CN116128034A - Optical fiber structure two-dimensional feature extraction and optical characteristic prediction method based on convolutional neural network - Google Patents
Optical fiber structure two-dimensional feature extraction and optical characteristic prediction method based on convolutional neural network Download PDFInfo
- Publication number
- CN116128034A CN116128034A CN202210659164.XA CN202210659164A CN116128034A CN 116128034 A CN116128034 A CN 116128034A CN 202210659164 A CN202210659164 A CN 202210659164A CN 116128034 A CN116128034 A CN 116128034A
- Authority
- CN
- China
- Prior art keywords
- optical fiber
- optical
- dimensional
- neural network
- convolutional neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/067—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Testing Of Optical Devices Or Fibers (AREA)
Abstract
Aiming at the problems of low efficiency, high labor cost and the like in the conventional optical fiber structure optimization design method, the invention provides an optical fiber structure two-dimensional feature extraction and optical characteristic prediction method based on a convolutional neural network, wherein an optical fiber cross section two-dimensional free distribution structure is used as a parameter space, any free distribution structure of an optical fiber is converted into a two-dimensional numerical matrix, the two-dimensional structural feature is extracted by the convolutional neural network, and an efficient and accurate mapping network relation between the optical characteristic and the two-dimensional structure of the optical fiber cross section free distribution is established by regression calculation. The invention takes multi-core optical fiber as an example, and accurately predicts the mode field area, dispersion and effective refractive index. The method has the advantages of high calculation speed and high accuracy, avoids the limitation of the existing mapping network based on the fixed characteristic parameters on the optimal design of the optical fiber structure, can be combined with various optimization algorithms, and is expected to realize the optimal design of the optical fiber structure with high degree of freedom. The technical scheme is also applicable to the prediction of other optical structures.
Description
Technical Field
The invention belongs to the technical field of optical fibers, and relates to an optical fiber cross section two-dimensional structural feature extraction and optical characteristic prediction analysis method based on a convolutional neural network.
Background
The development of the field of optical fiber photonics is not separated from the development and design of various novel functional optical fiber photon structures and devices. The research and development of the optical fiber photon structure at present mainly depends on a numerical simulation and manual error testing framework of intensive calculation, utilizes various numerical simulation calculations through a priori physical mechanism, and attempts to obtain ideal optical characteristics by manually and repeatedly adjusting structural characteristic parameters. The structural design optimization based on the manual trial-and-error framework is very dependent on experience and intuition of researchers, because the relation between an optical structure and characteristics is often very complex and hidden, mutual influence and restriction exist among various structural parameters, the researchers are difficult to obtain clear optimization and design directions, a great deal of repeated redundancy work is often accompanied in the design process, the design method is time-consuming and laborious, physical intuition and system fine adjustment are relied on, great breakthrough or improvement is difficult, and the structural parameters with optimal or limit optical characteristics are difficult to find. On the other hand, the numerical simulation method (such as plane wave expansion method, finite difference method, finite element method, etc.) with intensive calculation has stable algorithm and high accuracy, but has low convergence rate, large calculation amount and low efficiency, and further affects the design and research and development efficiency.
In recent years, the neural network is used as a cross subject in multiple fields, and has important application value in various fields such as semantic recognition, image processing, biochemistry, micro-nano optics and the like. The complex nonlinear relation between mapping parameters can be quickly found through the strong analysis and calculation capabilities of the neural network, and powerful support is provided for the complex optical fiber structure design of multi-parameter optimization. By defining a plurality of optical structure characteristic parameters in advance, the neural network can quickly establish the mapping relation between the optical structure parameters and the optical characteristics, and the calculation speed is several orders of magnitude faster than that of the traditional numerical simulation, so that the optical structure calculation efficiency is greatly improved [ Optics Express,27 (25): 36414-36425 (2019); optics Express,28 (15): 21668-21681 (2020); optics Letters,46 (6): 1454-1457 (2021), in combination with other optimization algorithms [ Optics Express,28 (15): 21971-21981 (2020); photonic Research,9 (6): B247-B252 (2021), which is more favorable for the automatic global optimization and improved design of optical structures, has important significance for the research and development of novel optical devices and limit optical performance structures.
However, the mapping networks reported at present are all mapping calculations based on fixed characteristic parameters, and although the optical structure can be accurately and rapidly predicted, there are great limitations: several feature parameters (such as assuming that the optical fiber is a five-ring structure and defining the structural parameters of each ring, however, the optimal optical fiber structure is likely not to be a five-ring structure at all) need to be assumed in advance according to intuitive experiences, and these intuitive assumptions fix the optical fiber structure features in advance, but rather greatly limit the parameter design space of the optical structure and limit the design thought, because we often have difficulty in knowing the geometric feature parameters of the optimal target structure in advance.
Disclosure of Invention
Aiming at a plurality of problems existing in the conventional optical fiber structure optimization design, the invention provides an optical fiber structure two-dimensional feature extraction and optical characteristic prediction method based on a convolutional neural network.
The invention aims to accurately predict the optical characteristics of different optical fibers by adopting a trained convolutional neural network, taking a multi-core optical fiber as an example, wherein the multi-core optical fiber cladding material is pure silica, the fiber core material is doped silica, the fiber core part is formed by closely arranging a plurality of high-refractive-index fiber cores, and the multi-fiber-core structure forms a whole and can support the supermode transmission of a large mode field area. By varying the structural parameters, the optical properties of the different supermodes that the core can support for transmission may be varied. The optical fiber optical characteristic prediction analysis method based on the convolutional neural network can rapidly and accurately predict and calculate corresponding optical fiber optical parameters according to different material refractive indexes and geometric size distribution, wherein the optical fiber optical parameters comprise mode field area, chromatic dispersion, effective refractive index and the like which can support a transmission mode. The technical scheme is also suitable for predicting the characteristics of other optical structures.
The technical scheme adopted by the invention comprises the following steps:
1. calculating optical characteristics of different modes which can support transmission in an optical fiber model for collecting different structural parameters by using a traditional finite element simulation method, such as mode field area, chromatic dispersion, mode effective refractive index and the like;
2. selecting a proper sampling interval, and converting the refractive index free distribution of the cross section materials of different optical fiber structures into a two-dimensional data matrix, wherein the numerical value of each point in the data matrix represents the refractive index of the material at the corresponding optical fiber structure;
3. constructing a proper regression prediction network structure by utilizing a convolutional neural network;
4. training a regression prediction network by using the collected data set and storing a model;
5. testing the performance of the regression prediction network model by using a test set;
6. and storing the most suitable regression prediction network for rapidly and accurately predicting the optical characteristics of different multi-core optical fibers.
The invention provides a convolutional neural network-based optical fiber cross section two-dimensional structural feature extraction and optical characteristic prediction analysis method, which has the advantages that:
1. in the aspect of establishing a data set, the two-dimensional structural features of the free distribution of the cross section of the optical fiber are converted into a two-dimensional data matrix, the size of the matrix can be flexibly adjusted according to the sampling interval (similar to the image processing, the pixels and the resolution of pictures can be freely adjusted according to the needs), the acquired and calculated optical fiber structure is the refractive index free distribution of the material of the whole cross section, is not limited by fixed characteristic parameters or size variables, and provides thinking and approaches for the more flexible optical fiber structure optimization design.
2. The numerical definition of the two-dimensional matrix for the parameter conversion of the optical fiber cross section structure is very flexible, can be defined as the two-dimensional distribution of the refractive index of the material, can be defined as the absorptivity of the material or the characteristics of other special materials, and provides convenience for the optimal design of various optical fiber structures or special optical fibers.
3. Compared with the traditional numerical simulation method, the method has absolute advantages in terms of calculation speed, the time required for calculating the optical characteristics of a single optical fiber structure is between 0.02 and 0.03 seconds, and the traditional mathematical physical numerical simulation method usually needs minutes or even hours to finish the calculation of the optical fiber structure. Therefore, the method has an order of magnitude improvement in speed, is more beneficial to being combined with various intelligent optimization algorithms, and is beneficial to further utilizing a computer to automatically optimize and design the optical fiber structure.
4. Computing optical structures with common fully connected artificial neural networks [ Optics Express,27 (25): 36414-36425 (2019); optics Express,28 (15): 21668-21681 (2020); optics Letters,46 (6): 1454-1457 (2021), the two-dimensional matrix parameters input by the method represent the two-dimensional structural characteristics of the free distribution of the cross section of the whole optical fiber, and effectively avoid the limitation of the artificial assumption of the dimension characteristic parameters in advance to the free design optimization of the optical fiber structure.
5. Feature parameters are reduced in feature extraction of the image, unimportant features are ignored, important parameters are reserved, the number of training parameters can be greatly reduced when the model is trained, overfitting can be prevented, and meanwhile the model training speed is greatly increased.
6. Based on a convolutional neural network, the two-dimensional numerical matrix of the refractive index distribution of the material of the optical fiber is subjected to convolutional operation, characteristics are extracted, regression calculation is performed on the final fully-connected layer, and the predicted optical characteristics such as the mode field area, the dispersion, the effective refractive index and the like are very high in accuracy compared with the traditional numerical simulation method. The results prove that the optical characteristics of the optical fiber can be predicted according to the refractive index of the material with the two-dimensional free distribution of the cross section of the optical fiber, and the prediction accuracy is extremely high.
Drawings
Fig. 1: the predicted example optical fiber structure diagram comprises a plurality of fiber cores (5 to 8 fiber cores are unequal), a certain interval is reserved between the fiber cores, the positions of each fiber core from the center of a cladding are the same, the cladding is pure silicon dioxide, and the fiber cores form a whole and can support the overmode transmission of large mode area;
fig. 2: a two-dimensional data matrix schematic diagram of refractive index conversion of a two-dimensional structure material freely distributed on the cross section of the optical fiber;
fig. 3: the invention provides a two-dimensional characteristic extraction and optical characteristic prediction method of an optical fiber structure based on a convolutional neural network;
fig. 4: the convolutional neural network and the full-connection regression calculation network adopted by the invention are structured schematically;
fig. 5: testing by using a test set to obtain a comparison graph of predicted values and true values of the mode field areas capable of supporting transmission of 6 modes;
fig. 6: testing by using a test set to obtain a comparison graph of a dispersion predicted value and a true value which can support transmission of 6 modes;
fig. 7 and 8: the invention uses the test set to test, and obtains a comparison chart of predicted values and true values of the effective refractive indexes of 6 modes capable of supporting transmission;
Detailed Description
The invention and the technical scheme are further specifically described below with reference to the accompanying drawings.
A method for extracting two-dimensional characteristics of optical fiber structure and predicting optical characteristics based on convolutional neural network uses multi-core optical fiber as an example to predict the mode field area, dispersion and mode effective refractive index capable of supporting transmission supermode. The cross-sectional structure of the used multicore fiber is shown in fig. 1. The core portion is formed of a plurality of (5 to 8 cores are not equal) high refractive index doped silica cores, each of which is the same distance from the center of the fiber. By changing the parameters of the optical fiber structure (such as the radius, the number and the distribution of fiber cores, etc.), different optical fiber structures can be obtained, and the two-dimensional data matrixes of the cross sections corresponding to the different optical fiber structures and the mode field area, the dispersion and the effective refractive index of the transmission supermode can be acquired. Example acquired mode is HE that can support transmission 11 、HE 21 、HE 31 、HE 41 、EH 11 、EH 21 The mode field area, dispersion and mode effective refractive index of these 6 modes.
For example, a multicore fiber is selected, the number of fiber cores is set to 5, the radius of each fiber core is 4 μm, the refractive index of each fiber core is 0.03 higher than that of pure silica, and the distance between the center of each fiber core and the center of the fiber is 7 μm. The two-dimensional distribution of the refractive index of the fiber cross section material is converted into a two-dimensional numerical matrix, and the matrix size is set to be 227×227. With this parameter, a two-dimensional data matrix diagram of the refractive index distribution of the cross-sectional material of the multi-core fiber model is shown in fig. 2. Fig. 2 shows only one example of the conversion, and the conversion mode can be flexibly adjusted, for example, the size of the matrix can be adjusted according to the requirement, and the numerical meaning (refractive index, material absorptivity and the like) of the conversion can be adjusted.
Fig. 3 is a flowchart of the present invention, first, finite element simulation software COMSOL Multiphysics is used to collect the data sets of the convolutional neural network, which are corresponding to the mode field areas, chromatic dispersion, effective refractive indexes, etc. corresponding to several modes capable of supporting transmission under different wavelengths of different optical fiber structures, and meanwhile, the refractive index distribution of the two-dimensional material of the cross section of the optical fiber structure is converted into a two-dimensional data matrix. Then building a proper neural network model, putting the acquired training data set into the built neural network model for training, obtaining a neural network model with lower loss and higher prediction accuracy, and storing the model. And finally, testing the trained network model by using a test data set, calculating a predicted value, and comparing the predicted value with a true value of the test data set to prove the accuracy of the neural network prediction.
Fig. 4 shows a neural network model capable of predicting the optical characteristics of a two-dimensional free distribution structure of an optical fiber cross section, and the size of a two-dimensional data matrix collected is 227×227. Therefore, the input layer of the convolutional neural network is (227, 227,1), and the convolutional layer obtained through multiple times of network adjustment is 3 layers in total; after all convolution layers, the maximum pooling layer and the Relu activation function are used to further reduce the number of weight parameters. And finally, obtaining 179776 nodes by the two-dimensional matrix through the convolution layer. In the full-connection layer, nodes and weights of the network are adjusted for multiple times, and a regression network finally obtained comprises an input layer and an output layer; in the input layer, besides 179776 nodes obtained through the convolution layer, wavelength nodes are added, and meanwhile, the numerical value of the wavelength is subjected to characteristic scaling and normalized to be between 0 and 1. Finally, the network uses an Adam optimizer, the initial learning rate is adjusted to 0.00005, and 500 times of iterative training are carried out.
The input layer of the finally trained predictive neural network is a two-dimensional data matrix with an optical fiber cross-section structure, and the output layer is a corresponding mode field area, dispersion and effective refractive index under different wavelengths. Fig. 5, 6, 7 and 8 show the prediction results of the neural network test using the test data set, and the accuracy of the prediction network model is verified. The figure shows predictions of 6 types of mode optical characteristics that a multi-core fiber can support for transmission, including mode field area, dispersion, and mode effective refractive index at different wavelengths. The optical fiber structure two-dimensional characteristic extraction and optical characteristic prediction method based on the convolutional neural network can rapidly and accurately predict and calculate the corresponding optical characteristic according to the refractive index and geometric size distribution of any two-dimensional material of the optical fiber cross section, and compared with the traditional numerical simulation method, the error is small, and meanwhile the prediction calculation time is improved by orders of magnitude. The technical scheme steps are also applicable to the prediction of other optical structural characteristics.
The invention is not described in detail in the field of technical personnel common knowledge.
Claims (4)
1. A method for extracting two-dimensional characteristics of an optical fiber structure and predicting optical characteristics based on a convolutional neural network, the method comprising the following steps:
step 1: converting the two-dimensional freely distributed structural characteristics of the cross section of the optical fiber into a two-dimensional numerical matrix;
step 2: collecting an optical characteristic data set of the optical fiber by using a numerical simulation method (such as finite elements and the like);
step 3: constructing an optical fiber structural feature extraction and regression calculation prediction neural network structure;
step 4: training the network by using the collected data set and storing a model;
step 5: testing the predicted computing performance of the network model by using the test set;
step 6: the most suitable prediction network is saved for fast and accurate calculation of the optical characteristics of different optical fibers.
2. The convolutional neural network-based optical fiber structure two-dimensional feature extraction and optical property prediction method as claimed in claim 1, wherein: in step 1, the two-dimensional freely distributed structural features of the cross section of the optical fiber are converted into a two-dimensional numerical matrix by selecting a proper sampling interval (resolution), the numerical value of each point in the matrix represents the material characteristic (such as the refractive index, the absorptivity and the like of the material) at the corresponding optical fiber structure, and the size of the matrix can be freely adjusted according to the requirement.
3. The convolutional neural network-based optical fiber structure two-dimensional feature extraction and optical property prediction method as claimed in claim 1, wherein: in step 3, in the process of constructing the prediction network, after feature extraction is performed on the two-dimensional data matrix by using the convolutional neural network, the extracted structural features and the wavelength nodes are put into the fully-connected input layer together at the fully-connected layer part, so as to obtain regression prediction results under different wavelengths.
4. The convolutional neural network-based optical fiber structure two-dimensional feature extraction and optical property prediction method as claimed in claim 1, wherein: in step 4, before training is performed by using the collected data set, the data needs to be separated according to different types of modes, and the data is respectively stored separately according to three optical characteristics of mode field area, chromatic dispersion, effective refractive index and the like, and is respectively put into a regression prediction model for training.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210659164.XA CN116128034A (en) | 2022-06-13 | 2022-06-13 | Optical fiber structure two-dimensional feature extraction and optical characteristic prediction method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210659164.XA CN116128034A (en) | 2022-06-13 | 2022-06-13 | Optical fiber structure two-dimensional feature extraction and optical characteristic prediction method based on convolutional neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116128034A true CN116128034A (en) | 2023-05-16 |
Family
ID=86296051
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210659164.XA Pending CN116128034A (en) | 2022-06-13 | 2022-06-13 | Optical fiber structure two-dimensional feature extraction and optical characteristic prediction method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116128034A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116593989A (en) * | 2023-06-15 | 2023-08-15 | 宁波麦思捷科技有限公司武汉分公司 | Troposphere waveguide inversion method and system based on radar sea clutter |
-
2022
- 2022-06-13 CN CN202210659164.XA patent/CN116128034A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116593989A (en) * | 2023-06-15 | 2023-08-15 | 宁波麦思捷科技有限公司武汉分公司 | Troposphere waveguide inversion method and system based on radar sea clutter |
| CN116593989B (en) * | 2023-06-15 | 2023-11-21 | 宁波麦思捷科技有限公司武汉分公司 | Troposphere waveguide inversion method and system based on radar sea clutter |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110851782B (en) | Network flow prediction method based on lightweight space-time deep learning model | |
| CN113094860B (en) | Industrial control network flow modeling method based on attention mechanism | |
| JP2013074695A (en) | Device, method and program for predicting photovoltaic generation | |
| CN116128034A (en) | Optical fiber structure two-dimensional feature extraction and optical characteristic prediction method based on convolutional neural network | |
| CN110555546B (en) | Updating method and system for optical performance degradation trend prediction | |
| CN112989701B (en) | Wavelength tuning method of SGDBR tunable semiconductor laser | |
| CN111898316A (en) | Construction method and application of super-surface structure design model | |
| CN111752891B (en) | IP core mapping method for optical network on chip | |
| Zhenlong et al. | The prediction model of cotton yarn intensity based on the CNN-BP neural network | |
| CN111722139A (en) | Lithium battery health monitoring model self-learning method based on micro-continuous mapping | |
| CN113887694A (en) | Click rate estimation model based on characteristic representation under attention mechanism | |
| CN117371320A (en) | Photonic crystal fiber prediction method and system based on reinforced gray wolf optimization algorithm | |
| CN116502779A (en) | Traveling merchant problem generation type solving method based on local attention mechanism | |
| Chen et al. | Joint sparse neural network compression via multi-application multi-objective optimization | |
| Behera et al. | Machine Learning-based Inverse Model for Few-Mode Fiber Designs. | |
| CN111985153B (en) | Silicon-based photonic device simulation method and system based on neural network and genetic algorithm | |
| CN112505829B (en) | Design method of mode selective coupler | |
| CN113705031B (en) | Nano antenna array electromagnetic performance prediction method based on deep learning | |
| Jha et al. | Performance evaluation of DWDM optical transmission system using deep learning technique | |
| CN114967126A (en) | Silica-based optical micro-ring filter reverse design method based on sparsity calculation | |
| CN114970336A (en) | Traffic flow prediction method based on improved GCN-attention algorithm | |
| CN114547473A (en) | Crowdsourcing task data recommendation method based on decision model and genetic matrix decomposition method | |
| Niu et al. | Neural architecture search based on particle swarm optimization | |
| Xiao et al. | Predicting optical properties of different photonic crystal fibers from 2D structural images using convolutional neural network and transfer learning | |
| Xue et al. | RARTS: an efficient first-order relaxed architecture search method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |