Multi-mode optical fiber gyroscope based on deep learning

Abstract

The invention relates to a multi-mode optical fiber gyroscope based on deep learning. The multi-mode optical fiber gyroscope comprises a laser light source, two beam splitters, an optical fiber loop, a CCD camera, a computer, a multi-mode optical fiber and a signal line, the laser light source is connected with the first beam splitter through multimode optical fibers, the first beam splitter is connected with the second beam splitter through multimode optical fibers, the first beam splitter is connected with the CCD camera through multimode optical fibers, the CCD camera is connected with the computer through signal lines, and a speckle recognition model based on a variational auto-encoder is arranged in the computer. Potential features in a speckle image can be automatically learned and extracted, the speckle image is mapped to probability distribution parameters in a potential space by using an encoder, and an image with features similar to those of an original image is reconstructed from the parameters through a decoder. Through the process, complex image features can be effectively captured and represented, and then the rotation angular velocity information contained in the speckle image is accurately solved by using the potential variables, so that high-precision rotation speed measurement is realized.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and fiber optic sensing, and particularly to a multimode fiber optic gyroscope based on deep learning Background Art

[0002] Multimode fibers allow multiple optical modes to propagate in parallel, and these modes carry different phase information, thus greatly increasing the information capacity. However, as a scattering medium, each optical mode inside a multimode fiber propagates at a different phase velocity. When the light beam in the fiber reaches the output end, due to the phase difference between the modes, beam interference generates a complex speckle image. The speckle image is a high-dimensional and irregular image data, which contains a large amount of phase and intensity information. However, due to the complexity and disorder of these images, it is difficult to effectively and accurately extract and solve the useful information carried in the speckle image using traditional signal processing algorithms such as Fourier transform or wavelet analysis

[0003] In recent years, as an important branch of machine learning, deep learning has demonstrated powerful capabilities in processing complex high-dimensional data. The variational autoencoder is a deep learning generative model, which is particularly suitable for feature extraction and modeling tasks of high-dimensional data. Compared with traditional convolutional neural networks, the variational autoencoder can efficiently characterize the distribution characteristics of input data in the latent space, while extracting key features and retaining their statistical laws. This ability is particularly prominent when dealing with complex speckle images, and can effectively capture the significant patterns and changing features in the input image

[0004] Based on the deficiencies of the above existing data processing methods and the characteristics of the variational autoencoder in processing data, it is intended to design a multimode fiber optic gyroscope based on deep learning to solve the deficiencies of the existing technology Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention proposes a multimode fiber optic gyroscope based on deep learning that can improve the accuracy of speckle image processing

[0006] The above object of the present invention is achieved by the following technical solutions

[0007] A multimode fiber optic gyroscope based on deep learning includes a laser light source, a first beam splitter, a second beam splitter, a fiber optic loop, a CCD camera, a computer, a multimode fiber, and a signal line; among them, multimode fibers are used to connect between the laser light source and the first beam splitter, between the first beam splitter and the second beam splitter, and between the first beam splitter and the CCD camera, and a signal line is used to connect between the CCD camera and the computer, and a speckle recognition model based on a variational autoencoder is built in the computer

[0008] The laser light source is used to emit laser light

[0009] The first beam splitter and the second beam splitter are used to split the emitted laser beams and transmit them to the optical fiber ring, and to combine the two beams output by the optical fiber ring and output them to the CCD camera;

[0010] The CCD camera is used to capture tiny changes in light intensity and convert the speckle image into electrical signals and output them to the computer;

[0011] The computer calculates the rotational angular velocity information contained in the speckle image through a speckle recognition model.

[0012] Moreover, the speckle recognition model consists of an encoder, a latent space, a decoder and a speed prediction module;

[0013] The encoder is used to extract features from the input two-dimensional speckle image and compress it into potential distribution parameters, wherein the potential distribution parameters include a mean μ and a logarithmic variance logσ 2 ;

[0014] The latent space, the mean μ and log variance logσ generated by the encoder 2 Representation, from which latent variables z are sampled through reparameterization techniques to represent the low-dimensional feature distribution of the input data;

[0015] The decoder is used to decode the latent variable z into a two-dimensional speckle image of the same size as the input image to achieve reconstruction of the input image;

[0016] The rotation speed prediction module is used to extract information related to the rotation speed of the fiber optic gyroscope from the potential variable z and directly output a scalar value.

[0017] Moreover, the encoder consists of four groups of convolutional layers and pooling layers and two fully connected layers; each convolutional layer uses a 4×4 convolution kernel, the input channel is 1, the output channel increases from 32 to 256 layer by layer, and the activation function is the relu function; the convolutional layer output generates the parameters of the potential distribution through two fully connected layers: mean z and logarithmic variance logσ 2 .

[0018] Moreover, the sampling latent variable z process adopts the reparameterization method, which is expressed as follows:

[0019] z=μ+∈·exp(0.5·logσ 2 )

[0020] in, is noise sampled from a standard normal distribution, and the dimension of the latent variable z is 64.

[0021] Moreover, the decoder consists of a fully-connected layer and four groups of deconvolution layers and upsampling layers; the fully-connected layer converts the latent variable z into a convolutional feature map, and then, an image with the same size as the input is generated through upsampling layers with gradually increasing resolution.

[0022] Moreover, the rotational speed prediction module consists of two layers of fully-connected networks, and the relu activation function is used after each layer of the network to introduce non-linearity.

[0023] Moreover, the working wavelength of the laser light source is 532 nm.

[0024] Moreover, the size of the speckle image collected by the CCD is 612*612.

[0025] The advantages and positive effects of the present invention are as follows:

[0026] 1. Application of multimode fiber: The present invention uses multimode fiber to wind the sensitive loop. Multimode fiber allows multiple optical modes to be transmitted in parallel, and each mode carries different phase information. This design greatly increases the phase information capacity of the system, so that compared with the traditional design using single-mode fiber, a shorter fiber length can be used under the same accuracy requirements. Therefore, the application of multimode fiber not only improves the sensitivity of the system, but also significantly reduces the volume and weight of the fiber optic gyroscope, which is beneficial to the miniaturization and integration of the system.

[0027] 2. Introduction of variational autoencoder: The present invention uses a speckle recognition model based on variational autoencoder to calculate the rotational angular velocity information in the speckle image. Different from traditional signal processing algorithms, variational autoencoder can automatically learn the complex feature distribution in the speckle image and generate a continuous and smooth representation in the latent space. Variational autoencoder not only improves the generalization ability of the system, but also has higher generation ability and can effectively extract useful information from the latent space. Through this feature, the present invention can more accurately analyze the speckle image generated by multimode interference and significantly improve the calculation accuracy of the rotational angular velocity.

[0028] 3. Robustness and stability: Since the variational autoencoder model can capture the latent feature distribution in the speckle image, the present invention can maintain high robustness and stability in the face of external environmental changes (such as temperature fluctuations, mechanical vibrations, etc.). This advantage enables the multimode fiber optic gyroscope of the present invention to still provide stable and reliable measurement results in complex and changing working environments.

[0029] In summary, through the innovative combination of multimode fiber and deep learning technology, especially variational autoencoder, the present invention realizes higher-precision rotational speed measurement and a more compact system design, providing a new approach for the development of fiber optic gyro technology. Description of the Drawings

[0030] Figure 1 is the structural diagram of the multimode fiber optic gyroscope based on deep learning of the present invention;

[0031] Figure 2 is the structural diagram of the speckle recognition model of the present invention; Specific Embodiments

[0032] The structure of the present invention will be further described below in conjunction with the accompanying drawings and through embodiments. It should be noted that this embodiment is narrative rather than restrictive.

[0033] The present invention proposes a multimode fiber optic gyroscope based on deep learning. By using a speckle recognition model based on variational autoencoders, it can automatically learn and extract latent features in speckle images. The model uses an encoder to map the speckle image to the probability distribution parameters (mean and variance) in the latent space, and then a decoder reconstructs an image with features similar to the original image from these parameters. Through this process, the variational autoencoder can effectively capture and represent complex image features, and then use the latent variables to accurately calculate the rotational angular velocity information contained in the speckle image, thereby achieving high-precision rotational speed measurement.

[0034] Please refer to Figure 1 , a multimode fiber optic gyroscope based on deep learning,: including a laser light source 1, a first beam splitter 2, a second beam splitter 3, an optical fiber loop 4, a CCD camera 5, a computer 6, a multimode optical fiber 7, and a signal line 8. This system processes speckle images based on a speckle recognition model. Among them, multimode optical fibers are used to connect between the laser light source and the first beam splitter, between the first beam splitter and the second beam splitter, and between the first beam splitter and the CCD camera. The computer is built-in with a speckle recognition model based on variational autoencoders.

[0035] The working wavelength of the laser light source is 532 nm. This wavelength selection is because of its low-loss characteristics in optical fiber transmission and high detection sensitivity in common optoelectronic devices.

[0036] The optical fiber loop is wound by multimode optical fiber. The multimode optical fiber can support the parallel transmission of multiple optical modes, thereby enhancing the phase information capacity of the system. When the fiber optic gyro rotates, two beams of light transmitted clockwise and counterclockwise in the loop will generate a Sagnac phase difference, and the magnitude of the phase difference is proportional to the rotational speed.

[0037] The working process of the above device is as follows: The laser source emits laser light. After the light beam passes through the first beam splitter, it is split into two parts at the second beam splitter. These two parts of the light beam enter the fiber optic loop respectively and propagate in opposite directions (i.e., clockwise and counterclockwise). Inside the loop, each optical mode in the two beams of light will generate different phases when reaching the other end of the loop due to different propagation constants. After leaving the fiber optic loop, the two beams of light are combined at the second beam splitter and output through the input port of the second beam splitter. This port is a reciprocal port. This reciprocal optical path structure where the input and output share a port can make the inherent phase shifts experienced by the two beams of light propagating in opposite directions equal when splitting and combining at the second beam splitter. Then the output light is separated from the input light by the first beam splitter and finally reaches the CCD camera. The high sensitivity of the CCD camera allows it to capture tiny light intensity changes and convert the speckle image into an electrical signal, where the size of the speckle image collected by the CCD is 612*612. Finally, the computer uses the speckle recognition model to solve the rotational angular velocity information contained in the speckle image.

[0038] As Figure 2 shown, the speckle recognition model is a model based on a variational autoencoder. It consists of an encoder 9, a latent space 10, a decoder 11, and a rotational speed prediction module 12.

[0039] The encoder is responsible for extracting features from the input two-dimensional speckle image and compressing them into latent distribution parameters (mean μ and log variance logσ 2 ). The encoder consists of four groups of convolutional layers and pooling layers, as well as two fully connected layers. Each convolutional layer uses a 4×4 convolutional kernel, the input channel is 1 (grayscale image), and the output channels increase layer by layer from 32 to 256. The activation function is the relu function. The output of the convolutional layer generates the parameters of the latent distribution: mean μ and log variance logσ 2 .

[0040] The latent space is used to sample the latent variable z from the mean μ and log variance logσ 2 generated by the encoder. The sampling process uses the reparameterization method to ensure differentiability:

[0041] z = μ + ∈·exp(0.5·logσ 2 )

[0042] where is the noise sampled from the standard normal distribution. The dimension of the latent variable z is 64, which is a feature representation in the low-dimensional space. It not only retains the key information of the input image but also has compactness, facilitating the processing of subsequent modules.

[0043] The decoder decodes the latent variable z into a two-dimensional speckle image with the same size as the input, realizing the reconstruction of the input image. The decoder consists of a fully connected layer and four groups of transposed convolution layers and upsampling layers. The fully connected layer converts the latent variable z into a convolutional feature map. Subsequently, an image with the same size as the input is generated through upsampling layers that gradually increase the resolution.

[0044] The rotational speed prediction module uses two layers of fully connected layers to output the latent variable z with a size of 64 as rotational speed information.

[0045] During the model training process, the variational autoencoder uses the reconstruction loss and the KL divergence loss to optimize the model parameters. The reconstruction loss uses the mean square error to measure the pixel difference between the original image and the reconstructed image, while the KL divergence loss is used to measure the difference between the latent distribution generated by the encoder and the standard normal distribution. In addition, the difference between the output of the rotational speed prediction module and the true rotational speed is measured using the mean square error loss. Generally speaking, the loss function during the training of the speckle recognition model includes:

[0046] L total =L reconstruction +βL KL +λL speed

[0047] where L reconstruction is the reconstruction loss, L KL is the KL divergence loss, and L speed is the mean square error loss of rotational speed prediction. The parameters β and λ respectively adjust the weights of reconstruction and KL divergence, and rotational speed prediction.

[0048] In summary, the system of the present invention systematically combines optical sensing technology and advanced machine learning algorithms for precise measurement of the rotational angular velocity. Specifically, the present invention applies a variational autoencoder (VAE) to process the complex speckle images generated by multimode optical fibers, which can effectively extract the latent features in the speckle images and use these features to accurately calculate the rotational angular velocity of the fiber optic gyroscope. The present invention not only improves the accuracy of speckle image processing, but also enhances the robustness and adaptability of the system, and can maintain high performance in complex environments.

[0049] The technology of the present invention has important application values in the fields of navigation, attitude control, seismic exploration, etc., especially in occasions that require high-precision angular velocity measurement, such as aerospace, ship navigation, and high-precision instrument manufacturing.

[0050] Through this deep learning method based on variational autoencoder, the present invention can effectively extract rotational speed-related information from complex speckle images, improving the calculation accuracy and robustness of the fiber optic gyroscope.

[0051] Although embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art can understand that various substitutions, changes, and modifications are possible without departing from the spirit of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the content disclosed in the embodiments and drawings.

Claims

1. A multimode fiber optic gyroscope based on deep learning, characterized in that: It includes a laser light source, a first beam splitter, a second beam splitter, an optical fiber ring, a CCD camera, a computer, a multimode optical fiber, and a signal line; wherein the laser light source and the first beam splitter, the first beam splitter and the second beam splitter, and the first beam splitter and the CCD camera are all connected by multimode optical fiber, the CCD camera and the computer are connected by a signal line, and the computer is built with a speckle recognition model based on a variational autoencoder; The laser light source is used to emit laser light; The first beam splitter and the second beam splitter are used to split the emitted laser beams and transmit them to the optical fiber ring, and to combine the two beams output by the optical fiber ring and output them to the CCD camera; The CCD camera is used to capture tiny changes in light intensity and convert the speckle image into electrical signals and output them to the computer; The computer calculates the rotational angular velocity information contained in the speckle image through a speckle recognition model.

2. The multimode fiber optic gyroscope based on deep learning according to claim 1, characterized in that: The speckle recognition model consists of an encoder, a latent space, a decoder and a speed prediction module; The encoder is used to extract features from the input two-dimensional speckle image and compress it into potential distribution parameters, wherein the potential distribution parameters include a mean μ and a logarithmic variance logσ 2 ; The latent space, the mean μ and log variance logσ generated by the encoder 2 Representation, from which latent variables z are sampled through reparameterization techniques to represent the low-dimensional feature distribution of the input data; The decoder is used to decode the latent variable z into a two-dimensional speckle image of the same size as the input image to achieve reconstruction of the input image; The rotation speed prediction module is used to extract information related to the rotation speed of the fiber optic gyroscope from the potential variable z and directly output a scalar value.

3. The multimode fiber optic gyroscope based on deep learning according to claim 2, characterized in that: The encoder consists of four groups of convolutional layers and pooling layers and two fully connected layers; each convolutional layer uses a 4×4 convolution kernel, the input channel is 1, the output channel increases from 32 to 256 layer by layer, and the activation function is the relu function; the convolutional layer output generates the parameters of the potential distribution through two fully connected layers: mean μ and logarithmic variance logσ 2 .

4. The multimode fiber optic gyroscope based on deep learning according to claim 2, characterized in that: The sampling latent variable z process adopts the reparameterization method, and the expression is as follows: z=μ+∈·exp(0.5·logσ 2 ) in, is noise sampled from a standard normal distribution, and the dimension of the latent variable z is 64.

5. The multimode fiber optic gyroscope based on deep learning according to claim 2, characterized in that: The decoder consists of a fully connected layer and four groups of deconvolutional layers and upsampling layers; the fully connected layer converts the latent variable z into a convolutional feature map, and then generates an image of the same size as the input through upsampling layers with gradually increasing resolution.

6. The multimode fiber optic gyroscope based on deep learning according to claim 2, characterized in that: The speed prediction module consists of two layers of fully connected networks, and the relu activation function is used after each layer of the network to introduce nonlinearity.

7. The multimode fiber optic gyroscope based on deep learning according to claim 1, characterized in that: The operating wavelength of the laser light source is 532 nm.

8. The multimode fiber optic gyroscope based on deep learning according to claim 1, characterized in that: The size of the speckle image collected by the CCD is 612*612.

Images