Star light interferometer optical path difference detection method based on deep learning

Through the deep learning-based neural network model FD-ViT, the optical path difference is directly calculated from a single-frame interference map with high accuracy, which solves the problem that the optical path difference detection frequency and energy utilization rate cannot be taken into account in the stellar light interferometer, and achieves high efficiency , high-speed optical path difference detection improves the energy utilization rate and signal-to-noise ratio of the interferometer.

CN119089791APending Publication Date: 2024-12-06NANJING INST OF ASTRONOMICAL OPTICS & TECH NAT ASTRONOMICAL OBSE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411217880.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-02
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

In the prior art, the optical path difference detection of stellar interferometers is difficult to take into account high frequency, large dynamic range and high accuracy, and the energy utilization rate is low, and the traditional method has low detection frequency and poor real-time performance.

Method used

Using deep learning-based methods, the neural network model FD-ViT is constructed, the optical path difference is calculated through interference graphs, the channel spectral interference simulation model is established, the simulation data set is produced, and the neural network model is trained to achieve high-precision optical path difference detection and adaptive Output the final result.

Benefits of technology

It realizes efficient and high-speed optical path difference detection, improves the energy utilization rate and signal-to-noise ratio of the interferometer, and does not require time or space phase shift, ensuring the accuracy and frequency of optical path difference detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089791A_ABST
    Figure CN119089791A_ABST
Patent Text Reader

Abstract

The invention relates to a fixed star light interferometer optical path difference detection method based on deep learning. The method comprises the following steps: constructing a neural network model; the method comprises the following steps: establishing a channel spectrum interference simulation model based on an optical interference theory, setting simulation parameters according to an actual interference system, and making a simulation data set which comprises a simulation interference pattern and two normalized fringe periodicity; taking the simulation data set as training data of a neural network model, taking the simulation interferogram as the input of the model, taking the normalized fringe periodicity as the output of the model, and obtaining a mapping relation between the interference data and the normalized fringe periodicity; and comprehensively evaluating the number of two stripe cycles output by the neural network model, and adaptively outputting a final optical path difference detection result in combination with a detection wave band. According to the method, the optical path difference can be detected at high precision only by a single interferogram, the requirements of fringe tracking and phase tracking on optical path difference detection are met in a self-adaptive manner, and the method has the advantages of high efficiency, high precision, large dynamic range, high signal-to-noise ratio and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of optical precision measurement, astronomical instruments and methods, and in particular to a method for detecting optical path difference of a stellar light interferometer based on deep learning. Background Art

[0002] Stellar light interferometers use the principle of interference to achieve ultra-high-resolution astronomical imaging and astronomical measurement, and have played a significant role in the field of astronomical observation. However, during astronomical observations, the optical path of starlight from different telescopes will change rapidly due to factors such as atmospheric disturbances, the rotation of the earth, and telescope tracking errors, resulting in reduced interferometric measurement accuracy and reliability. Therefore, how to achieve high-frequency, large dynamic range, and high-precision optical path difference detection is the key to rapid capture and stable tracking of interference fringes.

[0003] In traditional stellar light interferometers, optical path difference detection is usually divided into two parts: fringe tracking and phase tracking. Fringe tracking is mainly used to detect and correct large optical path differences, and control the optical path difference to the wavelength level, so as to ensure that the system always forms stable interference fringes; phase tracking is used to perform precise phase (optical path difference) detection on stable interference fringes, and its result directly determines the accuracy of astronomical measurement, which usually needs to be controlled within 0.1λ (λ, detection wavelength). Therefore, stellar light interferometers put forward requirements for optical path difference detection with large dynamic range, high precision and high frequency. In addition, in the process of astronomical measurement, energy is limited, and traditional detection methods are difficult to take into account the above requirements. E. Wilson et al. adopted an optical path difference detection algorithm based on discrete Fourier transform (DFT) on IOTA. By scanning the white light interference signal, the DFT results of multiple sampling windows are calculated, and the interference fringe envelope is fitted to obtain the optical path difference. However, this method relies on optical path scanning, has low detection frequency and poor real-time performance. The VLTI interferometer achieves 250nm precision optical path difference detection for four 8m aperture telescopes through spatial phase shifting combined with Kalman filtering technology. It is currently the device with the highest optical path difference detection accuracy among stellar light interferometers. However, its spatial phase shifting process requires light splitting, which reduces the interference signal-to-noise ratio, and the detection optical path is complex, so it is only suitable for interferometers composed of large-aperture telescopes.

[0004] In recent years, with the rapid development of deep learning methods, they have been widely used in various fields. In the field of optical astronomy, deep learning algorithms have been successfully applied to image preprocessing, phase detection, super-resolution imaging, interferometric signal processing, etc., and have achieved many outstanding results. However, in the prior art, an optical path difference detection method that takes into account frequency, accuracy and energy utilization has not yet been realized. Therefore, it is necessary to provide an optical path difference detection method based on deep learning. Summary of the invention

[0005] In order to solve the problem in the prior art that the detection frequency and energy utilization cannot be taken into account at the same time, the present invention provides a stellar light interferometer optical path difference detection method based on deep learning.

[0006] The specific content of the present invention is as follows: A method for detecting optical path difference of a stellar light interferometer based on deep learning, comprising the following steps:

[0007] Step 1: Construct a neural network model FD-ViT, through which the number of fringe periods in an interference pattern can be calculated, and then the corresponding optical path difference can be solved;

[0008] Step 2: Based on the theory of optical interference, a channel spectral interference simulation model is established, simulation parameters are set according to the actual interference system, and a simulation data set is produced. The simulation data set includes a simulated interference pattern and two normalized fringe period numbers;

[0009] Step 3: Use the simulated data set as the training data of the neural network model FD-ViT, use the simulated interference pattern as the input of the model, and use the normalized fringe period number as the output of the model to obtain the mapping relationship between the interference data and the normalized fringe period number;

[0010] Step 4: Comprehensively evaluate the two fringe period numbers output by the neural network model FD-ViT, combine the detection band, and adaptively output the final optical path difference detection result.

[0011] Furthermore, in step 1, the structure of the constructed neural network model FD-ViT includes:

[0012] Input image and image block module: divide the input image into multiple fixed-size image blocks, flatten each image block into a one-dimensional vector, and then map the flattened image blocks to a feature space of fixed dimension D through a fully connected layer;

[0013] Position encoding module: adds position encoding to each image block to retain the position information of each block in the image;

[0014] Transformer encoder layer: The Transformer encoder layer is obtained by stacking multiple identical encoder layers. Each encoder layer contains the following 3 sublayers:

[0015] Multi-head self-attention mechanism MHSA: Multiple self-attention heads are used to parallely calculate the correlation between each image block embedding and other image block embeddings to obtain self-attention scores, thereby learning the relationship between image blocks globally.

[0016] Feedforward neural network FFN: It includes two fully connected layers and a nonlinear activation function to perform nonlinear transformation on the features of each image block;

[0017] Layer normalization and residual connection: Layer normalization is used before the multi-head self-attention mechanism MHSA and the feedforward neural network FFN to stabilize the training process, and the input and output are skipped in each sub-layer to ensure the backward propagation of shallow information and model convergence.

[0018] Multilayer perceptron classification head: A learnable class label is added before the image block sequence. After passing through the Transformer encoder layer, the class label is input into the multilayer perceptron, and the predicted value is output. The output value is then normalized to [-1, 1] through the tanh activation function to stabilize the output result and accelerate model convergence.

[0019] Furthermore, in step 2, making a simulation data set includes:

[0020] After two beams of starlight from the same star interfere, the intensity distribution of the interference field is expressed as:

[0021]

[0022] Where x and y are spatial position coordinates, I1(x, y) and I2(x, y) are the intensity distributions of the two coherent light beams, λ(x) is the distribution of wavelength along the dispersion direction, d is the optical path difference to be detected, φ is the phase constant, and σ(x, y) is the system noise;

[0023] I1(x,y) and I2(x,y) are expressed as normalized light intensity distribution:

[0024]

[0025] Among them, I0 is the average light intensity, γ is the ratio of the intensity of the two beams, ε1(x, y) and ε2(x, y) are the non-uniform noise of the two beams respectively;

[0026] The wavelength distribution λ(x) along the dispersion direction is expressed as:

[0027] λ0+Cx(3)

[0028] Where λ0 is the central detection wavelength and C is the dispersion coefficient associated with the dispersion element;

[0029] Combining equations (1)-(3), within the range of parameter values ​​under experimental conditions, random functions are used to generate I0, γ, ε1(x, y), ε2(x, y), d, φ and σ(x, y), and simulated interference patterns under different conditions are obtained;

[0030] In formula (1), the relationship between the optical path difference d to be measured and the detection wavelength satisfies:

[0031]

[0032] Among them, λmin , max are the minimum and maximum wavelengths of the detection band, Δλ is the detection bandwidth, m is the number of interference fringes, and the number of interference fringes cycles is obtained according to the optical path difference and the detection band;

[0033] The normalized number of interference fringe periods is used as the true value in the simulation data set.

[0034] Further, in step 4, the adaptive output of the final optical path difference detection result includes: the two fringe period numbers output by the neural network model FD-ViT are respectively the predicted value O1 that meets the fringe tracking detection requirements and the predicted value O2 that meets the phase tracking detection requirements. According to the size of O1 and O2, the final result is adaptively output. The specific steps include:

[0035] According to the detection requirements, set the stripe number segmentation threshold N. When the stripe number N f >N, fringe tracking is performed to control the optical path difference to within N fringes; when N f When ≤N, phase tracking is performed;

[0036] The maximum number of resolvable fringes in the interference pattern is M, and the criteria for selecting the values ​​of O1 and O2 are:

[0037]

[0038] The output value of the neural network model FD-ViT is judged. When O2=1, the number of stripes is calculated according to O1. When O1=-1, the number of stripes is calculated according to O2.

[0039] Substituting the number of fringes into equation (4), we can obtain the final optical path difference calculation result.

[0040] Furthermore, in the training process of step three, the mean square error is used as the loss function, and the adaptive moment estimation method is used to optimize the weights in the training process.

[0041] Beneficial effects of the present invention: The present invention realizes optical path difference detection based on deep learning algorithm. After training the model with data set, the optical path difference can be directly calculated from a single interference pattern, which has extremely high measurement efficiency. The present invention uses a neural network model to calculate the optical path difference from a single interference pattern with high precision, without time or space phase shifting, and can effectively improve the energy utilization rate and interference signal-to-noise ratio of the interferometer while ensuring the accuracy of optical path difference detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The specific implementation of the present invention will be further explained below in conjunction with the accompanying drawings.

[0043] Figure 1 This is a flow chart of the optical path difference detection method of a stellar light interferometer based on deep learning of the present invention;

[0044] Figure 2 is a network structure diagram of the present invention;

[0045] Figure 3 is a structural diagram of the Transformer encoder of the present invention;

[0046] Figure 4 is a portion of the input interferogram used for testing. DETAILED DESCRIPTION

[0047] Combination Figure 1-Figure 4 The present invention provides a method for detecting optical path difference of a stellar light interferometer based on deep learning, comprising the following steps:

[0048] Step 1: Construct a neural network model FD-ViT to calculate the number of fringe periods from an interference pattern and then solve the corresponding optical path difference. Figure 2 As shown, the FD-ViT structure consists of the following parts:

[0049] (1) Input image and image patching: The input image is divided into multiple image patches of fixed size, each image patch is flattened into a one-dimensional vector, and then the flattened image patch is mapped to a feature space of fixed dimension D (Linear Projection) through a fully connected layer;

[0050] (2) Positional Embedding: Add positional encoding to each image block to retain the position information of each block in the image.

[0051] (3) Transformer encoder layer: The Transformer encoder layer is obtained by stacking multiple identical encoder layers, such as Figure 3 As shown, each encoder layer contains the following 3 sub-layers:

[0052] (a) Multi-Head Self-Attention (MHSA): Multiple self-attention heads are used to parallely compute the correlation between each image patch embedding and other image patch embeddings to obtain a self-attention score, thereby learning the relationship between image patches globally.

[0053] (b) Feed-Forward Network (FFN): It consists of two fully connected layers and a nonlinear activation function, which performs nonlinear transformation on the features of each image block;

[0054] (c) Layer Normalization and Residual Connection: Use layer normalization before MHSA and FFN to stabilize the training process, and skip connections are made between input and output in each sub-layer to ensure that shallow information is propagated backward and the model converges;

[0055] (4) Multilayer Perceptron (MLP) classification head: A learnable class token is added before the image block sequence. After passing through the Transformer encoder layer, the class token is input into the MLP and two prediction values ​​are output. The prediction values ​​are then normalized to [-1, 1] through the tanh activation function to stabilize the output results and accelerate model convergence.

[0056] Step 2: Based on the theory of optical interference, a channel spectral interference simulation model is established. The simulation parameters are set according to the actual interference system, and a simulation data set with a distribution similar to the actual interference data is produced. The simulation data set includes a simulated interference pattern and two normalized fringe period numbers. The specific process is as follows:

[0057] When two beams of starlight from the same star interfere with each other, the intensity distribution of the interference field formed by the two beams of starlight can be expressed as

[0058]

[0059] Where x and y are the spatial position coordinates, I1(x, y) and I2(x, y) are the intensity distributions of the two coherent light beams, λ(x) is the wavelength distribution along the dispersion direction, d is the optical path difference to be detected, φ is the phase constant, and σ(x, y) is the system noise.

[0060] In a stellar light interferometer, the interference signal and the intensity distribution of the two coherent light beams are usually collected simultaneously and normalized. The normalized spot intensity is approximately uniformly distributed, so I1(x, y) and I2(x, y) can be expressed as

[0061]

[0062] Among them, I0 is the average light intensity, γ is the ratio of the intensity of the two beams, ε1(x, y) and ε2(x, y) are the non-uniform noise of the two beams.

[0063] The wavelength distribution λ(x) along the dispersion direction can be expressed as

[0064] λ(x)=λ0+Cx (3)

[0065] Where λ0 is the central detection wavelength and C is the dispersion coefficient associated with the dispersion element.

[0066] Combining equations (1)-(3), within the range of parameter values ​​under experimental conditions, random functions are used to generate I0, γ, ε1(x, y), ε2(x, y), d, φ and σ(x, y), and the simulated interference patterns under different conditions can be obtained.

[0067] In formula (1), the relationship between the optical path difference d to be measured and the detection wavelength satisfies

[0068]

[0069] Among them, λ min , max are the minimum and maximum wavelengths of the detection band, Δλ is the detection bandwidth, m is the number of interference fringes, is the equivalent optical path difference of each interference fringe, so the number of interference fringe periods can be obtained based on the optical path difference and the detection band.

[0070] In order to facilitate the training of the neural network model, the number of interference fringe cycles is normalized as the true value in the data set. The two values ​​are

[0071]

[0072] Among them, O1 and O2 are two normalized stripe numbers, N f is the actual number of fringes, M is the maximum number of fringes that can be distinguished in the interference pattern, when N f >N, the optical path difference is large, and fringe tracking is required to stabilize the interference fringes. At this time, O1 is used as the effective true value and O2 is used as the auxiliary value. f When ≤N, the optical path difference is small, the fringes are stable, and phase tracking is required for precise measurement. At this time, O2 is used as the effective true value and O1 is used as an auxiliary quantity.

[0073] Step 3: Use the simulation data set as the training data of the neural network model, the interference pattern as the input of the neural network model, and the normalized fringe period number as the output of the neural network model to obtain the mapping relationship between the interference data and the normalized fringe period number.

[0074] The data set used in this embodiment contains a total of 100,000 sets of data, which are allocated to the training set, validation set, and test set at a ratio of 0.8, 0.1, and 0.1, respectively. The training process uses the mean squared error (MSE) as the loss function, and uses the adaptive moment estimation method (Adam) to optimize the weights in the training process. The initial value of the learning rate lr is 3×10 -4, the rate decays at 0.1 after every 100 rounds of training, the network is trained for a total of 300 rounds, and the batch size is set to 512. After 300 rounds of training, the loss functions on the test set and the validation set converge synchronously.

[0075] Step 4: Based on the comprehensive evaluation of the model prediction value and the optical path difference resolution l, the final predicted optical path difference is obtained:

[0076]

[0077] The prediction accuracy of the model was verified on 10,000 sets of data in the test set, and some interference patterns and optical path difference detection results are shown in the figure. Figure 4 As shown in Table 1, the results show that the RMS value of the optical path difference accuracy required for fringe tracking is The RMS value of the fringe number detection accuracy required for phase tracking is

[0078] Table 1 Optical path difference detection results of some input interference patterns

[0079]

[0080] Step 5: Input the interference pattern collected by the interference system into the trained model and output the optical path difference.

[0081] In summary, the present invention does not require phase shifting and can directly calculate the optical path difference from a single interference pattern with high precision. It only takes 0.0001 seconds to calculate the optical path difference for each interference pattern on average, and has high measurement efficiency, measurement accuracy and energy utilization.

[0082] The present invention proposes a method for detecting optical path difference of stellar light interferometer based on deep learning, which realizes high-precision detection of optical path difference by using a single interference pattern, and solves the problem that detection frequency and energy utilization cannot be taken into account at the same time in traditional methods. On the premise of ensuring the accuracy and frequency of optical path detection, the overall energy utilization of the interferometer can be effectively improved, thereby improving its limit of star detection, and has broad application prospects in the fields of stellar light interferometry and telescope common phase.

[0083] Many specific details are described in the above description to facilitate a full understanding of the present invention. However, the above description is only a preferred embodiment of the present invention. The present invention can be implemented in many other ways different from those described herein, so the present invention is not limited to the specific implementation disclosed above. At the same time, any person familiar with the art can make many possible changes and modifications to the technical solution of the present invention using the methods and technical contents disclosed above without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A method for detecting optical path difference of a stellar light interferometer based on deep learning, characterized in that: The steps include: Step 1: Construct a neural network model FD-ViT, through which the number of fringe periods in an interference pattern can be calculated, and then the corresponding optical path difference can be solved; Step 2: Based on the theory of optical interference, a channel spectral interference simulation model is established, simulation parameters are set according to the actual interference system, and a simulation data set is produced. The simulation data set includes a simulated interference pattern and two normalized fringe period numbers; Step 3: Use the simulated data set as the training data of the neural network model FD-ViT, use the simulated interference pattern as the input of the model, and use the normalized fringe period number as the output of the model to obtain the mapping relationship between the interference data and the normalized fringe period number; Step 4: Comprehensively evaluate the two fringe period numbers output by the neural network model FD-ViT, combine the detection band, and adaptively output the final optical path difference detection result.

2. The method for detecting optical path difference of a stellar light interferometer based on deep learning according to claim 1, characterized in that: In step 1, the structure of the constructed neural network model FD-ViT includes: Input image and image block module: divide the input image into multiple fixed-size image blocks, flatten each image block into a one-dimensional vector, and then map the flattened image blocks to a feature space of fixed dimension D through a fully connected layer; Position encoding module: adds position encoding to each image block to retain the position information of each block in the image; Transformer encoder layer: The Transformer encoder layer is obtained by stacking multiple identical encoder layers. Each encoder layer contains the following 3 sublayers: Multi-head self-attention mechanism MHSA: Multiple self-attention heads are used to parallely calculate the correlation between each image block embedding and other image block embeddings to obtain self-attention scores, thereby learning the relationship between image blocks globally. Feedforward neural network FFN: It includes two fully connected layers and a nonlinear activation function to perform nonlinear transformation on the features of each image block; Layer normalization and residual connection: Use layer normalization before the multi-head self-attention mechanism MHSA and the feed-forward neural network FFN to stabilize the training process, and skip the input and output in each sub-layer to ensure the backward propagation of shallow information and model convergence; Multilayer perceptron classification head: A learnable class label is added before the image block sequence. After passing through the Transformer encoder layer, the class label is input into the multilayer perceptron, and the predicted value is output. The output value is then normalized to [-1, 1] through the tanh activation function to stabilize the output result and accelerate model convergence.

3. The method for detecting optical path difference of a stellar light interferometer based on deep learning according to claim 1, characterized in that: In step 2, making a simulation data set includes: After two beams of starlight from the same star interfere, the intensity distribution of the interference field is expressed as: Where x and y are spatial position coordinates, I1(x, y) and I2(x, y) are the intensity distributions of the two coherent light beams, λ(x) is the distribution of wavelength along the dispersion direction, d is the optical path difference to be detected, φ is the phase constant, and σ(x, y) is the system noise; I1(x,y) and I2(x,y) are expressed as normalized light intensity distribution: Among them, I0 is the average light intensity, γ is the ratio of the intensity of the two beams, ε1(x, y) and ε2(x, y) are the non-uniform noise of the two beams respectively; The wavelength distribution λ(x) along the dispersion direction is expressed as: λ(x)=λ0+Cx (3)where λ0 is the central detection wavelength, and C is the dispersion coefficient associated with the dispersion element; Combining equations (1)-(3), within the range of parameter values ​​under experimental conditions, random functions are used to generate I0, γ, ε1(x, y), ε2(x, y), d, φ and σ(x, y), and simulated interference patterns under different conditions are obtained; In formula (1), the relationship between the optical path difference d to be measured and the detection wavelength satisfies: Among them, λ min , max are the minimum and maximum wavelengths of the detection band, Δλ is the detection bandwidth, m is the number of interference fringes, and the number of interference fringes cycles is obtained according to the optical path difference and the detection band; The normalized number of interference fringe periods is used as the true value in the simulation data set.

4. The method for detecting optical path difference of a stellar light interferometer based on deep learning according to claim 3, characterized in that: In step 4, the final optical path difference detection result is adaptively outputted, including: the two fringe period numbers outputted by the neural network model FD-ViT are respectively the predicted value O1 that meets the fringe tracking detection requirements and the predicted value O2 that meets the phase tracking detection requirements. According to the size of O1 and O2, the final result is adaptively outputted. The specific steps include: According to the detection requirements, set the stripe number segmentation threshold N. When the stripe number N f >N, fringe tracking is performed to control the optical path difference to within N fringes; when N f When ≤N, phase tracking is performed; The maximum number of resolvable fringes in the interference pattern is M, and the criteria for selecting the values ​​of O1 and O2 are: The output value of the neural network model FD-ViT is judged. When O2=1, the number of stripes is calculated according to O1. When O1=-1, the number of stripes is calculated according to O2. Substituting the number of fringes into equation (4), we can obtain the final optical path difference calculation result.

5. The method for detecting optical path difference of a stellar light interferometer based on deep learning according to claim 1, characterized in that: In the training process of step 3, the mean square error is used as the loss function, and the adaptive moment estimation method is used to optimize the weights during the training process.