An optical path transmission quality prediction method based on multi-task learning

Abstract

The application discloses a kind of based on multi-task learning's light path transmission quality prediction method, network reliability and prediction model accuracy are considered simultaneously, a kind of multi-task loss function is designed, and model is trained using the way of joint training, further reduce the error and design allowance of model, improve the accuracy of model, while guaranteeing network reliability, realize network capacity promotion.Compared with the traditional QoT estimation model based on single task, the scheme proposed in the patent can improve the accuracy while ensuring the reliability of the model, thereby achieving the purpose of improving the allocation efficiency of optical path modulation format and improving the capacity of optical network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical path transmission, and in particular to an optical path transmission quality prediction method based on multi-task learning. BACKGROUND

[0002] The explosive growth of Internet data services worldwide means that optical networks need to develop towards greater capacity, which has increasingly high demands on network operators and the network itself. At the same time, the emergence of more and more emerging technologies such as artificial intelligence brings new opportunities and challenges to optical networks.

[0003] In traditional optical transmission networks, in order to meet the smooth deployment of services in the life cycle of the optical network, sufficient margin needs to be reserved for the optical network during the deployment phase of the optical network. The margin mainly includes design margin, unallocated margin and system margin, wherein the design margin is the margin reserved due to inaccurate estimation of signal transmission quality. Due to the existence of the margin, the selection of the modulation format of the optical signal is low, the spectral efficiency of the optical path is reduced, and therefore the available capacity of the optical link is reduced.

[0004] In order to cope with the diversity of future services, optical networks need to have the characteristics of flexibility and reconfigurability, so that different routes will be taken by different services. If the margin required by the route taken by the service can be accurately predicted, the modulation format can be flexibly allocated, and the available capacity of the optical link can be improved, and the transmission efficiency of the network can be improved. Therefore, an accurate optical path transmission quality prediction method is one of the key technologies to realize a low-margin optical network, and predicting the transmission quality (QoT) of the optical path before deployment is a crucial step for the optimal design of the optical network.

[0005] Each optical path in the optical network is composed of one or more links, and let the undirected graph G = (V, E) represent the optical transport network, wherein V represents the set of nodes in the network, and E represents the set of network links. P represents the state of the current network, i.e. the state of the established optical paths, and p represents each optical path therein. Let Q(p|P) represent the QoT value of the optical path p under the network state P, and Q(P) represent the set of QoT values of all optical paths under the network state P. Let ω represent the optical path to be established, and the new optical path QoT value needs to be predicted, i.e. Q * (ω|P). The problem of extracting relevant information from the established optical paths to predict the QoT of the new optical path can be represented as Q(p|P)→Q * (ω|P).

[0006] Traditional signal quality estimation models are divided into two categories. One is an accurate analytical model, which can obtain accurate QoT values by solving the optical transmission equation, but the calculation time is long, which is not suitable for large networks and dynamic networks. The other is an approximate estimation model, which has faster calculation speed, but the design margin is introduced due to the uncertainty of the model itself, which reduces the spectral efficiency. In view of the successful application of machine learning in network layer and physical layer, in order to well balance the calculation complexity and accuracy of QoT prediction, based on the data collected from the optical network, the method of machine learning can be used to implicitly capture the physical layer behavior through the change of data, so as to predict the QoT value of the unestablished connection link according to the link information of the established connection

[0007] At present, due to the inaccuracy of the optical path transmission quality prediction model, operators and equipment manufacturers will allocate modulation formats to the optical path by using the difference between the optical path signal quality prediction value and the design margin to ensure the reliability of the optical path. If the design margin is too high, it will lead to the selection of low modulation format when deploying the optical path, which will further lead to the reduction of network capacity.

[0008] Therefore, how to improve the accuracy of transmission quality prediction while ensuring the reliability of the optical path is the key problem to reduce the network design margin. However, the existing research mainly focuses on the average accuracy of the model, and improves the accuracy of the prediction model by reducing the average error of all samples, ignoring the analysis and modeling of positive and negative deviations of the model. When the network positive deviation is large, it will lead to unreliable optical path; when the network negative deviation is large, it will lead to the selection of low modulation format of the optical path, which will lead to the reduction of optical network capacity. Considering the actual network demand, the reliability of the optical path has a greater impact on the network performance, so it is necessary to reduce the impact of the prediction positive deviation on the network reliability as much as possible. However, the existing QoT estimation model based on single task learning method is difficult to improve the network capacity while ensuring the network reliability by reducing the average error of the sample.

[0009] The principle of QoT estimation tool based on machine learning (ML) is as follows Figure 1The network control module predicts the optical path transmission quality by using the QoT evaluation tool according to the current resource state of the network after receiving the optical path request in the traditional optical network deployment scheme. However, due to the inaccuracy of the QoT evaluation tool, a higher margin m needs to be reserved in the design. By introducing the machine learning training process and the prediction process, the trained model collects the physical layer information of the optical network. Then, the collected information is processed by using the machine learning technology to extract the nonlinear relationship between the physical layer information and the optical path QoT, establish the corresponding model, and set the loss function of the model. At present, the MSE and MAE are mostly used in the literature. The model parameters that make the loss function value minimum are found through supervised learning training. Then, when a new optical path request arrives, the QoT value of the new request optical path is predicted by the trained model. Due to the high accuracy of the model, the new design margin is smaller than the initial design margin.

[0010] A large number of methods for QoT estimation using supervised learning models have been studied in existing literature, such as maximum likelihood estimation, least squares method, gradient boosting decision tree, and neural network method. Among them, the neural network has high accuracy and strong generalization ability, and the trained model can be migrated to other network scenarios, and has been more widely applied.

[0011] The existing ML-based QoT estimation model is an independent model, and a symmetric loss function, i.e., mean square error (MSE) and absolute error (MAE), is used to train the model. Both of them are symmetric loss functions, and give the same penalty for overestimation and underestimation. The main concern is the average accuracy of the model, i.e., the average error of all samples, without additional attention to the maximum positive deviation of the model that affects the network reliability, i.e., the design margin of the network. Considering the actual network demand, the optical path reliability in the network has a greater impact on the network performance, and it is necessary to reduce the impact of the prediction positive deviation on the network reliability while improving the model accuracy.

[0012] Therefore, to solve this problem, a joint optimization model needs to be proposed to ensure the reliability of the model while improving the accuracy of the model, thereby maximizing the network capacity. SUMMARY

[0013] The purpose of the present application is to provide an optical path transmission quality prediction method based on multi-task learning. A multi-task loss function is designed considering network reliability and prediction model accuracy, and the model is trained by using joint training, further reducing the error and design margin of the model, and improving the accuracy of the model, while ensuring the network reliability and realizing the improvement of the network capacity.

[0014] To achieve the above purpose, the present application provides the following technical solutions:

[0015] A kind of optical path transmission quality prediction method based on multi-task learning, multi-task learning mechanism is applied to the QoT estimation process based on ANN, model input is signal and optical path related information, and model output is the QoT value of optical path;A kind of multi-task loss function is designed, the loss function Loss1 of one task is mean square error function MSE, as shown in formula (1), for paying attention to improve the average accuracy of model;The loss function of another task is Loss2, as shown in formula (3);Regularization term I (x) is introduced on the basis of MSE, as shown in formula (2), to realize the purpose of guaranteeing network reliability, reducing design allowance, i.e.the maximum positive deviation of model;The total loss function is the weighted sum of two loss functions, and is expressed as Loss, as shown in formula (4);

[0016]

[0017]

[0018]

[0019] Loss=βLoss1+(1-β)Loss2 (4)

[0020] Wherein y represents true value, Predicted value, α is the penalty coefficient greater than 0, and β is a constant between 0 and 1.

[0021] Further, the joint training steps of the model are as follows:

[0022] First step: initialize multi-task learning model structure, set the hyperparameters of the model, and randomly initialize the weight of each layer;

[0023] Second step: the input value of training data and its corresponding label value are output to the model, and the corresponding predicted value of the model is obtained according to the forward propagation algorithm;

[0024] Third step: according to the loss function Loss1 and Loss2, the error between the predicted value of the model and the given label value is calculated, and the joint loss function Loss value is calculated;

[0025] Fourth step: according to the Loss value, the weight of neuron is adjusted by using stochastic gradient descent method, so that the joint error of the model is minimized;

[0026] Fifth step: the forward transmission and back propagation between training data set in model are repeated, and the weight of neuron is updated, so that the joint loss value of each learning iteration is minimized;After a certain learning process iteration, the complete multi-task learning model is applied to new input data set to predict target value.

[0027] Compared with the prior art, the beneficial effects of the present application are:

[0028] The multi-task learning-based optical path transmission quality prediction method of the present application simultaneously considers network reliability and prediction model accuracy, designs a multi-task loss function, and trains the model using a joint training method, further reducing the model error and design margin, improving the accuracy of the model, and achieving network capacity improvement while ensuring network reliability. Compared with the traditional single-task-based QoT estimation model, the scheme proposed in the present patent can improve the accuracy while ensuring the reliability of the model, thereby achieving the purpose of improving the allocation efficiency of the optical path modulation format and improving the capacity of the optical network. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0030] Figure 1 is a ML-based QoT model training flowchart.

[0031] Figure 2 is a single-task learning schematic diagram.

[0032] Figure 3 is a multi-task learning schematic diagram.

[0033] Figure 4 is a multi-task learning schematic diagram based on hard parameter sharing.

[0034] Figure 5 is a multi-task learning-based optical path transmission quality estimation model result diagram. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0036] From the above analysis, it can be seen that the existing QoT estimation scheme mainly focuses on the average accuracy of the model, ignoring the maximum positive deviation term affecting the reliability of the model. Therefore, the present application mainly proposes a light path transmission quality prediction method based on multi-task learning, designs a multi-task loss function considering network reliability and prediction model accuracy, and trains the model by using the joint training method, further reduces the error and design margin of the model, improves the accuracy of the model, and realizes the improvement of network capacity while ensuring the reliability of the network.

[0037] Multi-task learning (Multi-task learning) is a machine learning method opposite to single-task learning (Single-task learning). As shown in Figure 2 and Figure 3 , single-task learning learns one task at a time, and multi-task learning is a kind of joint learning, multiple tasks are learned in parallel, and the results affect each other. Multi-task learning belongs to a kind of transfer learning, and traditional transfer learning emphasizes the order of learning, that is, the knowledge learned in one field is transferred to another field, and the process of knowledge transfer is serial. In multi-task learning, information between tasks is shared, and knowledge is transferred between different tasks, so multi-task learning is also called parallel transfer learning. Multi-task learning method improves the overall learning effect through information sharing, and theory and experiment have proved that when all tasks, at least part of the tasks, are related, joint learning of multiple tasks can get better performance than learning them separately.

[0038] Based on the sharing of hard parameters, one of the main forms of multi-task learning, it shares the hidden layer in all tasks while retaining the output layer of several specific tasks, and different tasks can be trained with different loss functions, as shown in Figure 4 . This way helps to reduce the risk of overfitting and improve model accuracy, because the more tasks learned at the same time, the more difficult it is for the model to find a representation containing all tasks, and the less likely it is to overfit to a specific task.

[0039] To further ensure the reliability of the model while improving the accuracy of the model, the multi-task learning mechanism described above is applied to the ANN-based QoT estimation process, and the model structure is as shown in Figure 5As shown, the model input is the signal and the light path related information, and the model output is the QoT value of the light path. A multi-task loss function is designed, in which the loss function Loss1 of one task is the mean square error function MSE, as shown in formula (1), mainly focusing on improving the average accuracy of the model, and the loss function of another task is Loss2, as shown in formula (3), introducing a regularization term I(x) on the basis of MSE, as shown in formula (2), to achieve the purpose of ensuring network reliability and reducing design margin, i.e. the maximum positive deviation of the model. The total loss function is the weighted sum of the two loss functions, denoted as Loss, as shown in formula (4).

[0040]

[0041]

[0042]

[0043] Loss=βLoss1+(1-β)Loss2 (4)

[0044] where y represents the true value, the predicted value, and a is a penalty coefficient greater than 0, and β is a constant between 0 and 1.

[0045] The following are the joint training steps of the model:

[0046] Step 1: Initialize the multi-task learning model structure, set the model's hyperparameters, and randomly initialize the weights for each layer.

[0047] Step 2: Output the input values and corresponding label values of the training data to the model, and obtain the predicted values of the model according to the forward propagation algorithm.

[0048] Step 3: Calculate the error between the predicted values of the model and the given label values according to the loss functions Loss1 and Loss2, and calculate the joint loss function Loss value.

[0049] Step 4: Adjust the weights of the neurons according to the Loss value using the stochastic gradient descent method to minimize the joint error of the model.

[0050] Step 5: Repeat the forward and backward propagation of the training data set between the models to update the weights of the neurons, so that the joint loss value of each learning iteration is minimized. After a certain learning process iteration, the complete multi-task learning model is applied to new input data sets to predict target values.

[0051] The application mainly provides a light path transmission quality prediction method based on multi-task learning, a multi-task loss function is designed considering network reliability and prediction model accuracy, and the model is trained by using a joint training method, so that the error and design allowance of the model are further reduced, the accuracy of the model is improved, the network capacity is improved while the network reliability is ensured. Compared with the traditional single-task-based QoT estimation model, the scheme provided in the patent can improve the accuracy while ensuring the reliability of the model, so as to realize the purposes of improving the allocation efficiency of the light path modulation format and improving the capacity of the optical network.

[0052] It should be noted that, in this document, the terms such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between such entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0053] Each of the embodiments in the specification is described in a related manner, and the same and similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. Especially, for the device embodiment, the electronic device embodiment, the computer readable storage medium embodiment and the computer program product embodiment, since they are basically similar to the method embodiment, the description is relatively simple, and the related parts can be referred to the part of the method embodiment.

[0054] The above-described embodiments are merely specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit the same. The protection scope of the present application is not limited thereto, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical scope disclosed by the present application, or make equivalent replacements to part of the technical solutions; and these modifications, changes or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application. All should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1.A method for optical path transmission quality prediction based on multi-task learning, characterized in that, The multi-task learning mechanism is applied to the ANN-based QoT estimation process, the model input is a signal and optical path related information, and the model output is a QoT value of the optical path; a multi-task loss function is designed, and the model is trained by using a joint training method, wherein a loss function Loss1 of one task is a mean square error function MSE, as shown in formula (1), which is used to focus on improving the average accuracy of the model; the loss function of another task is Loss2, as shown in formula (3); a regularization term I(x) is introduced on the basis of MSE, as shown in formula (2), to realize the purpose of guaranteeing network reliability and reducing design margin, that is, the maximum positive deviation of the model; the total loss function is the weighted sum of the two loss functions, denoted as Loss, as shown in formula (4); Loss=βLoss1+(1-β)Loss2(4) where y represents the true value, represents the predicted value, a is a penalty coefficient greater than 0, and β is a constant between 0 and 1. The joint training steps of the model are as follows: Step 1: initialize the multi-task learning model structure, set the model hyperparameters, and randomly initialize the weight values for each layer; Step 2: output the input value and its corresponding label value of the training data to the model, and obtain the corresponding prediction value of the model according to the forward propagation algorithm; Step 3: calculate the error between the model prediction value and the given label value according to the loss functions Loss1 and Loss2, and calculate the joint loss function Loss value; Step 4: adjust the neuron weights according to the Loss value by using the stochastic gradient descent method, so that the joint error of the model is minimized; Step 5: repeat the forward and backward propagation of the training data set between the models, update the neuron weights, and minimize the joint loss value of each learning iteration; after a certain learning process iteration, the complete multi-task learning model is applied to the new input data set to predict the target value.

Images