A Non-destructive Characterization Method for Hollow-core Anti-resonant Optical Fiber Based on Multimodal Deep Learning

Abstract

This invention provides a non-destructive characterization method for hollow anti-resonant optical fibers based on multimodal deep learning. The specific implementation path includes: (1) acquiring microscopic images (image modalities) of the end face of hollow anti-resonant optical fibers covering differentiated microstructures and their corresponding original transmission spectral data to construct a multimodal original sample set; (2) performing feature extraction and standardization processing on the multimodal data; (3) building a cascaded deep learning model integrating a visual feature capture subnetwork and a transmission spectral feature fusion regression subnetwork; (4) implementing joint training, using a convolutional neural network (CNN) to automatically extract the geometric feature parameters of the fiber core and cladding microstructures in the microscopic images, and fusing them with the transmission spectrum features, constructing a mapping relationship through a deep neural network (DNN); (5) inputting the microscopic image and transmission spectral data of the optical fiber under test into the model to achieve non-destructive prediction of loss characteristics. This invention features high characterization accuracy, fast speed, and zero material loss.

Description

Technical Field

[0001] This invention belongs to the field of advanced optics, optical communication and artificial intelligence, and in particular relates to a non-destructive characterization method for loss of hollow anti-resonant optical fiber based on a multimodal deep learning architecture. Background Technology

[0002] Hollow-core fiber confines most of the light energy within an air core for transmission, fundamentally overcoming the intrinsic defects of traditional solid-core silica fiber. It possesses excellent characteristics such as low delay, low dispersion, low nonlinearity, and a high laser damage threshold. Among them, hollow-core anti-resonant fibers (HC-ARFs), with their significant advantages of simple structure, low transmission loss, large optical bandwidth, and flexible tunable modes, have become a cutting-edge research focus in fields such as next-generation high-capacity optical fiber communication, ultra-high-power laser flexible transmission, and fiber optic sensing. The light guiding mechanism of hollow-core anti-resonant fibers is mainly based on the anti-resonant reflection waveguide theory. Its loss characteristics are highly correlated with the structural parameters of the fiber microstructure, particularly exhibiting extremely high sensitivity to core diameter, number of cladding capillaries, number of cladding capillary layers, cladding capillary outer diameter, cladding capillary wall thickness, and cladding capillary gap.

[0003] Transmission loss is one of the most crucial and fundamental indicators for evaluating the performance of hollow-core antiresonant optical fibers. Currently, the most common and classic method for characterizing fiber loss in the field of optical testing is the destructive truncation method. This physical testing method calculates the average loss across the entire wavelength range by measuring the difference in output spectral power of the same fiber at a longer length and at a shorter length after truncation. However, for extremely low-loss hollow-core antiresonant fibers, the traditional truncation method has revealed significant limitations in practical applications and industrialization: First, due to the limitations of current fiber drawing processes and fabrication technologies, the single-fiber length of high-performance hollow-core antiresonant fibers drawn in a single operation is usually short, and their intrinsic loss is extremely low. This directly leads to random errors introduced during the testing process due to spatial optical coupling instability, which can easily mask the true loss value, resulting in inaccurate test results. If this traditional method is to be used to achieve high-precision, wide-band testing of hollow-core anti-resonant optical fibers, the operational skills required of the testers are extremely demanding, and multiple repeated destructive tests must be performed to barely ensure the reliability of the test results. Secondly, the research and development and manufacturing costs of high-quality hollow-core anti-resonant optical fibers are extremely high, and the truncation method will inevitably cause irreversible physical length damage to the optical fiber, resulting in a large waste of materials.

[0004] The loss characteristics of hollow antiresonant optical fibers exhibit complex nonlinear physical relationships with their microstructure geometry and transmission spectrum. For example, the wall thickness of the cladding tube directly determines the center wavelength position of the antiresonant low-loss window. Even slight inhomogeneities or asymmetric deformations in the wall thickness of the cladding capillaries can significantly compress the effective bandwidth of actual transmission and lead to a surge in loss. However, this multidimensional matching relationship between microstructure and loss characteristics involves extremely complex nonlinear physical processes such as wave optics and resonant coupling, making real-time, accurate calculations impossible using simple analytical mathematical formulas or traditional empirical models. In recent years, with the rapid development of computer vision and heterogeneous data fusion processing technologies in the field of multimodal deep learning, utilizing spatially perceptive convolutional neural networks to automatically extract high-dimensional geometric parameters from digital images and combining them with deep data regression networks to process spectral signals has become a cutting-edge technology. By systematically collecting a large number of physically representative optical fiber end-face microscopic images, transmission transmission spectra, and loss data obtained by the corresponding truncation method, and conducting joint deep learning of multi-mode data, a nonlinear prediction model is established that directly maps "heterogeneous multi-mode information (image mode + data mode)" end-to-end to "full-band loss characteristics". This is a key technical approach to achieve lossless characterization of hollow-core anti-resonant optical fibers.

[0005] In summary, developing a non-destructive evaluation method that requires no manual intervention and does not damage the optical fiber itself—which, by simply inputting a high-magnification microscopic image of the fiber end face and corresponding transmission spectral information—can automatically capture structural features and perform multi-modal nonlinear fusion calculations based on intelligent algorithms, thereby accurately predicting loss spectral lines, has significant practical value for reducing the testing cost of hollow-core anti-resonant fibers, achieving non-destructive screening, and improving the automation and intelligence of testing processes. Summary of the Invention

[0006] The technical problem that this invention aims to solve is to overcome the shortcomings of existing technologies, such as permanent damage to optical fibers caused by truncation testing, large measurement errors due to limited fabrication length, and low efficiency and subjective bias in manually extracting microstructure parameters. This invention proposes a non-destructive characterization method for loss of hollow anti-resonant optical fibers based on multimodal deep learning.

[0007] The core innovation of this invention lies in the introduction of a cascaded network architecture that combines computer vision capture with deep feature fusion of multi-source data. This architecture enables the automatic extraction of comprehensive microstructure parameters from two-dimensional images of the fiber end face and their collaborative modeling with corresponding transmission spectral data. This intelligent architecture can quickly and accurately predict the full-band continuous loss spectrum of a given hollow-core antiresonant fiber, greatly improving the automation and economic efficiency of testing.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is detailed below:

[0009] This invention provides a non-destructive characterization method for hollow anti-resonant optical fibers based on multimodal deep learning, specifically encompassing the following steps:

[0010] (1) Construct a cascaded multimodal deep evaluation model:

[0011] A cascaded deep learning model is established, comprising a visual feature extraction subnetwork (CNN) and a multi-source data fusion regression subnetwork (DNN). The model uses a microscopic image of the cross-section of an optical fiber as the first input mode (spatial visual information), the corresponding transmission spectral data as the second input mode (wavelength-intensity data information), and the numerical sequence of the full-band loss characteristic spectrum of the optical fiber as the output target.

[0012] (2) Obtain a multimodal heterogeneous learning sample set:

[0013] Hollow-core antiresonant fiber samples with different structural parameters were collected. High-resolution microscopic images of their end faces were acquired using high-definition electronic imaging equipment, and the transmission spectrum of the fibers was obtained using a spectrometer. The loss spectrum data was then calibrated using the truncation method. These three types of data (microscopic images, transmission spectra, and loss spectra) are strictly aligned and together constitute a three-in-one multimodal supervised learning sample dataset that drives the evolution of the model.

[0014] (3) Joint training and parameter optimization of multimodal cascaded networks:

[0015] An error backpropagation optimization algorithm is used to optimize the cascaded model architecture end-to-end. First, a visual feature capture subnetwork is trained to automatically quantize and extract key geometric features such as core diameter, number of cladding layers, and wall thickness. Then, these feature parameters are concatenated with the transmission spectrum sequence of the second mode and fed into a deep regression hidden layer for multivariate nonlinear coupling fitting to complete the iterative update of the weight matrix.

[0016] (4) Perform lossless prediction:

[0017] High-resolution images and transmission energy spectrum data of any hollow-core anti-resonant fiber with unknown loss can be directly input into the system. Based on the forward operation of the model, the system directly outputs the predicted loss spectrum results for the fiber under test.

[0018] As a further preferred embodiment of the present invention, the visual feature capture subnetwork constructed in step (1) adopts a structure with alternating layers of deep convolutional layers and pooling layers, which has the ability to perform feature generalization extraction on optical fiber end face images under different magnification.

[0019] As a further preferred embodiment of the present invention, the multi-source data fusion regression subnetwork in step (1) uses a complex fully connected structure to strictly align the visual spatial features with the frequency response features of the spectrum in terms of data dimensions. In this process, the hidden layer of the regression subnetwork specifically uses GELU (Gaussian error linear unit) as the activation function to effectively suppress the gradient vanishing phenomenon in the propagation of deep networks, thereby simulating the complex loss mechanism of hollow anti-resonant optical fiber.

[0020] As a further preferred embodiment of the present invention, when acquiring training sample data in step (2), the database must include data on hollow anti-resonant optical fibers under different structural types (such as single-ring capillary structure, multi-layer nested capillary structure, 5-tube type, 7-tube type, etc.) and data on optical fiber microstructure deformation (such as adhesion between capillary walls, asymmetric distribution of capillary) to enhance the robustness of the model.

[0021] As a further preferred embodiment of the present invention, during the truncation method test described in step (2), it is necessary to ensure the absolute stability of the fiber coupling, use a mode stripper to filter out high-order mode interference, and obtain accurate fundamental mode loss data for algorithm training.

[0022] As a further preferred embodiment of the present invention, in step (3), when optimizing network weights, the traditional static learning strategy is no longer relied upon. Instead, the Adam optimizer with momentum term and adaptive learning rate dynamic adjustment strategy is introduced. The L2 regularization constraint term is explicitly introduced in the loss function. The purpose is to greatly accelerate the convergence speed of the model in the high-dimensional nonlinear space, completely avoid getting trapped in local optima, and effectively prevent the model from overfitting to a specific single fiber sample.

[0023] As a further preferred embodiment of the present invention, in order to preserve possible negative feedback edge features in the input image, the convolutional module of the visual feature capture subnetwork uses Leaky ReLU (Leaky Modified Linear Unit) as the local activation function in step (3), which effectively ensures the gradient survival rate of the fiber microtube wall defect features during the back propagation process.

[0024] As a further preferred embodiment of the present invention, in step (3), in response to the characteristic of a sharp increase in loss often occurring near a specific resonance band in hollow-core anti-resonant optical fibers, the present invention creatively sets an integrated dynamic weight matrix based on the change in loss gradient in the global error function. This matrix automatically assigns a larger error penalty weight factor to the high-loss abrupt change region, forcibly guiding the model to actively learn and overcome the representation difficulties on these physical boundaries.

[0025] As a further preferred embodiment of the present invention, the resolution of the predicted continuous loss spectrum output by the intelligent system in step (4) can be dynamically adjusted according to the acquisition accuracy configuration of the spectrometer. It supports both continuous curve fitting prediction of the ultrawide full band (e.g., from 400 nm in the ultraviolet to 2000 nm in the mid-infrared) and high-precision point-to-point numerical extraction for specific industrial laser working wavelengths (e.g., 1064 nm, 1550 nm, etc.).

[0026] As a further preferred embodiment of the present invention, this method also incorporates an automatic reverse evaluation step for the microstructure parameters of the optical fiber within its logical architecture. That is, while outputting the loss spectrum, it simultaneously outputs the quantitative evaluation index values ​​of the average cladding wall thickness, structural ellipticity, and overall symmetry of the optical fiber, extracted through a visual network, providing intuitive data support for process improvement. The significant beneficial effects of this invention are summarized as follows:

[0027] (1) Complete non-destructive testing process: This invention replaces the traditional physical method of "truncation measurement" with "image-based combined with spectral algorithm prediction". It completely solves the problems of characterization difficulties and test result distortion caused by the short single-processing length of high-performance, low-loss hollow-core anti-resonant optical fibers, saving optical fiber material costs for research institutions and enterprises.

[0028] (2) Extremely high characterization efficiency and objectivity: The complex microstructure parameters are automatically extracted and quantified by the convolutional neural network (CNN), which replaces the previous inefficient and error-prone manual microscopic measurement steps, and can significantly shorten the testing and verification cycle of new optical fibers.

[0029] (3) Excellent multimodal nonlinear fitting capability: This invention abandons the traditional single-input shallow network and innovatively constructs a deep network architecture based on multimodal data cascading and fusion. This architecture deeply explores the intrinsic relationship between the cladding inhomogeneity and loss of hollow antiresonant optical fibers. It has extremely high identification accuracy for loss fluctuations caused by complex wall thickness fluctuations and tube gap changes, demonstrating high physical feature identification accuracy and wide applicability. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the core technical route of a non-destructive characterization method for hollow anti-resonant optical fiber based on multimodal deep learning proposed in this invention.

[0031] Figure 2 This is a schematic diagram of the end face geometry of the first typical optical fiber structure (nested tube hollow anti-resonant optical fiber) involved in the embodiments of the present invention.

[0032] Figure 3The cascaded network architecture diagram with dual-modal input channels (covering a visual feature capture convolutional network and a multi-source data fusion deep network) is specially designed and adopted for this invention.

[0033] Figure 4 This is a schematic diagram of the end face geometry of the second typical optical fiber structure (single-ring hollow anti-resonant optical fiber) involved in the embodiments of the present invention. Detailed Implementation

[0034] To make the technical problems to be solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings. It should be clearly understood that the specific embodiments described in detail herein are only used to explain the engineering feasibility and internal logic of the multimodal deep learning technical solution of this invention, and do not constitute any limitation on the scope of protection of the claims of this invention.

[0035] See Figure 1 This paper presents a flowchart illustrating the overall technical roadmap of a unique, non-destructive characterization method for hollow-core antiresonant fiber loss based on multimodal deep learning, a method developed in this invention. The core operational steps of the intelligent characterization method are divided into highly coordinated stages S1 to S4:

[0036] Step S1: Construction of the multimodal raw sample database (fiber microscopic images, transmission spectra, loss spectra).

[0037] This phase requires the extensive collection of hollow-core antiresonant optical fiber samples with significant differences in microstructural parameters. For each physical sample, high-magnification electron microscopy is used to acquire high-precision images of its end face (as the visual modality). Simultaneously, broadband light sources and spectral analyzers are used to acquire its transmission spectrum within the target band (as the data modality). Subsequently, benchmark tests are conducted using a precise physical truncation method to calibrate the loss spectrum data for each fiber in the corresponding band. These three dimensions of data are interconnected, collectively forming the foundational dataset supporting the subsequent large-scale training of deep neural networks.

[0038] Step S2: Architecture design and training of multimodal cascaded deep networks (image parameter extraction CNN, data fusion DNN).

[0039] A cascaded architecture integrating Convolutional Neural Networks (CNNs) and Deep Regressive Neural Networks (DNNs) was designed. The CNN front-end vision module was specifically configured to automatically parse and extract microstructural features that determine light-guiding properties from 2D images; while the DNN back-end module was responsible for multimodal fusion of the geometric parameter features output by the CNN and the input spectral features. Using a dataset established in S1, the cascaded network was trained via a computer cluster using forward inference and backward bias correction until the global loss error function reached its minimum and tended to a stable convergence state.

[0040] Step S3: Data acquisition of the hollow anti-resonant optical fiber under test (microscopic image, transmission spectrum).

[0041] Non-destructive acquisition of end-face images and transmission spectra of any hollow anti-resonant optical fiber with unknown loss is performed, and the internal microstructure parameters are automatically identified and quantified by a trained intelligent model.

[0042] Step S4: Use the model to predict the loss spectrum to achieve non-destructive characterization.

[0043] Based on the relationship between the identified microstructure parameters and spectral information, the model ultimately outputs the predicted loss spectrum of the fiber under test through forward mapping calculation.

[0044] During the detailed implementation of step S1, the end-face microscopic image used as the input for the first mode can be captured using a high-end optical microscope or a high-resolution scanning electron microscope (SEM). To ensure the robustness of the deep learning model in extracting visual features to the greatest extent, the fiber end face cut by the cleaver must have sufficient flatness and the outlines of all the thin-walled quartz capillary structures inside must be clear during the imaging operation.

[0045] The original transmission spectrum described in S1 reflects the macroscopic transmission characteristics of the optical fiber within the anti-resonance window. The fluctuation trend of the transmission spectrum contains key physical information such as the inhomogeneity of the cladding wall thickness.

[0046] The loss data calibrated using the truncation method described in S1 is used as "Ground Truth" to guide network learning. Loss calculation is based on the formula for the power difference between long and short fibers: Among them, P short With P long L represents the output transmission power of the short fiber after truncation and the long fiber before truncation at a specific wavelength, respectively. long With L short These represent the physical lengths of the corresponding long and short optical fibers, respectively, and the unit of loss is dB / km or dB / m.

[0047] Entering the core algorithm implementation stage of step S2, the cascaded deep model architecture designed in this invention (such as...) Figure 3 (As detailed) It innovatively adopts the input characteristics of convergence after dual-modal parallel processing.

[0048] The first branch channel is the high-dimensional image visual analysis channel: This channel utilizes a deep convolutional network (CNN) array configured with Leaky ReLU (Leaky Rectified Linear Unit) activation functions to accurately extract multi-scale spatial geometric features of optical fibers. Its core mathematical expression for two-dimensional convolution operations is: in, The input is a two-dimensional feature map from the previous layer. The current layer size is The convolution kernel weight matrix, For bias terms, This is the feature map output after convolution. The algorithm can completely eliminate human intervention and precisely locate the sub-micron boundary contour between the solid quartz fiber and the air pores. Specifically, refer to... Figure 3 As shown, the CNN front-end visual feature extraction module contains at least four consecutive alternating convolutional layers (Conv) and pooling layers (Pooling). To capture subtle tube wall features, the convolutional layers preferably use lightweight 3×3 two-dimensional convolutional kernels, with a stride of 1. To preserve weak negative feedback edge features such as structural defects, the cascaded parameters after each convolutional operation are set to... The Leaky ReLU activation function; its nonlinear mapping formula is defined as: Subsequently, a 2×2 max pooling layer is applied for feature dimensionality reduction. The feature selection mechanism of the pooling operation is defined as follows: in, This is a pooling sliding window. Finally, a flattening operation maps the high-dimensional feature map into a one-dimensional geometric feature vector.

[0049] Its second branch is the numerical processing channel: it uses deep neural networks (DNNs) containing multiple layers of neurons to process spectral sequences. For example... Figure 3 As shown in the intermediate layer, the system uses the feature concatenation module to align and connect the one-dimensional visual geometric feature vector output by the CNN with the original one-dimensional transmission spectrum data, forming a multimodal hybrid feature stream of "fiber microstructure parameters and transmission spectrum information", which is then input into the data fusion regression network (DNN).

[0050] Combined with appendix Figure 2As shown, when faced with complex nested tubular hollow anti-resonant optical fibers, the CNN front-end module is trained to identify and automatically and accurately extract or regress the following parameters: core diameter, number of cladding capillaries, inner cladding tube diameter, outer cladding tube diameter, inner cladding tube wall thickness, outer cladding tube wall thickness, tube gap, quartz outer tube wall thickness, and fiber outer diameter. Core diameter, number of cladding capillaries, number of cladding capillary layers, outer diameter of cladding capillary, wall thickness of cladding capillary, gap between cladding capillaries. Combined with appendix Figure 4 As shown, for a relatively simple single-loop hollow anti-resonant fiber, the CNN front-end module still performs a full-parameter scan and autonomously identifies, extracts or regresses parameters including but not limited to: core diameter, number of cladding capillaries, cladding tube diameter, cladding tube wall thickness, tube gap, quartz outer tube wall thickness, and fiber outer diameter.

[0051] In the iterative calculation and parameter update training process of step S2, this system employs an error backpropagation optimization algorithm equipped with an Adam advanced optimizer. Regarding the gradient update mechanism, the Adam optimizer combines momentum acceleration from first-order moment estimation with adaptive learning rate adjustment from second-order moment estimation, effectively improving the convergence speed of model training and enhancing the stability of the optimization process. Its core mathematical objective is to minimize the expected deviation between the predicted loss given by the network calculation and the loss value precisely measured by the truncation method in multidimensional space by adjusting the weight parameters. To effectively resist the inevitable system noise in the dataset and prevent severe overfitting of the model, a crucial L2 weight regularization penalty mechanism is injected into the objective loss function E, whose standard calculation paradigm is defined as: Among them, T i This represents the measured loss value, O i This represents the model's predicted output value, where m is the total number of wavelength sampling points. For regularization constraints; Here, W represents the weight decay coefficient, and W represents the weight matrix of each layer in the network. This mechanism, by restricting the norm of the weight matrix, greatly improves the model's generalization and prediction ability for samples with unknown microstructural variations.

[0052] In terms of model structure design, such as Figure 3 As shown, the data fusion regression (DNN) subnetwork adopts a deep multilayer perceptron architecture, specifically consisting of an input layer, three consecutive hidden layers (hidden layer 1, hidden layer 2, and hidden layer 3), and a regression output layer. The computational paradigm for fully connected forward propagation between the hidden layers is as follows: To achieve layer-by-layer abstraction and nonlinear compression of features, the neuron size of each hidden layer adopts a stepped shrinking distribution (e.g., but not limited to: 256 nodes in the first hidden layer, 128 nodes in the second hidden layer, and 64 nodes in the third hidden layer). This topology forces the network to extract the core physical features most sensitive to loss from redundant multimodal inputs. The nonlinear activation mechanism of the deep hidden layer neurons in the multi-source data fusion regression subnetwork is preferably the GELU (Gaussian Error Linear Unit) function. GELU combines nonlinear activation and stochastic regularization properties, and its precise mathematical definition depends on the cumulative distribution function of the standard normal distribution. : in, The Gaussian error function is used. The GELU activation function provides a smooth stochastic regularization effect to the input neurons by introducing the cumulative distribution characteristics of Gaussian error. This activation function has continuous and smooth nonlinear mapping characteristics, which helps to improve the expressive power and numerical stability of the network in complex nonlinear modeling problems. Furthermore, within each hidden layer, this invention introduces Layer Normalization (LN) technology. The normalization calculation process is as follows: in, and Within the same floor Mean and variance of activation values ​​of individual neurons and These are dynamically learnable scaling and translation parameters. To prevent the use of tiny constants with zero denominators, real-time mean and variance normalization is performed on the activation values ​​of each layer of neurons. This helps reduce the impact of differences in feature scales on the training process, ensuring faster gradient convergence under the Adam optimizer and enhancing the model's robustness to input perturbations to some extent.

[0053] The terminal output layer neurons are configured with non-decaying linear transfer functions (Linear Activation). This is designed to ensure that the loss predictions maintain continuous, smooth, and highly accurate numerical regression capabilities across the entire range (from the low-loss window to the high-loss resonance region).

[0054] To address the extreme loss sensitivity of hollow-core antiresonant optical fibers near the resonant band, this invention incorporates an integrated weight matrix based on the loss gradient in the training algorithm. When the spectrum is identified as being at the edge of antiresonance, the weight coefficients in that region are dynamically increased to guide the model to learn the physical boundary more accurately.

[0055] In step S3, when acquiring the data to be tested, the system needs to perform standardization processing on the image to be tested, including grayscale conversion, noise reduction, and geometric correction, to eliminate system bias caused by different shooting devices. This ensures the image clarity entering the CNN network.

[0056] The loss prediction process in step S4 does not involve any cutting or mechanical damage to the optical fiber. After the model completes the forward calculation, it can output continuous loss curves covering the entire wavelength range (e.g., from ultraviolet to mid-infrared).

[0057] The advantage of this intelligent technology approach is that even if the structural parameters of the hollow anti-resonant fiber undergo asymmetric deformation due to fluctuations in the drawing process, the CNN module can still capture these subtle changes through pixel-level features and reflect them in the loss prediction results.

[0058] The non-destructive characterization technique described in this invention is particularly suitable for online quality assessment of high-performance hollow-core antiresonant optical fibers. During the fiber drawing process, only real-time sampling and imaging of the fiber cross-section and monitoring of the transmission power are needed to accurately and in real-time assess the loss level of the currently drawn fiber.

[0059] The "image + spectrum" dual-mode input architecture proposed in this invention significantly reduces the random errors caused by manual parameter extraction compared to single manual structural parameter input, and improves the objectivity of representation.

[0060] Since this method can achieve lossless prediction, it provides a practical and feasible quantitative characterization scheme for preparing high-performance hollow anti-resonant optical fibers with limited length and extremely low intrinsic loss.

[0061] Furthermore, this solution can be deployed on portable computing platforms or in the cloud after training, greatly reducing the reliance on expensive optical hardware testing platforms.

[0062] In summary, this invention simulates complex fluctuation optical loss mechanisms using deep learning technology, providing an efficient and low-cost technical path for the research, screening, and evaluation of novel hollow optical fibers.

[0063] Finally, the technical approach and network architecture described in this application are merely preferred embodiments of the present invention. Any modifications, equivalent substitutions, logical improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for non-destructive characterization of the loss properties of a hollow core anti-resonant optical fiber, characterized in that, Includes the following steps: (1) Constructing a cascaded multimodal deep learning model: A cascaded model architecture is constructed, which includes a visual feature capture subnetwork and a multi-source data fusion regression subnetwork. The optical fiber cross-section microscopic image is used as the first input mode, the original transmission spectral data is used as the second input mode, and the corresponding full-band loss spectral data is used as the desired output target. (2) Obtain a multimodal heterogeneous sample set: By collecting hollow anti-resonant fiber samples of different structural types through experiments, obtaining their end-face microscopic image information and corresponding transmission spectrum information, and using the truncation method to test the corresponding actual loss spectrum data, together forming a multi-modal training dataset. (3) End-to-end joint training of multimodal models: Based on the model in step (1), iterative training is performed using the dataset from step (2): First, the convolutional neural network (CNN) is used as a visual feature capture subnetwork to automatically extract the geometric feature parameters of the fiber microstructure from the first input mode; then, the extracted microstructure parameters are concatenated with the transmission spectrum, i.e., the second input mode, and input into the deep neural network (DNN) for iterative training. The model weight matrix is ​​adjusted through the backpropagation algorithm to minimize the target loss function E. ; where T i is the measured loss value, O i is the model predicted output value, is the regularization constraint term; (4) Implement loss non-destructive characterization prediction: The end-face microscopic image of the hollow anti-resonant fiber under test and the corresponding transmission spectrum are input into the trained model. The model automatically identifies the microstructure features and performs forward mapping calculations in combination with the spectrum to directly predict the loss spectrum of the fiber across the entire wavelength range, thus achieving characterization.

2. The non-destructive characterization method for hollow anti-resonant optical fiber based on multimodal deep learning according to claim 1, characterized in that: Step (1) The first input mode is a digital image of the cross-section of an optical fiber obtained by an optical microscope or a scanning electron microscope; Step (3) The features automatically extracted by the visual feature capture sub-network include the core diameter, the number of cladding capillaries, the number of cladding capillary layers, the outer diameter of the cladding capillary, the wall thickness of the cladding capillary, the gap between cladding capillaries, the wall thickness of the quartz outer tube, and the outer diameter of the optical fiber.

3. The non-destructive characterization method for hollow anti-resonant optical fiber based on multimodal deep learning according to claim 1, characterized in that: The transmission spectral data of the second input mode mentioned in step (1) is the relative power distribution curve reflecting the anti-resonance window characteristics in a specific band measured by a spectrometer.

4. The non-destructive characterization method for hollow anti-resonant optical fiber based on multimodal deep learning according to claim 1, characterized in that: In step (3), the visual feature capture subnetwork selects Leaky ReLU, i.e., the modified linear unit function with leakage, as the activation function to identify and extract fiber microstructure feature information; The hidden layer of the data fusion regression subnetwork uses GELU (Gaussian error linear unit) as the activation function; the output layer uses linear regression unit to achieve continuous and high-precision prediction of loss values.

5. The non-destructive characterization method for hollow anti-resonant optical fiber based on multimodal deep learning according to claim 1, characterized in that: In step (3) during the training process, an integrated weight matrix based on the loss change gradient is introduced. According to the abrupt change sensitivity of the loss spectrum at the edge of the anti-resonance window and the resonant band, the weights of different wavelength sampling points in the loss function are dynamically adjusted.

6. The non-destructive characterization method for hollow anti-resonant optical fiber based on multimodal deep learning according to claim 1, characterized in that: The characterization process described in step (4) is implemented automatically by a computer intelligent system. The system integrates automatic image recognition, multimodal feature alignment and parallel computing inference modules to complete the closed-loop evaluation from raw data input to loss spectrum output.

Images