A method for extracting spatiotemporal characteristics of photonic crystal spatial transmission spectrum

Abstract

The present invention provides a method for extracting spatiotemporal characteristics of a spatial transmission spectrum of a photonic crystal, the method comprising: obtaining a training data set containing a spatial transmission spectrum of a photonic crystal and structural parameters; constructing a neural network model based on ECA-CNN fused with a dual-channel RNN; using the acquired training data set to train the constructed neural network model, and using an optimization algorithm to improve the model feature extraction and parameter prediction performance; and then evaluating the model performance through a variety of regression prediction evaluation indicators. In the training process, the spatial pixel features extracted by the ECA-CNN convolution module and the image time series features extracted by the dual-channel RNN are fused, and the image is comprehensively analyzed based on the fused features to complete the structural design of the photonic crystal. By adopting the present invention, the image can be analyzed using more comprehensive image features after fusion, thereby improving the parameter prediction performance based on complex images.

Description

Technical Field

[0001] The present invention relates to the fields of regression prediction and artificial intelligence, and in particular to a method for extracting the spatiotemporal characteristics of a photonic crystal spatial transmission spectrum in parallel by fusing an ECA-CNN with a dual-channel RNN. Background Art

[0002] Photonic crystals are structures formed by periodically arranging dielectric materials with different dielectric constants in space, which can control the light waves propagating in them to form a band structure. People usually use numerical calculations based on electromagnetic theory to analyze the characteristics of photonic crystals. Common methods include scattering matrix method and finite difference time domain method, but they all inevitably have some defects.

[0003] The development of deep learning has provided a new perspective for photonic crystal research, and many optical applications have gradually combined and benefited from deep network models. Previous researchers have used deep convolutional networks to study the beam pattern of few-mode optical fibers, and predicted the weight and phase difference of the mode field by analyzing its spatial pixel characteristics; they have also used recurrent networks to study the transmission and reflection spectra of photonic crystals, and designed the structure of photonic crystals by analyzing their timing characteristics.

[0004] The spatial transmission spectrum of photonic crystals is a special image that contains both temporal features and spatial pixel features. Due to the complexity of the image, a single network model is not sufficient to extract complete feature information. The network model based on fusion features presented in this article is a good idea. Summary of the invention

[0005] The purpose of the present invention is to provide a deep neural network of ECA-CNN fused with dual-channel RNN, which can fuse the spatial pixel features extracted by the convolutional network and the image temporal features extracted by the dual-channel recurrent network, and use the more comprehensive image features after fusion to analyze the image to improve the parameter prediction performance based on complex images, thereby effectively solving the problem that the current regression prediction method is not comprehensive enough in extracting features of complex images, resulting in low parameter prediction performance.

[0006] The present invention provides a method for extracting the spatiotemporal characteristics of the spatial transmission spectrum of a photonic crystal by fusing an ECA-CNN with a dual-channel RNN in parallel. The method comprises:

[0007] (1) obtaining spatial transmission spectrum samples of a one-dimensional photonic crystal and corresponding photonic crystal structural parameters as training data sets, and performing data preprocessing operations on the structural parameters;

[0008] (2) Design a neural network model that combines ECA-CNN with dual-channel RNN;

[0009] (3) Use the acquired data set to train the constructed neural network model and use optimization algorithms to improve the model feature extraction and parameter prediction performance;

[0010] (4) The performance of the network model is evaluated by calculating the mean square error, root mean square error, mean absolute error, mean relative error, mean absolute percentage error and symmetric mean absolute percentage error indicators.

[0011] Furthermore, the training data set is obtained by collecting different types of photonic crystal spatial transmission spectra as training samples, dividing the corresponding structural parameters by their respective maximum values ​​for normalization operation as label data, and verifying the universality of the network model through multiple groups of experiments.

[0012] Furthermore, the neural network model structure includes a feature fusion module and a fully connected layer. The feature fusion module is used to extract the spatial pixel features and image temporal features of the spatial transmission spectrum and perform fusion transmission; the fully connected layer is used to flatten the fusion features output by the last feature fusion module and map them into structural parameters of the photonic crystal.

[0013] Furthermore, the feature fusion module includes a convolution module and a dual-channel recurrent layer. In the convolution module, an ECA attention mechanism is connected after the convolution layer to enhance the convolution layer's ability to extract spatial pixel features with a local cross-channel interaction strategy without dimensionality reduction; the dual-channel recurrent layer extracts and fuses the row and column time series features of the input feature map through two parallel recurrent layers to obtain the overall time series features of the feature map. The new feature map obtained after the two features are fused in a 1:1 ratio is used as the input of the next layer of feature fusion module to continue feature extraction.

[0014] Furthermore, the neural network model uses the Sigmoid nonlinear activation function during the training process, and its mathematical expression is:

[0015]

[0016] Among them, x represents each weight value of the flattened feature sequence. Through this formula, each weight value can be limited between 0 and 1. In addition, the adaptive algorithm ReduceLROnPlateau is used to dynamically adjust the learning rate. When the verification loss does not change within a certain iteration cycle, the learning rate will be attenuated at a specified rate. A variety of optimization methods are used to improve the feature extraction and parameter prediction performance of the model.

[0017] Furthermore, the multiple network model performance evaluation indicators, including mean square error, root mean square error, mean absolute error, mean relative error, mean absolute percentage error and symmetric mean absolute percentage error, are used to evaluate the prediction performance of the model:

[0018]

[0019] Where n represents the number of samples, Represents the predicted value, y i Represents the actual value. The six evaluation indicators reflect the differences between the real data and the predicted data in different ways. The higher the model performance, the closer the evaluation indicator value is to 0.

[0020] The beneficial effects brought about by the technical solution provided by the present invention include at least:

[0021] In the present invention, a training set including the spatial transmission spectrum of photonic crystals and related structural parameters is obtained; a feature fusion network model based on ECA-CNN fused with dual-channel RNN is constructed; the constructed network model is trained using the acquired data set, and an optimization algorithm is used to improve the model feature extraction and parameter prediction performance; and the model performance is evaluated using a variety of regression prediction evaluation indicators. In this process, the spatial pixel features extracted by the ECA-CNN convolution module and the image time series features extracted by the dual-channel RNN are fused, and the image is analyzed using the more comprehensive image features after fusion to improve the parameter prediction performance based on complex images, thereby effectively solving the problem that the current regression prediction method is not comprehensive enough in extracting features of complex images, resulting in low parameter prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of a method provided by the present invention for fusing ECA-CNN with dual-channel RNN to extract the spatiotemporal characteristics of the spatial transmission spectrum of a photonic crystal in parallel.

[0023] Figure 2 It is a schematic diagram of neural network modeling of ECA-CNN fused dual-channel RNN provided by the present invention.

[0024] Figure 3 It is a detailed schematic diagram of the network model feature fusion module provided by the present invention.

[0025] Figure 4 It is a detailed schematic diagram of the ECA-CNN module of the network model provided by the present invention.

[0026] Figure 5 It is a detailed schematic diagram of the dual-channel RNN module of the network model provided by the present invention.

[0027] Figure 6 It is a schematic diagram of the process of obtaining a photonic crystal spatial transmission spectrum data set provided by the present invention. DETAILED DESCRIPTION

[0028] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0029] like Figure 1 As shown, the present invention provides a method for extracting the spatiotemporal characteristics of the spatial transmission spectrum of photonic crystals in parallel by fusing ECA-CNN with dual-channel RNN. Specific embodiments include:

[0030] (1) obtaining spatial transmission spectrum samples of ABAB-type one-dimensional photonic crystals and corresponding photonic crystal structural parameters as training data sets, and performing data preprocessing operations on the structural parameters;

[0031] (2) Design a neural network model that combines ECA-CNN with dual-channel RNN;

[0032] (3) Use the acquired data set to train the constructed neural network model and use optimization algorithms to improve the model feature extraction and parameter prediction performance;

[0033] (4) The performance of the network model is evaluated by calculating indicators such as mean square error, root mean square error, mean absolute error, mean relative error, mean absolute percentage error and symmetric mean absolute percentage error.

[0034] In the specific implementation of the aforementioned method for extracting the spatiotemporal characteristics of the spatial transmission spectrum of photonic crystals in parallel based on ECA-CNN fusion dual-channel RNN, further, the acquisition of the training data set collects different types of spatial transmission spectra of photonic crystals as training samples, and the corresponding structural parameters are divided by their respective maximum values ​​for normalization operation as label data, and the universality of the network model is verified through multiple groups of experiments.

[0035] In the above-mentioned specific implementation of the method for extracting the spatiotemporal characteristics of the spatial transmission spectrum of photonic crystals in parallel based on ECA-CNN fusion dual-channel RNN, further, as Figure 2 As shown, the neural network model structure includes a feature fusion module and a fully connected layer. The feature fusion module is used to extract the spatial pixel features and image temporal features of the spatial transmission spectrum and perform fusion transmission; the fully connected layer is used to flatten the fusion features output by the last feature fusion module and map them into the structural parameters of the photonic crystal.

[0036] In the above-mentioned specific implementation of the method for extracting the spatiotemporal characteristics of the spatial transmission spectrum of photonic crystals in parallel based on ECA-CNN fusion dual-channel RNN, further, as Figure 3 , 4As shown in Figure 5, the feature fusion module includes a convolution module and a dual-channel recurrent layer. In the convolution module, the ECA attention mechanism is connected after the convolution layer to enhance the convolution layer's ability to extract spatial pixel features with a local cross-channel interaction strategy without dimensionality reduction; the dual-channel recurrent layer extracts and fuses the row and column time series features of the input feature map through two parallel recurrent layers and performs a reshape operation to obtain the overall time series features of the feature map. The new feature map obtained by fusing the two features in a 1:1 ratio is used as the input of the next layer of feature fusion module to continue feature extraction.

[0037] The acquisition process of the photonic crystal spatial transmission spectrum data set provided in the present invention is shown in the attached Figure 6 , including randomly taking values ​​of structural parameters; normalizing structural parameters; collecting csv parameter labels; calculating the spatial transmission spectrum of layered photonic crystals using the propagation matrix method; and collecting mart spectrum samples.

[0038] In the above-mentioned specific implementation of the method for extracting the spatiotemporal characteristics of the spatial transmission spectrum of photonic crystals in parallel based on ECA-CNN fusion dual-channel RNN, further, the neural network model uses the Sigmoid nonlinear activation function during the training process, and its mathematical expression is:

[0039]

[0040] Among them, x represents each weight value of the flattened feature sequence. Through this formula, each weight value can be limited between 0 and 1. In addition, the adaptive algorithm ReduceLROnPlateau is used to dynamically adjust the learning rate. When the verification loss does not change within a certain iteration cycle, the learning rate will be decayed at a specified rate. The feature extraction and parameter prediction performance of the model are improved through a variety of optimization methods.

[0041] In the above-mentioned specific implementation of the method for extracting the spatiotemporal characteristics of the spatial transmission spectrum of photonic crystals in parallel based on ECA-CNN fusion dual-channel RNN, further, the multiple network model performance evaluation indicators, including mean square error, root mean square error, mean absolute error, mean relative error, mean absolute percentage error and symmetric mean absolute percentage error, evaluate the prediction performance of the model:

[0042]

[0043] Where n represents the number of samples, Represents the predicted value, y i Represents the actual value. The six evaluation indicators reflect the differences between the real data and the predicted data in different ways. The higher the model performance, the closer the evaluation indicator value is to 0.

[0044] This example compares the prediction performance of different neural network models through different evaluation indicators, as shown in Table 1.

[0045] MSE RMSE MAE MRE MAPE SMAPE ResNet 4.50e-06 2.10e-03 1.70e-03 2.93e-03 0.2829％ 0.2830％ CRNN 1.50e-05 3.90e-03 3.02e-03 5.01e-03 0.4906％ 0.4905％ Ours 1.6e-06 1.30e-03 8.90e-04 1.47e-03 0.1468％ 0.1468％

[0046] It can be seen from Table 1 that among all the neural network models, the prediction performance of the neural network model based on ECA-CNN fusion dual-channel RNN is the highest, and the comparison of different evaluation indicators is 1.6e-06, 1.30e-03, 8.90e-04, 1.47e-03, 0.1468%, and 0.1468%, respectively. This proves the effectiveness of the method of the present invention and also shows that different image features play a complementary role.

[0047] It should be noted that the above description is only an embodiment of the present invention and is only used to explain the present invention. It does not limit the patent scope of the present invention. Only obvious changes that belong to the technical concept of the present invention are also within the protection scope of the present invention.

Claims

1. A method for extracting the spatiotemporal characteristics of a photonic crystal spatial transmission spectrum, characterized in that: The following steps are involved: (1) obtaining spatial transmission spectrum samples of a one-dimensional photonic crystal and corresponding photonic crystal structural parameters as training data sets, and performing data preprocessing operations on the structural parameters; (2) Design a neural network model that combines ECA-CNN with dual-channel RNN; (3) Use the acquired data set to train the constructed neural network model and use optimization algorithms to improve the model feature extraction and parameter prediction performance; (4) evaluating the performance of the network model by calculating mean square error, root mean square error, mean absolute error, mean relative error, mean absolute percentage error and symmetric mean absolute percentage error indicators; In the step (2), the neural network model structure includes a feature fusion module and a fully connected layer, the feature fusion module is used to extract the spatial pixel features and image temporal features of the spatial transmission spectrum and perform fusion transmission; the fully connected layer is used to flatten the fusion features output by the last feature fusion module and map them into the structural parameters of the photonic crystal; The feature fusion module includes a convolution module and a dual-channel recurrent layer. In the convolution module, the ECA attention mechanism is connected after the convolution layer to enhance the convolution layer's ability to extract spatial pixel features with a local cross-channel interaction strategy without dimensionality reduction. The dual-channel recurrent layer extracts and fuses the row and column time series features of the input feature map through two parallel recurrent layers to obtain the overall time series features of the feature map. The new feature map obtained by fusing the two features in a 1:1 ratio is used as the input of the next layer feature fusion module to continue feature extraction. In the step (3), the neural network model uses the Sigmoid nonlinear activation function, whose mathematical expression is: Among them, x represents each weight value of the flattened feature sequence. This formula limits each weight value to between 0 and 1. In addition, the adaptive algorithm ReduceLROnPlateau is used to dynamically adjust the learning rate. When the verification loss does not change within a certain iteration cycle, the learning rate will be decayed at a specified rate. A variety of optimization methods are used to improve the feature extraction and parameter prediction performance of the model.

2. The method according to claim 1, characterized in that In the step (1), different types of photonic crystal spatial transmission spectra are collected as training samples, and the corresponding structural parameters are divided by their respective maximum values ​​for normalization operation to serve as label data. The universality of the network model is verified through multiple groups of experiments.

3. The method according to claim 1, characterized in that In the step (4), the prediction performance of the model is evaluated by using multiple evaluation indicators such as mean square error, root mean square error, mean absolute error, mean relative error, mean absolute percentage error and symmetric mean absolute percentage error: Where n represents the number of samples, Represents the predicted value, y i Represents the actual value. The six evaluation indicators reflect the differences between the real data and the predicted data in different ways. The higher the model performance, the closer the evaluation indicator value is to 0.

Images