number of divided intervals, the lower the final prediction error, but too many intervals may affect the accuracy of the model.

[0125] In this embodiment, the fully connected layer adopts an interval regression strategy to map the feature of each position of the feature map X dec generated by the decoder to the indices of K discrete intervals in step S12. According to the obtained indices of the discrete intervals, the median of the discrete intervals is calculated as the predicted value of the corresponding optical path QoT index, reducing the influence of outliers on the model and improving the robustness of the MAAResU-Net model; among them, the output interval index represents the prediction interval of the corresponding optical path QoT index value.

[0126] In this embodiment, the constructed MAAResU-Net model is trained using a training set. The objective function for model training adopts the Cross-Entropy Loss function, and the Adam optimizer is selected. The prediction effect is verified through the classification accuracy verification interval. Training can be stopped when the model converges or the accuracy reaches the target. A validation set is used to evaluate the interval prediction accuracy of the model, and finally a test machine is used to verify the model performance.

[0127] S3. Use the trained MAAResU-Net model to predict the QoT metric values of the optical path.

[0128] In summary, the method for predicting the transmission quality of multiple optical paths in an optical network based on deep learning provided by the embodiments of the present invention has at least the following beneficial effects:

[0129] 1) Deeply extract multi-scale features of the optical network state data through a residual neural network, and perform feature fusion through a bottleneck layer to reduce the computational complexity;

[0130] 2) The adaptive mask attention mechanism combines the physical meaning of the optical network to mask the unopened optical paths, thereby achieving adaptive filtering of the unopened optical paths, eliminating the interference of invalid unopened optical paths, and applying the channel and spatial dual attention mechanisms to calculate the importance of different features and the dependencies between links and channels, enhancing the learning of key features, and enabling the MAAResU-Net model to focus on effective optical paths and features;

[0131] 3) Adopt the U-Net structure for upsampling and residual fusion to gradually restore the feature space, so that the MAAResU-Net model integrates the rich detailed information of the encoder feature map, enhances the global information learning, and helps to improve the prediction accuracy of the model.

[0132] 4) Adopt the interval regression strategy to predict the QoT of each optical path, which can effectively avoid the influence of outliers on the model, improve the robustness of the model, quickly achieve high-precision prediction of the QoT of multiple optical paths in the optical network, thereby helping to guide the selection of modulation formats, reduce the network margin, and achieve reliable transmission of the optical network.

[0133] Figure 5 It is a schematic structural diagram of an optical network multi-optical path transmission quality prediction device provided by an embodiment of the present invention. As Figure 5 shown, the optical network multi-optical path transmission quality prediction device may include the above-mentioned Figure 3 optical network multi-optical path transmission quality prediction device shown. Optionally, the optical network multi-optical path transmission quality prediction device 510 may include a first processor 2001.

[0134] Optionally, the optical network multi-optical path transmission quality prediction device 510 may further include a memory 2002 and a transceiver 2003.

[0135] Among them, the first processor 2001 is connected to the memory 2002 and the transceiver 2003, such as through a communication bus.

[0136] Next, in combination with Figure 5 Specific introductions will be made to the various components of the optical network multi-optical path transmission quality prediction device 510:

[0137] Among them, the first processor 2001 is the control center of the optical network multi-optical path transmission quality prediction device 510, which can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 is one or more central processing units (CPUs), or can be an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention, such as: one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs).

[0138] Optionally, the first processor 2001 can execute various functions of the optical network multi-optical path transmission quality prediction device 510 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0139] In a specific implementation, as an embodiment, the first processor 2001 may include one or more CPUs, such as Figure 5 the CPU0 and CPU1 shown in

[0140] In a specific implementation, as an embodiment, the optical network multi-optical path transmission quality prediction device 510 may also include multiple processors, such as Figure 5 the first processor 2001 and the second processor 2004 shown in

[0141] Among them, the memory 2002 is used to store the software program for implementing the solution of the present invention and is controlled by the first processor 2001 for execution. The specific implementation manner can refer to the above method embodiments and will not be elaborated here.

[0142] Optionally, the memory 2002 can be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or can also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 2002 can be integrated with the first processor 2001 or exist independently and is coupled to the first processor 2001 through the interface circuit ( Figure 5 not shown) of the optical network multi-path transmission quality prediction device 510. The embodiments of the present invention do not make specific limitations on this.

[0143] The transceiver 2003 is used to communicate with a network device or communicate with a terminal device.

[0144] Optionally, the transceiver 2003 can include a receiver and a transmitter ( Figure 5 not shown separately). Among them, the receiver is used to implement the receiving function, and the transmitter is used to implement the sending function.

[0145] Optionally, the transceiver 2003 can be integrated with the first processor 2001 or exist independently and is coupled to the first processor 2001 through the interface circuit ( Figure 5 not shown) of the optical network multi-path transmission quality prediction device 510. The embodiments of the present invention do not make specific limitations on this.

[0146] It should be noted that Figure 5 the structure of the optical network multi-path transmission quality prediction device 510 shown does not constitute a limitation on the router. The actual knowledge structure recognition device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0147] In addition, for the technical effects of the optical network multi-optical path transmission quality prediction device 510, reference can be made to the technical effects of the optical network multi-optical path transmission quality prediction method based on deep learning described in the above method embodiments, which will not be elaborated here.

[0148] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), and this processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or this processor may also be any conventional processor, etc.

[0149] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable ROM (PROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of random access memory (RAM) are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchlink DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0150] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that contains one or more collections of available media. The available media can be magnetic media (such as floppy disks, hard disks, magnetic tapes), optical media (such as DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0151] It should be understood that the term "and / or" in this document is merely a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. Here, A and B can be singular or plural. In addition, the character " / " in this document generally represents an "or" relationship between the associated objects before and after, but it may also represent an "and / or" relationship, which can be specifically understood by referring to the context.

[0152] In the present invention, "at least one" means one or more, and "a plurality" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, a - b, a - c, b - c, or a - b - c, where a, b, and c can be single or multiple.

[0153] It should be understood that in various embodiments of the present invention, the magnitudes of the sequence numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0154] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0155] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the devices, apparatuses, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0156] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be electrical, mechanical, or other forms.

[0157] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place, or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0158] In addition, the functional units in each embodiment of the present invention can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0159] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of this technical solution can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0160] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for predicting the transmission quality of multi-optical paths in optical networks based on deep learning, characterized in that: The method comprises: Obtain optical network status data for processing to obtain a training set; A MAAResU-Net model based on residual neural network, adaptive masked attention mechanism and U-shaped symmetric network is constructed, and the constructed MAAResU-Net model is trained using the training set; MAAResU-Net represents adaptive masked attention symmetric residual network. During the training process, after the residual neural network deeply extracts the characteristics of the optical network status data, the adaptive masked attention mechanism is applied to adaptively filter the unopened optical paths, calculate the importance of different features and the dependency between links and channels, and then the U-shaped symmetric network is used for upsampling and residual fusion to restore the feature space, and then the interval regression strategy is used to predict the QoT of each optical path; The trained MAAResU-Net model is used to predict the QoT index value of the light path.

2. The optical network multi-path transmission quality prediction method based on deep learning according to claim 1 is characterized in that: The optical network status data is obtained and processed to obtain a training set including: The obtained optical network status data is combined into a four-dimensional shape Xori; wherein the obtained optical network status data includes: S data samples, each sample contains N physical characteristics of L×F optical paths corresponding to all L links and their F channels in the entire optical network, Xori∈R S×N×L×F , S is the number of samples, N is the number of physical features collected, L is the number of optical links, and F is the number of channels per link; Divide the QoT index value of each optical path in the optical network into K discrete intervals according to the range, and generate corresponding K category labels; Select N from the N physical characteristics of the acquired optical network status data sel The items constitute the optical network state characteristic matrix Among them, B represents the size of each batch of data; The data in the optical network state feature matrix X is preprocessed to obtain a training set; wherein the preprocessing includes: filtering abnormal data and normalizing.

3. The optical network multi-path transmission quality prediction method based on deep learning according to claim 1 is characterized in that: Before training the constructed MAAResU-Net model using the training set, the method further includes: Generate the mask matrix M based on whether the channel f optical path of link l is open:

4. The method for predicting the quality of multi-path transmission in optical networks based on deep learning according to claim 3 is characterized in that: The adaptive masked attention symmetric residual network model includes: an encoder, a bottleneck layer, an adaptive attention mechanism module, a decoder and a fully connected layer; The encoder uses a ResNet structure to deeply extract multi-scale features of input data; ResNet is a residual neural network, and the input data of the encoder during training is a training set; The bottleneck layer is used to fuse the multi-scale features extracted by the encoder and output the feature map X bott , providing deep feature information for decoder upsampling; Adaptive attention mechanism module, used to combine the physical meaning of optical network and apply mask attention mechanism to feature map X bott Perform mask operations, adaptively filter unopened light paths, and apply channel and spatial dual attention mechanisms to calculate the importance of different features and the dependencies between links and channels, so that the MAAResU-Net model focuses on valid light paths and features, and outputs feature maps X sa ; Decoder, used to decode the input feature map X sa Upsample to restore spatial resolution and output feature map X dec ; The fully connected layer is used to use the interval regression strategy to transform the feature map X output by the decoder dec The feature of each position is mapped to the index of K discrete intervals, and according to the obtained index of the discrete interval, the median of the discrete interval is calculated as the predicted value of the corresponding optical path QoT indicator; wherein the output interval index represents the predicted interval of the corresponding optical path QoT indicator value.

5. The method for predicting the quality of optical network multi-path transmission based on deep learning according to claim 4 is characterized in that: The basic structure of the ResNet is a residual block Res-Block, each of which is followed by a maximum pooling layer; Among them, each residual block contains two convolutional layers, and each convolutional layer is followed by batch normalization and ReLU activation function.

6. The optical network multi-path transmission quality prediction method based on deep learning according to claim 4 is characterized in that: The adaptive attention mechanism module is used to make the mask matrix M and the feature map X through the broadcast mechanism bott The dimensions of M are matched to perform effective mask operations. The specific method is to downsample M using convolution or pooling operations to obtain a dimension-matched matrix M′; Use matrix M′ to transform feature map X bott Mask and get feature X masked : X masked =X bott ×M′ Apply channel attention mechanism to feature map X masked Each layer of the learning channel global information to calculate the importance of each channel, weight each channel to enhance the MAAResU-Net model to pay attention to important channels, and obtain the feature map X ca : X ca =X masked ×M c (X mased ) M c (X mased )=sigmod(MLP(AvgPool(X mased )+MLP(MaxPool(X mased )) Among them, M c () represents channel attention, sigmod() represents sigmod activation function, MLP() represents multi-layer perceptron, AvgPool() represents average pooling layer, MaxPool() represents maximum pooling layer; Apply spatial attention to the feature map X ca Each position in is weighted to focus on more important spatial regions and obtain the feature map X sa : X sa =X ca ×M s (X ca ) M s (X ca )=sigmod(f 3×3 ([AvgPool(X ca );MaxPool(X ca )])) Among them, M s () represents spatial attention, f 3×3 () indicates that a convolution operation is performed using a 3×3 convolution kernel.

7. The optical network multi-path transmission quality prediction method based on deep learning according to claim 4 is characterized in that: The decoder uses the U-Net upsampling structure to input the feature map X sa Upsampling is performed to restore the spatial information of the feature map, and the feature map with restored spatial information is spliced ​​with the feature map transmitted by the corresponding Res-Block in the encoder, so that the MAAResU-Net model integrates the rich detail information of the encoder feature map, where U-Net is a U-shaped symmetric network.

8. The optical network multi-path transmission quality prediction method based on deep learning according to claim 7 is characterized in that: The basic structure of U-Net is Unet-Block, and the number of Unet-Blocks is the same as the number of residual blocks Res-Blocks in the encoder.

9. An optical network multi-path transmission quality prediction device, characterized in that: The optical network multi-optical path transmission quality prediction device comprises: processor; A memory having computer-readable instructions stored thereon, wherein when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program codes, which can be called by a processor to execute the method according to any one of claims 1 to 8.

Images