阅读说明：本技术 一种osnr检测方法及装置 (OSNR detection method and device ) 是由 于文海 程勇鹏 于 2021-08-27 设计创作，主要内容包括：本发明涉及WDM光层技术领域,提供了一种OSNR检测方法及装置。方法包括使用N种OSNR计算子模型分别得到各自在各OMS段；从而得到N种OSNR计算子模型的各个模型在相应光信道中的OSNR-(1),OSNR-(2),…,OSNR-(N)；通过机器学习的权重分配模型,计算出应用与当前光信道组合的所述N种OSNR计算子模型的各个权重W-(j),其中j为[1,N]区间内的自然数；将所述各个权重W-(j)分别与相应乘积并求和得到当前光信道的OSNR值。本发明利用多种OSNR计算子模型,利用各自的计算得到的OSNR值,提升不同波道组合最差场景下的OSNR计算精度。(The invention relates to the technical field of WDM optical layers, and provides an OSNR detection method and device. The method comprises the steps of respectively obtaining each OMS section by using N OSNR calculation submodels; thereby obtaining OSNR of each model of the N OSNR calculation submodels in the corresponding optical channel 1 ，OSNR 2 ，…，OSNR N (ii) a Calculating respective weights W of the N OSNR calculation submodels applied in combination with the current optical channel by a machine-learned weight assignment model j Wherein j is [1, N ]]A natural number within the interval; the respective weights W j And respectively multiplying the products by the corresponding products and summing to obtain the OSNR value of the current optical channel. According to the invention, various OSNR calculation submodels are utilized, and OSNR values obtained by respective calculation are utilized to improve the OSNR calculation precision under the worst scene of different channel combinations.)

1. An OSNR detection method, comprising an optical channel OCH to be calculated including k optical multiplexed OMS segments, a calculation model including N OSNR calculation submodels, the method comprising:

obtaining Δ OSNR at each OMS segment using N OSNR computation submodels_{OMS_1},ΔOSNR_{OMS_2}，...，ΔOSNR_{OMS_k}(ii) a Thereby obtaining OSNR of each model of the N OSNR calculation submodels in the corresponding optical channel₁，OSNR₂，…，OSNR_N；

Calculating respective weights W of the N OSNR calculation submodels applied in combination with the current optical channel by a machine-learned weight assignment model_jWherein j is [1, N ]]A natural number within the interval;

the respective weights W_jAnd respectively multiplying the products by the corresponding products and summing to obtain the OSNR value of the current optical channel.

2. The OSNR detection method of claim 1, wherein the k optically multiplexed OMS segments comprise:

OSNR calculation for each traffic OCH, which is divided into k OMS segments_{OMS_i}A combination of (1); wherein i is [1, k ]]A natural number within the interval;

wherein the output OSNR of the previous OMS is the input OSNR of the next OMS.

3. The OSNR detection method according to claim 2, wherein the OSNR of each of the N OSNR computation submodels in the corresponding optical channel is obtained₁，OSNR₂，…，OSNR_NThe method comprises the following steps:

each OSNR calculation submodel respectively adopts the following formula, and the OSNR calculation submodel under the corresponding optical channel is calculated in a piecewise recursion mode₁，OSNR₂，…，OSNR_N；

OSNR_{out,OMS_i}Represents the OSNR value at the output of the OMSi segment;

OSNR_{in,OMS_i}represents the OSNR value at the input end of the OMSi section;

ΔOSNR_{OMS_i}the segment of OMSi is represented by OSNR change value calculated and predicted according to the corresponding OSNR calculation submodel;

b is the channel bandwidth of each channel; bn is a constant of 12.5GHz as determined by the OSNR definition.

4. The OSNR detection method of claim 1, wherein the weighting W is based on the respective weights_jAnd respectively multiplying the optical signals by corresponding products and summing to obtain the OSNR value of the current optical channel, which specifically comprises the following steps:

is obtained by solving a formula, wherein the formula is as follows:

OSNR＝W₁*OSNR₁+W₂*OSNR₂+W₃*OSNR₃+…+W_N*OSNR_N。

5. the OSNR detection method according to claim 1, wherein the weight W of each OSNR computation submodel of the current input channel combination is computed by the machine-learned weight assignment model_jThe method comprises the following steps:

reading a transmitting-end OPM spectrum of the corresponding OMS section;

inputting the power of each channel read by the OPM as an input vector to a weight distribution model of machine learning;

the weight distribution model of machine learning adopts a multilayer DNN model, and inputs M-dimensional vector P of total number M of channels of corresponding optical channels OCH₁,P₂,...,P_MThe N-dimensional weight vector W corresponding to the number of OSNR computation submodels is output₁,W₂,...,W_N。

6. The OSNR detection method according to claim 5, wherein the DNN model specifically comprises:

the input layer is M-dimensional vector, the middle layer contains a specified number of full connection layers FC, and finally the SOFTMAX layer is used as the output layer to output N-dimensional vector W₁,W₂,...,W_N。

7. The OSNR detection method of claim 5 wherein the DNN model is trained by:

setting a first waveguide combination, and calculating a sub-model by using N different OSNR to obtain an OSNR predicted value OSNR of the OMS terminal_modeli；

Calculating the mean square error mse between the predicted value and the measured value of each model:wherein Nch is the number of channels of the system; OSNR_measure,ichOSNR value of the ich channel measured using experimental means; OSNR_model1,ichIch-th expert calculated using model1The OSNR value of the lane;

assuming that the mse of the ith model is minimum, an N-dimensional label (label) vector (0,0,0, …,1, … 0,0,0) is generated, wherein the ith number is 1 and the remaining N-1 numbers are 0;

read the single wave power vector P ═ (P) of the originating OPM1 at this time₁,p₂,…p_M)；

P and label (label) read by OPM1 at this time were saved₁,label₂,…label_N) Completing the collection of the first case;

switching the rest waveguide combinations one by one, and correspondingly collecting the use cases corresponding to the waveguide combinations;

inputting all collected use cases as a training set into a DNN model; with power vector P as an input quantity and tag label as a desired output quantity.

8. The OSNR detection method of claim 7, further comprising:

and constructing a typical OMS section as a training data acquisition platform, and realizing different input channel combinations by changing a wavelength selective switch WSS.

9. The OSNR detection method of claim 7, further comprising:

collecting spectrum through OSA at the tail end, and calculating OSNR of each wave by using a spectrum scanning method or a wave dropping method_measure。

10. An OSNR detecting apparatus, comprising:

at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor to perform the OSNR detecting method of any one of claims 1-9.

[ technical field ] A method for producing a semiconductor device

The invention relates to the technical field of WDM optical layers, in particular to an OSNR detection method and device.

[ background of the invention ]

The conventional Optical Signal Noise Ratio (OSNR) detection adopts an out-of-band Noise interpolation mode, and the OSNR of a wavelength division multiplexing system is a key parameter for measuring the transmission performance of the wavelength division system, and is defined as dividing the Signal power in a channel by the Noise power in a 0.1nm range at the Signal wavelength. Under DWDM dense wavelength division, it cannot be guaranteed that all channel combination scenarios have sufficient out-of-band spectrum for noise detection. With the increasing single wave rate of the wavelength division multiplexing system, the frequency spectrum utilization rate is gradually increased, and an algorithm for estimating the OSNR by using the out-of-band noise cannot be used.

In-band noise in the existing network is difficult to measure directly, and the existing measuring mode can not be effectively applied to the existing network.

The gain and noise index spectrum of the optical amplifier can be theoretically calculated to obtain a more accurate OSNR value, but the gain and noise index spectrum of the optical amplifier in the existing network changes along with input, so that the theoretically calculated OSNR is inaccurate. Polarization extinction methods exist in the industry, for example, but are not available for polarization multiplexing systems today.

In view of the above, overcoming the drawbacks of the prior art is an urgent problem in the art.

[ summary of the invention ]

The technical problem to be solved by the invention is that under DWDM dense wavelength division, sufficient out-of-band spectrum for noise detection can not be ensured under all channel combination scenes; in-band noise in the existing network is difficult to measure directly, and the existing measuring mode can not be effectively applied to the existing network.

The invention adopts the following technical scheme:

in a first aspect, the present invention provides an OSNR detecting method, including that an optical channel OCH to be calculated includes k optical multiplexed OMS segments, and a calculation model includes N OSNR calculation submodels, the method including:

obtaining Δ OSNR at each OMS segment using N OSNR computation submodels_{OMS_1},ΔOSNR_{OMS_2}，...，ΔOSNR_{OMS_k}(ii) a Thereby obtaining OSNR of each model of the N OSNR calculation submodels in the corresponding optical channel₁，OSNR₂，…，OSNR_N；

Calculating respective weights W of the N OSNR calculation submodels applied in combination with the current optical channel by a machine-learned weight assignment model_jWherein j is [1, N ]]A natural number within the interval;

the respective weights W_jAnd respectively multiplying the products by the corresponding products and summing to obtain the OSNR value of the current optical channel.

Preferably, the k optical multiplexing OMS sections include:

OSNR calculation for each traffic OCH, which is divided into k OMS segments_{OMS_i}A combination of (1); wherein i is [1, k ]]A natural number within the interval;

wherein the output OSNR of the previous OMS is the input OSNR of the next OMS.

Preferably, the OSNR of each model of the N OSNR calculation submodels in the corresponding optical channel is obtained₁，OSNR₂，…，OSNR_NThe method comprises the following steps:

each OSNR calculation submodel respectively adopts the following formula, and the OSNR calculation submodel under the corresponding optical channel is calculated in a piecewise recursion mode₁，OSNR₂，…，OSNR_N；

OSNR_{out,OMS_i}Represents the OSNR value at the output of the OMSi segment;

OSNR_{in,OMS_i}represents the OSNR value at the input end of the OMSi section;

ΔOSNR_{OMS_i}the segment of OMSi is represented by OSNR change value calculated and predicted according to the corresponding OSNR calculation submodel;

b is the channel bandwidth of each channel; bn is a constant of 12.5GHz as determined by the OSNR definition.

Preferably, the weighting W is_jAnd respectively multiplying the optical signals by corresponding products and summing to obtain the OSNR value of the current optical channel, which specifically comprises the following steps:

is obtained by solving a formula, wherein the formula is as follows:

OSNR＝W₁*OSNR₁+W₂*OSNR₂+W₃*OSNR₃+…+W_N*OSNR_N。

preferably, the weight W of each OSNR computation submodel of the current input channel combination is computed by the machine-learned weight assignment model_jThe method comprises the following steps:

reading a transmitting-end OPM spectrum of the corresponding OMS section;

inputting the power of each channel read by the OPM as an input vector to a weight distribution model of machine learning;

the weight distribution model of machine learning adopts a multilayer DNN model, and inputs M-dimensional vector P of total number M of channels of corresponding optical channels OCH₁,P₂,...,P_MThe N-dimensional weight vector W corresponding to the number of OSNR computation submodels is output₁,W₂,...,W_N。

Preferably, the DNN model specifically is:

the input layer is M-dimensional vector, the middle layer contains a specified number of full connection layers FC, and finally the SOFTMAX layer is used as the output layer to output N-dimensional vector W₁,W₂,...,W_N。

Preferably, the DNN model training method includes:

setting a first waveguide combination, and calculating a sub-model by using N different OSNR to obtain an OSNR predicted value OSNR of the OMS terminal_modeli；

Calculating the mean square error mse between the predicted value and the measured value of each model:wherein Nch is the number of channels of the system; OSNR_measure,ichOSNR value of the ich channel measured using experimental means; OSNR_model1,ichThe OSNR value for the ich channel calculated using model 1;

assuming that the mse of the ith model is minimum, an N-dimensional label (label) vector (0,0,0, …,1, … 0,0,0) is generated, wherein the ith number is 1 and the remaining N-1 numbers are 0;

read the single wave power vector P ═ (P) of the originating OPM1 at this time₁,p₂,…p_M)；

P and label (label) read by OPM1 at this time were saved₁,label₂,…label_N) Completing the collection of the first case;

switching the rest waveguide combinations one by one, and correspondingly collecting the use cases corresponding to the waveguide combinations;

inputting all collected use cases as a training set into a DNN model; with power vector P as an input quantity and tag label as a desired output quantity.

Preferably, the method further comprises:

and constructing a typical OMS section as a training data acquisition platform, and realizing different input channel combinations by changing a wavelength selective switch WSS.

Preferably, the method further comprises:

collecting spectrum through OSA at the tail end, and calculating OSNR of each wave by using a spectrum scanning method or a wave dropping method_measure。

In a second aspect, the present invention further provides an OSNR detecting apparatus for implementing the OSNR detecting method in the first aspect, where the apparatus includes:

at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor for performing the OSNR detecting method of the first aspect.

In a third aspect, the present invention also provides a non-transitory computer storage medium storing computer-executable instructions for execution by one or more processors for performing the OSNR detection method of the first aspect.

According to the invention, various OSNR calculation submodels are utilized, the OSNR values obtained by respective calculation are matched with the weights obtained by a machine-learned weight distribution model, the SANR value of the final optical channel is obtained by a weighted summation mode, and the OSNR calculation precision under the worst scene of different channel combinations is improved.

In the preferred implementation scheme, the DNN model can be trained offline, online learning does not need to be performed again aiming at different network architectures and code patterns, and the use cost is low.

[ description of the drawings ]

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required to be used in the embodiments of the present invention will be briefly described below. It is obvious that the drawings described below are only some embodiments of the invention, and that for a person skilled in the art, other drawings can be derived from them without inventive effort.

Fig. 1 is a schematic flow chart of an OSNR detecting method according to an embodiment of the present invention;

fig. 2 is a schematic flow chart of an OSNR detecting method according to an embodiment of the present invention;

FIG. 3 is a diagram of a weight assignment DNN model architecture according to an embodiment of the present invention;

fig. 4 is a schematic flow chart of an OSNR detecting method according to an embodiment of the present invention;

fig. 5 is a schematic diagram of a training data collection networking according to an embodiment of the present invention;

fig. 6 is a schematic flow chart of an OSNR detecting method according to an embodiment of the present invention;

FIG. 7 is a schematic diagram of OMS segmentation calculation according to an embodiment of the present invention;

FIG. 8 is a flow chart of a calculation provided by an embodiment of the present invention;

fig. 9 is a schematic structural diagram of an OSNR detecting apparatus according to an embodiment of the present invention.

[ detailed description ] embodiments

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

Example 1:

the embodiment 1 of the invention provides an OSNR detection method, which comprises the steps that an Optical channel OCH to be calculated comprises k Optical Multiplexing (OMS) sections, and the OSNR calculation of each service OCH divides the OSNR into OSNR variable quantity delta OSNR of the k OMS sections_{OMS_i}A combination of (1); wherein i is [1, k ]]A natural number within the interval; wherein the output OSNR of the previous OMS is the input OSNR of the next OMS. The calculation model comprises N OSNR calculation submodels, the OSNR calculation submodel generally refers to various existing single-wave OSNR calculation models, such as a spectrum scanning model, a system simulation evolution model and a related detection model, and as shown in FIG. 1, the method comprises the following steps:

in step 201, the N OSNR computation submodels are used to obtain Δ OSNR at each OMS segment_{OMS_1},ΔOSNR_{OMS_2}，...，ΔOSNR_{OMS_k}(ii) a Thereby obtaining OSNR of each model of the N OSNR calculation submodels in the corresponding optical channel₁，OSNR₂，…，OSNR_N。

Here, it should be noted that Δ OSNR is respectively provided in each OMS section_{OMS_1},ΔOSNR_{OMS_2}，...，ΔOSNR_{OMS_k}Is present in the form of a set of Δ OSNR's as described above for each OSNR computation submodel_{OMS_1},ΔOSNR_{OMS_2}，...，ΔOSNR_{OMS_k}Calculating a value; the larger the value of N is, the higher the corresponding calculation complexity is, and the higher the accuracy of the finally obtained OSNR calculation result is.

In step 202, the respective weights W of the N OSNR calculation submodels applied in combination with the current optical channel are calculated by a machine-learned weight assignment model_jWherein j is [1, N ]]A natural number within the interval.

The weight distribution model is obtained by performing a learning process through a large amount of test data, so that the weights W matched with the channel combination condition of the current optical channel and the number of the set OSRN calculation submodels can be generated_jHere, the maximum value of j is the parameter value corresponding to N in step 201.

In the following embodiments of the present invention, the selection operation of various channel combinations of the current optical channel may be involved, because one or more optical switches are disposed in the optical channel, and the configuration and switching action of the corresponding optical switch generate new channel combinations.

In step 203, the weights W are applied_jAnd respectively multiplying the products by the corresponding products and summing to obtain the OSNR value of the current optical channel.

Optionally, the method is obtained by solving a formula, where the formula is: OSNR ═ W₁*OSNR₁+W₂*OSNR₂+W₃*OSNR₃+…+W_N*OSNR_N。

According to the embodiment of the invention, various OSNR calculation submodels are utilized, the OSNR values obtained by respective calculation are matched with the weights obtained by a machine-learned weight distribution model, the SANR value of the final optical channel is obtained by a weighted summation mode, and the OSNR calculation accuracy under the worst scene of different channel combinations is improved.

In the embodiment of the present invention, the OSNR obtained in step 201 for each of the N OSNR computation submodels in the corresponding optical channel₁，OSNR₂，…，OSNR_NA specific implementation is given below, including:

each OSNR calculation submodel respectively adopts the following formula (1) to calculate and obtain the OSNR of the OSNR calculation submodel under the corresponding optical channel in a piecewise recursion mode₁，OSNR₂，…，OSNR_N；

Wherein, i is 1,2,3,. k;

OSNR_{out,OMS_i}represents the OSNR value at the output of the OMSi segment; OSNR_{in,OMS_i}Represents the OSNR value at the input end of the OMSi section; Δ OSNR_{OMS_i}The segment of OMSi is represented by OSNR change value calculated and predicted according to the corresponding OSNR calculation submodel; b is the channel bandwidth of each channel; bn is a constant of 12.5GHz as determined by the OSNR definition.

Here, the above-mentioned calculation process is presented in a more complete manner (with the first OSNR calculation submodel as an example) as follows:

for segment 1 OMS, the input OSNR value (i.e., O)SNR_{in,OMS_1}) Is known, then calculates each OMS segment Δ OSNR through the current first OSNR calculation submodel_{OMS_1},ΔOSNR_{OMS_2}，...，ΔOSNR_{OMS_k}Output OSNR value (i.e., OSNR) with respect to 1 st OMS segment_{out,OMS_1}) Can be calculated by the following formula (in this case, OSNR is the formula)_{out,OMS_1}As unknown):

for the 2 nd OMS segment, the input OSNR value is the output OSNR value of the 1 st OMS segment, and the corresponding 2 nd OMS segment output value can be calculated by the above formula (1):

at this time, the OSNR calculated by the above equation 2_{out,OMS_1}OSNR as a parameter in the above formula (3)_{in,OMS_2}Substituting the known quantity into equation (3) yields the following deformation equation (4):

at this time, in equation 4, only the OSNR exists_{out,OMS_2}An unknown quantity, the corresponding OSNR can be obtained by solving the equation_{out,OMS_2}The result value of (a); by analogy, the output OSNR value (i.e. OSNR) of the last OMS section is obtained in a recursion mode_{out,OMS_k}) The OSNR output value OSNR of the whole optical channel (which can also be described as an optical channel) is obtained₁。

The OSNR is obtained by calculating the OSNR output of the first OSNR calculation submodel on the optical channel, and the OSNR outputs of the other N-1 OSNR calculation submodels on the optical channel₂，…，OSNR_N。

In combination with the embodiment of the present invention, the weight distribution model through machine learning is also calculatedCalculating weights W of OSNR calculation submodels of the current input channel combination_jAn alternative implementation is given, as shown in fig. 2, comprising:

in step 301, an originating Optical Performance Monitor (OPM) spectrum of the corresponding OMS segment is read.

In step 302, the respective channel powers read by the OPM are input to the machine-learned weight assignment model as input vectors.

In step 303, the machine-learned weight assignment model adopts a multi-layer DNN model, and an M-dimensional vector P corresponding to the total number M of channels possessed by the optical channel OCH is input₁,P₂,...,P_MThe N-dimensional weight vector W corresponding to the number of OSNR computation submodels is output₁,W₂,...,W_N。

The DNN model is shown in fig. 3, and specifically includes: the input layer is M-dimensional vector, the middle layer contains a specified number of full connection layers FC, and finally the SOFTMAX layer is used as the output layer to output N-dimensional vector W₁,W₂,...,W_N。

The DNN model training method is shown in fig. 4, and includes:

in step 401, a first waveguide combination is set, and a sub-model is calculated by using N different OSNR to obtain an OSNR predicted value OSNR of the OMS terminal_modeli. Wherein, OSNR_modeliIn the implementation process, the OSNR value of each channel terminal is expressed as a corresponding model modeli, for example, the OSNR described below_model1,ich。

In step 402, the mean square error mse between the predicted value and the measured value of each model is calculated:wherein Nch is the number of channels of the system, wherein, in an optional implementation, the number of channels Nch is consistent with the number of channels M; OSNR_measure,ichThe OSNR value of the ich channel measured by using experimental means (the spectrum can be collected by OSA at the tail end, and the OSNR of each wave is calculated by using a spectrum scanning method or a wave dropping method_measure)；OSNR_model1,ichThe OSNR value for the ich channel calculated using model 1.

Assuming that the mse of the ith model is minimum, an N-dimensional label (label) vector (0,0,0, …,1, … 0,0,0) is generated, where the ith number is 1 and the remaining N-1 numbers are 0.

In step 403, the single-wave power vector P ═ P (P) of the originating OPM1 at this time is read₁,p₂,…p_M)。

In step 404, P and label (label) read by OPM1 at this time are saved₁,label₂,…label_N) And finishing the acquisition of the first case.

In step 405, the remaining waveguide combinations are switched one by one, and the use cases corresponding to each waveguide combination are correspondingly collected.

In step 406, all collected use cases are used as a training set and input into the DNN model; with power vector P as an input quantity and tag label as a desired output quantity. Preferably, the model parameter optimization training can also be carried out by a back propagation gradient descent method.

In the training process, a typical OMS section can be set up as a training data acquisition platform, and different input channel combinations are realized by changing a Wavelength Selective Switch (WSS).

Example 2:

the embodiment of the present invention takes 3 OMS segments, 96-wave systems, and 3 OSNR computation submodels as examples:

we use 3 typical submodels:

1. OPM scanning model: and scanning the spectrum by using an OPM device, and calculating the OSNR of the signal channel in an out-of-band interpolation mode. The advantages are that: a few-wave scene and a scene with larger channel interval. Disadvantages are that: continuous channel scanning is difficult.

2. Gain, NF spectrum theoretical model: and the OSNR is deduced by theoretical calculation by using the gain and NF spectrum of factory calibration, and the factory calibration can only calibrate a full spectrum and has larger spectrum type error when the wave is small. The advantages are that: full wave, multi-wave scenes. Disadvantages are that: the few-wave gain NF spectrum has larger deviation.

3. Pilot-assisted model: the OSNR is estimated using the received SNR of the pilot and the non-linear noise cannot be excluded. The advantages are that: a few-wave scene and a less nonlinear scene. Disadvantages are that: the multi-wave scene crosstalk is large.

An off-line training data acquisition platform as shown in fig. 5 was first set up.

First, data tag generation:

as shown in fig. 5, after a 2-span-3 optical amplifier OMS segment is built as a training data acquisition platform, as shown in fig. 6, the embodiment of the present invention includes:

in step 501, different input channel combinations (typically using, for example, 1000 different channel combinations) are implemented by changing the WSS.

In step 502, spectra are collected through OSA at the end, and the OSNR of each wave is calculated using spectral scanning or evanescent wave method_measure。

In step 503, the predicted OSNR value OSNR of the OMS terminal is obtained by using 3 different OSNR calculation submodels_modeli。

In step 504, the mean square error mse between the predicted value and the measured value of each model is calculated:

assuming that the mse of the 2 nd model is minimal, a 3-dimensional label (label) vector (0,1,0) is generated, the second labeled 1, and the rest 0.

In step 505, the single-wave power vector P ═ P (P) of the originating OPM1 at this time is read₁,p₂,…p₉₆)。

In step 506, P and label (label) read by OPM1 at this time are saved₁,label₂,…label₃)。

In step 507, the WSS is modified again, and switched to another channel combination to collect the next use case in step 501.

Second, DNN model training:

DNN model As shown in FIG. 3, the input layer is a 96-dimensional vector, the tundishComprising a plurality of fully connected layers (FC) and finally a SOFTMAX layer, outputting a 3-dimensional vector (W)₁,W₂,W₃)。

And inputting all the use cases generated in the step one into the DNN model as a training set. With power vector P as an input quantity and tag label as a desired output quantity.

And performing model parameter optimization training by a back propagation gradient descent method.

Third, model using method:

as shown in fig. 7, the OSNR calculation for each traffic OCH can be divided into OSNR variations (Δ OSNR) of 3 OMS segments_{OMS_i}) Combinations of (a) and (b). (addressing the issue of out-of-band and in-band ASE inconsistency) where the output OSNR of the previous OMS segment is equal to the input OSNR of the next segment, it can be used for segment-wise recursion.

The specific formula is as follows for calculating the output OSNR of a single OMS segment:

OSNR_{out,OMS_i}represents the OSNR value at the output of the OMSi segment; OSNR_{in,OMS_i}Represents the OSNR value at the input end of the OMSi section; Δ OSNR_{OMS_i}The segment of OMSi is represented by OSNR change value calculated and predicted according to the corresponding OSNR calculation submodel; b is the channel bandwidth of each channel; bn is a constant of 12.5GHz as determined by the OSNR definition.

Obtaining respective 3 groups of OMS segments delta OSNR by using 3 OSNR calculation submodels_{OMS_1},ΔOSNR_{OMS_2},ΔOSNR_{OMS_3}. The corresponding method process may also refer to the flowchart shown in fig. 8.

Calculating each model weight W of the current input channel combination through the weight distribution model of machine learning of off-line training₁,W₂,W₃. The calculation flow of the weight distribution model of machine learning is as follows:

the originating OPM spectrum P of the corresponding OMS segment is read.

And inputting the power of each channel read by the OPM as an input vector to a machine learning model.

The machine learning model adopts a multilayer DNN model and the input is 96-dimensional vector (P)₁,P₂,...,P₉₆) The output is a 3-dimensional vector (W)₁,W₂,W₃)。

From the weight W of the previous step₁,W₂,W₃And OSNR of the respective models₁,OSNR₂,OSNR₃The result is superposed to obtain the final calculated OSNR ═ W of each channel₁*OSNR₁+W₂*OSNR₂+W₃*OSNR₃。

Example 3:

fig. 9 is a schematic structural diagram of an OSNR detecting apparatus according to an embodiment of the present invention. The OSNR detecting apparatus of the present embodiment includes one or more processors 21 and a memory 22. In fig. 9, one processor 21 is taken as an example.

The processor 21 and the memory 22 may be connected by a bus or other means, and fig. 9 illustrates the connection by a bus as an example.

The memory 22, which is a non-volatile computer-readable storage medium, may be used to store non-volatile software programs and non-volatile computer-executable programs, such as the OSNR detecting method in embodiment 1. The processor 21 executes the OSNR detection method by executing non-volatile software programs and instructions stored in the memory 22.

The memory 22 may include high speed random access memory and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid state storage device. In some embodiments, the memory 22 may optionally include memory located remotely from the processor 21, and these remote memories may be connected to the processor 21 via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The program instructions/modules are stored in the memory 22 and, when executed by the one or more processors 21, perform the OSNR detecting method of embodiment 1 described above, for example, perform the steps shown in fig. 1, fig. 2, fig. 4, fig. 6, and fig. 8 described above.

It should be noted that, for the information interaction, execution process and other contents between the modules and units in the apparatus and system, the specific contents may refer to the description in the embodiment of the method of the present invention because the same concept is used as the embodiment of the processing method of the present invention, and are not described herein again.

Those of ordinary skill in the art will appreciate that all or part of the steps of the various methods of the embodiments may be implemented by associated hardware as instructed by a program, which may be stored on a computer-readable storage medium, which may include: read Only Memory (ROM), Random Access Memory (RAM), magnetic or optical disks, and the like.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.