Deep learning-based orbital angular momentum pattern recognition method
By introducing adaptive learning rate scheduling and gradient pruning techniques, the problem of unstable training of convolutional neural networks in complex scattering environments is solved, achieving efficient and robust recognition of OAM patterns and improving the training efficiency and accuracy of the model.
Patent Information
- Application Number
- CN202511391088.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-26
- Publication Date
- 2026-01-09
AI Technical Summary
Existing OAM pattern recognition methods based on convolutional neural networks are unstable during training in complex scattering environments, have slow convergence speeds, and insufficient generalization capabilities. In particular, during the optimization of large-scale datasets and complex models, fixed or empirical learning rate strategies cannot adapt to constantly changing training conditions.
The model employs a Two-Phase Scheduler adaptive learning rate scheduler and a Dynamic Range SGD optimizer, combined with a conditional linear warm-up and validation accuracy-driven decay mechanism, to dynamically adjust the learning rate. Gradient clipping is also used to prevent gradient explosion, thereby improving model stability and convergence speed.
It significantly improves the training efficiency and accuracy of OAM pattern recognition, provides high-precision recognition capability under various scattering environments, and ensures robust convergence of the model in complex environments.
Smart Images
Figure CN121302045A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of optical communication technology, specifically relating to a method for orbital angular momentum pattern recognition based on deep learning. Background Technology
[0002] In recent years, vortex beams (OAM), as an emerging optical mode, have become a key physical carrier for achieving high-capacity transmission in optical communication due to their orbital angular momentum. However, turbulent (atmospheric / oceanic) scattering effects in actual transmission environments, as well as mode coupling and dispersion in optical fibers, can cause severe wavefront distortion and modal crosstalk, resulting in characteristically degraded speckle patterns at the receiver, which seriously affects the recognition accuracy of OAM modes.
[0003] Currently, Convolutional Neural Networks (CNNs) have been widely used in OAM recognition tasks, successfully addressing partial scattering effects thanks to their advantages in image classification and feature extraction. However, existing methods still face problems such as unstable training, slow convergence speed, and insufficient generalization ability when dealing with complex scattering environments. In particular, during the optimization of large-scale datasets and complex models, fixed or empirical learning rate strategies cannot effectively adapt to constantly changing training conditions.
[0004] Therefore, designing an adaptive learning rate adjustment mechanism that can dynamically adjust the learning rate according to changes in scattering effects, and enhancing the stability and convergence speed of the model through an improved optimizer, has become an important research challenge. Summary of the Invention
[0005] The purpose of this invention is to overcome the above-mentioned problems in traditional technologies and provide a deep learning-based method for orbital angular momentum pattern recognition.
[0006] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:
[0007] This invention provides a deep learning-based method for orbital angular momentum pattern recognition, comprising the following steps:
[0008] S1. The initial learning rate is dynamically and adaptively adjusted using the Two Phase Scheduler adaptive learning rate scheduler, which is based on two core mechanisms: conditional linear warm-up and verification accuracy-driven decay.
[0009] S2. Design a Dynamic Range SGD optimizer, based on the SGD optimizer, to solve the problem that the training gradient may escape local minima or stagnate.
[0010] Furthermore, in the aforementioned deep learning-based orbital angular momentum pattern recognition method, the Two PhaseScheduler adaptive learning rate scheduler is equipped with a linear preheating mechanism, which is activated only when the initial learning rate is less than a preset condition, in order to avoid unnecessary warming up of the already sufficiently large initial learning rate.
[0011] Furthermore, in the aforementioned deep learning-based orbital angular momentum pattern recognition method, the Two PhaseScheduler adaptive learning rate scheduler is equipped with a verification accuracy-driven decay mechanism, which is introduced on the basis of cosine annealing decay to solve the oscillation and local minima trap problems that occur in the later training stages.
[0012] Furthermore, in the aforementioned deep learning-based orbital angular momentum pattern recognition method, the Two PhaseScheduler adaptive learning rate scheduler sets an activation condition. When the initial learning rate is greater than or equal to the activation condition, the decay function in the optimizer is invoked; when the initial learning rate is less than the activation condition, the warm-up function is triggered, first warming up the learning rate to a preset value through the warm-up algorithm, and then using the decay function normally.
[0013] Furthermore, in the aforementioned deep learning-based orbital angular momentum pattern recognition method, the Two PhaseScheduler adaptive learning rate scheduler can improve the recognition capability at different initial learning rates between 0.1 and 0.001 when recognizing OAM scattering spot data on the ResNet50 deep learning network.
[0014] Furthermore, in the aforementioned deep learning-based orbital angular momentum pattern recognition method, the Dynamic RangeSGD optimizer incorporates gradient clipping techniques. By constraining the L2 norm of the gradient, gradient explosion is prevented during training, ensuring the training stability of the ResNet deep network in complex speckle OAM classification.
[0015] Furthermore, the aforementioned deep learning-based orbital angular momentum pattern recognition method ensures robust convergence of the model within different initial learning rate ranges, significantly accelerating the training process.
[0016] The beneficial effects of this invention are:
[0017] The OAM pattern recognition method of this invention significantly improves the efficiency and accuracy of model training by introducing adaptive learning rate scheduling and gradient pruning techniques. In experiments, the proposed Two Phase Scheduler and Dynamic RangeSGD optimizer ensure robust convergence of the model across different initial learning rate ranges, significantly accelerating the training process and effectively avoiding common convergence oscillation problems. This method not only provides high-precision recognition results in OAM speckle classification tasks but also exhibits good generalization ability under various scattering environments, providing an efficient and robust solution to the OAM pattern recognition problem. The successful application of this scheme provides a valuable reference for future deep learning tasks with other complex data characteristics (such as high dimensionality, non-convexity, and noise interference).
[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the above advantages at the same time. Attached Figure Description
[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a diagram of the experimental optical path device in the embodiment;
[0021] Figure 2 The above examples show OAM scattering patterns with topological charge numbers l of 1, 2, 3, and 4.
[0022] Figure 3 This is a graph showing the change of the learning rate as a function of epochs based on an adaptive learning rate architecture.
[0023] Figure 4 A comparison of the training performance of SGD and AdamW under three initial learning rates using the Two Phase Scheduler;
[0024] Among them, (a) lr-0.1; (b) lr-0.01; (c) lr-0.001. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] This embodiment proposes an OAM pattern recognition method based on deep learning and adaptive learning rate, including the following steps:
[0027] S1. The Two-Phase Scheduler adaptive learning rate scheduler is based on two core mechanisms: conditional linear warm-up and validation accuracy-driven decay. To address the issue of slow training (convergence) caused by an excessively low initial learning rate, this embodiment designs a linear warm-up mechanism and sets a learning rate conditional activation: warm-up is only performed when a certain initial learning rate is less than a set condition, avoiding unnecessary warm-up of an already sufficiently large initial learning rate. Furthermore, the periodic restart mechanism is explicitly abandoned, as it easily leads to convergence oscillations in the later stages. To address oscillations and local minima traps in later training, this embodiment introduces a validation accuracy-driven decay mechanism on top of cosine annealing decay. The designed Two-Phase Scheduler achieves more refined and dynamic control of the learning rate, more closely aligned with the training state (especially the model's learning progress on scattering data), through conditional warm-up based on an initial learning rate threshold and conditional decay driven by validation accuracy.
[0028] S2. Based on the SGD optimizer, an improved version of Dynamic Range SGD was designed to address the problem that training gradients may escape local minima or stagnate. Considering that ResNet deep networks may experience gradient explosion in complex speckle OAM classification, leading to unstable training, this embodiment considers gradient clipping as a key improvement in the design of Dynamic Range SGD.
[0029] In step S1, the Two Phase Scheduler adaptive learning rate scheduler used improves the recognition capability of identifying OAM scattering spot data on the ResNet50 deep learning network by using different initial learning rates between 0.1 and 0.001 through two activatable phases.
[0030] In step S2, the Dynamic Range SGD optimizer used incorporates gradient clipping techniques to prevent gradient explosion during training by constraining the L2 norm of the gradient.
[0031] like Figure 1 The diagram shown is a simulation of the optical path device used in this embodiment. When using the experimental setup, different modes of OAM images can be obtained using an SLM and computer A. After passing through a scattering sheet, a CCD and computer B are used to acquire the scattered images.
[0032] like Figure 2 As shown, the scattered light spots obtained after OAM images of different modes pass through the scattering sheet.
[0033] like Figure 3 As shown, the initial learning rate lr is dynamically and adaptively adjusted by the Two Phase Scheduler adaptive learning rate scheduler, which enables more precise and dynamic control of the learning rate that is more in line with the training state (especially the learning progress of the model on the scattering data).
[0034] In this embodiment, an activation condition lr equal to 0.01 is set to divide the initial learning rate. If lr is greater than or equal to the activation condition, the decay function in the optimizer is called; if the initial learning rate is less than the activation condition, activation is triggered and the warm-up function is used.
[0035] In this embodiment, if the warm-up function is not activated, the learning rate (lr) is adjusted based on the training effect after each cooling-off period, and the decay algorithm is used for adjustment. If the training effect is good, the current lr is maintained. If the warm-up function is activated, the learning rate is warmed up to lr equal to 0.01 using the warm-up algorithm before the decay function is used normally.
[0036] S2 is an improved version of Dynamic Range SGD that addresses the problem that training gradients may escape local minima or stagnate.
[0037] like Figure 4 The figure shows the ACC results of identifying and detecting OAM patterns with topological charge numbers l of 1, 2, 3, and 4 using the method of this embodiment. Figure 4 Using three initial learning rates employed in this study, during training, it was found that compared to the conventional fixed learning rate, the proposed dynamic learning rate architecture significantly accelerated loss convergence and achieved higher or comparable final accuracy in fewer epochs. For example, in... Figure 4 In case (a) where lr = 0.1 (an excessively large initial learning rate), the accuracy of the adaptive policy gradually converges and reaches its final value after the 42nd epoch; while the accuracy of the fixed policy oscillates repeatedly and fails to converge. Figure 4In example (b) with lr = 0.01 (close to the optimal learning rate), the final accuracy values using a fixed learning rate and an adaptive learning rate are essentially the same, but the adaptive learning rate clearly achieves higher or comparable final accuracy in fewer epochs and converges more stably. Figure 4 (c) In the case of lr = 0.001 (too small initial learning rate), the adaptive strategy is significantly better than the fixed strategy in terms of accuracy.
[0038] This embodiment proposes a deep learning optimization framework for OAM pattern recognition. By combining an adaptive learning rate scheduler and a dynamic range gradient pruning optimizer, it significantly improves the training efficiency and accuracy of the model in complex scattering environments. Experimental results show that the proposed method can achieve stable convergence with a wide range of initial learning rates and provides high-precision recognition results in multiple scattering scenarios. This method provides an efficient and robust solution for OAM pattern recognition, has broad application prospects, and provides a strong reference for deep learning tasks handling other complex data characteristics.
[0039] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. A deep learning-based method for orbital angular momentum pattern recognition, characterized in that, Includes the following steps: S1. The initial learning rate is dynamically and adaptively adjusted using the Two Phase Scheduler adaptive learning rate scheduler, which is based on two core mechanisms: conditional linear warm-up and verification accuracy-driven decay. S2. Design a Dynamic Range SGD optimizer, based on the SGD optimizer, to solve the problem that the training gradient may escape local minima or stagnate.
2. The orbital angular momentum pattern recognition method based on deep learning according to claim 1, characterized in that, The Two Phase Scheduler adaptive learning rate scheduler has a linear warm-up mechanism that is activated only when the initial learning rate is less than a preset condition, in order to avoid unnecessary warming up of an already sufficiently large initial learning rate.
3. The orbital angular momentum pattern recognition method based on deep learning according to claim 1, characterized in that, The Two Phase Scheduler adaptive learning rate scheduler has a verification accuracy-driven decay mechanism, which is introduced on the basis of cosine annealing decay to solve the oscillation and local minima trap problems that occur in the later stages of training.
4. The orbital angular momentum pattern recognition method based on deep learning according to claim 1, characterized in that, The Two Phase Scheduler adaptive learning rate scheduler sets an activation condition. When the initial learning rate is greater than or equal to the activation condition, the decay function in the optimizer is invoked. When the initial learning rate is less than the activation condition, the warmup function is triggered. The learning rate is first warmed up to a preset value through the warmup algorithm, and then the decay function is used normally.
5. The orbital angular momentum pattern recognition method based on deep learning according to claim 1, characterized in that, The Two Phase Scheduler adaptive learning rate scheduler can improve recognition capabilities when identifying OAM scattering spot data on the ResNet50 deep learning network with different initial learning rates between 0.1 and 0.
001.
6. The orbital angular momentum pattern recognition method based on deep learning according to claim 1, characterized in that, The Dynamic Range SGD optimizer incorporates gradient clipping techniques, which constrain the L2 norm of gradients to prevent gradient explosion during training and ensure the training stability of the ResNet deep network in complex speckle OAM classification.
7. The orbital angular momentum pattern recognition method based on deep learning according to claim 1, characterized in that, This orbital angular momentum pattern recognition method ensures robust convergence of the model within different initial learning rates, significantly accelerating the training process.