Speckle wavelength measurement method and device based on pattern embedding machine learning

Abstract

The invention discloses a speckle wavelength measurement method based on mode embedding machine learning. The method comprises the following steps: S1, acquiring a speckle pattern after a to-be-measured optical signal passes through a fixed scattering element; s2, extracting a modal coefficient vector from the speckle pattern; s3, embedding the modal coefficient vector input mode into the machine learning model so as to invert the wavelength of the optical signal to be measured; the mode embedded machine learning model is obtained through pre-training by the following method: collecting speckle patterns of a series of optical signals with known wavelengths after passing through the fixed scattering element, and extracting corresponding modal coefficient vectors; and training a machine learning regression model by taking the obtained modal coefficient vector and the corresponding wavelength as input and output respectively. The invention also discloses a speckle wavelength measurement device based on pattern embedding machine learning. According to the invention, the rapid change wavelength can be measured in a high-precision and large-range manner, the noise robustness is better, and the consumption of computing resources and time is greatly reduced.

Description

Technical Field

[0001] The invention relates to a speckle wavelength measurement method and belongs to the technical field of optical measurement. Background Art

[0002] Wavelength is a fundamental parameter describing light waves and other wave phenomena, related to photon energy, dispersion, and coherence properties. Wavelength measurement has become a core topic in modern scientific research due to its widespread application in multiple disciplines, including physics and precision measurement, chemical analysis, biomedicine, astronomy, and quantum optics. As application scenarios become increasingly diverse, new demands are being placed on wavelength measurement. Traditional interferometers require precise motion devices and mirrors to ensure accuracy, making them unsuitable for measuring rapidly changing wavelengths and prohibitively expensive due to their size. Although dispersive systems such as gratings and prisms offer compact structures and wide wavelength coverage, resolution is significantly compromised, and system size increases linearly with accuracy, failing to meet application requirements.

[0003] Emerging speckle wavelength meters utilize simple devices to map wavelength to speckle patterns and demodulate them, demonstrating miniaturization, high accuracy, and fast response. However, wavelength accuracy remains constrained by the speckle correlation limit, which is determined by the laser linewidth and the inherent properties of the dispersive medium. While advanced demodulation algorithms and compact integrated platforms have enabled speckle wavelength meters to approach theoretical resolution limits, overcoming this limitation remains a challenging problem. With the development of intelligent optimization algorithms and data-driven technologies, speckle wavelength inversion has evolved from transfer matrix analysis to machine learning. A compact wavelength meter based on multiple scattering in a silicon chip reconstructs wavelength after calibration using a combination of matrix pseudoinversion and nonlinear optimization. It achieves 0.75 nm resolution at 1500 nm, close to the speckle correlation limit of 0.6 nm. Similar reconstruction algorithms have been used on low-cost multimode optical fibers to achieve picometer-level resolution. These studies measure wavelength by inverting the transfer matrix. While these methods are widely applicable and interpretable, they often rely on multiple calibrations and require time-consuming iterative optimization or cumbersome denoising algorithms.

[0004] A chip-scale atomic wavelength meter based on machine learning has been proposed to improve measurement accuracy and stability. Using a K-nearest neighbor classification algorithm, it achieved a measurement bandwidth of 1 nm and an optimal resolution of 0.5 pm, outperforming the transfer matrix method. Furthermore, the use of a convolutional neural network (CNN) for speckle image classification provides the wavelength meter with high dynamic range and excellent noise robustness. The trained deep model achieved attometer-level accuracy on a dataset consisting of five discrete wavelengths. Furthermore, wavelength reconstruction using a convolutional long short-term memory network (CNN-LSTM) denoising algorithm completes at millisecond speeds with an accuracy of 0.5 nm within a 10 nm wavelength range, reaching the speckle correlation limit. These data-driven machine learning methods have enabled the speckle wavelength meter to achieve high accuracy and noise robustness. However, its performance is limited by the amount of data and the complexity of the model, and the speckle correlation limit cannot be ignored. In classification tasks, only a few specific wavelength states can be estimated from the speckle pattern, resulting in poor performance when dealing with continuously varying wavelengths. In regression tasks, wavelength accuracy relies on a deep model, which requires significant memory and computational resources. Therefore, it is always necessary to use a large amount of speckle pattern data and complex models for tedious training to achieve high-resolution and wide-bandwidth wavelength measurements. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the problems of existing speckle wavelength measurement technologies, such as the speckle correlation limit, poor noise robustness, difficulty in measuring continuously changing wavelengths, and heavy dependence of performance on large data sets and deep models. A speckle wavelength measurement method based on pattern embedding machine learning is provided. The present invention proposes a pattern embedding machine learning framework for the first time, which integrates the beam propagation model with the machine learning algorithm, constrains the input data with the physical model, and utilizes the continuous evolution characteristics of the modal coefficients to break through the speckle correlation limit, solving the accuracy bottleneck of previous data-driven models. While achieving high precision and measuring rapidly and continuously changing wavelengths, the wavelength meter's measurement accuracy and noise robustness are improved, the speckle data required for training the model and the training time are reduced, and the overall dependence of the speckle wavelength meter on complex networks and large data is avoided. The consumption of computing resources and the time for redeployment are reduced, providing a new approach to enhance speckle measurement.

[0006] The present invention specifically adopts the following technical solutions to solve the above technical problems:

[0007] The speckle wavelength measurement method based on pattern embedding machine learning includes the following steps:

[0008] S1, obtaining a speckle pattern of the light signal to be measured after passing through a fixed scattering element;

[0009] S2. extracting a modal coefficient vector from the speckle pattern;

[0010] S3. Inputting the modal coefficient vector into a pattern-embedded machine learning model to invert the wavelength of the optical signal to be measured; the pattern-embedded machine learning model is pre-trained by: collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting the corresponding modal coefficient vectors; and training a machine learning regression model using the obtained modal coefficient vectors and the corresponding wavelengths as input and output, respectively.

[0011] Preferably, a mode decomposition method based on a beam propagation model is used to extract the modal coefficient vector from the speckle pattern.

[0012] Preferably, the fixed scattering element is an optical fiber containing no more than 15 linear polarization modes.

[0013] Further preferably, the length of the optical fiber is 5 m.

[0014] Preferably, when collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting corresponding modal coefficient vectors, the wavelength tuning rate of the series of optical signals of known wavelengths is 6.5 pm / s.

[0015] Based on the same inventive concept, the following technical solutions can also be obtained:

[0016] The speckle wavelength measurement device based on pattern embedding machine learning includes:

[0017] A speckle pattern acquisition module is used to acquire the speckle pattern of the light signal to be measured after passing through the fixed scattering element;

[0018] a mode decomposition module for extracting a modal coefficient vector from the speckle pattern;

[0019] A mode embedding machine learning model is used to invert the wavelength of the optical signal to be measured using the modal coefficient vector as input. The mode embedding machine learning model is pre-trained by collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting the corresponding modal coefficient vectors; and training a machine learning regression model using the obtained modal coefficient vectors and the corresponding wavelengths as input and output, respectively.

[0020] Preferably, the mode decomposition module extracts the modal coefficient vector from the speckle pattern using a mode decomposition method based on a beam propagation model.

[0021] Preferably, the fixed scattering element is an optical fiber containing no more than 15 linear polarization modes.

[0022] Further preferably, the length of the optical fiber is 5 m.

[0023] Preferably, when collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting corresponding modal coefficient vectors, the wavelength tuning rate of the series of optical signals of known wavelengths is 6.5 pm / s.

[0024] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0025] This invention proposes the concept of pattern-embedded machine learning for the first time, using the beam propagation model as a physical constraint for the machine learning model, thus overcoming the accuracy bottleneck of traditional data-driven speckle wavelength meters. Based on the continuity and predictability of the evolution of the modal coefficient vector as the wavelength changes, high-precision measurement of continuously changing wavelengths over a large range is achieved, with accuracy exceeding the speckle correlation limit. Furthermore, the accuracy does not decrease when measuring rapidly changing wavelengths, enabling the speckle wavelength meter to adapt to more complex application scenarios.

[0026] Compared to existing speckle wavelength meter technology, the modal coefficient vectors retrieved from speckle patterns exhibit strong noise immunity. Thanks to the introduction of a physical model, the wavelength inversion from modal features is learned, achieving higher accuracy without the need to collect large amounts of diverse data to ensure generalization. This is due to the model's efficient training on modal data that follows the beam propagation model, rather than wasting time learning speckle noise. Furthermore, the significant reduction in model training time and the amount of training data required avoids the wavelength meter's over-reliance on complex networks and large amounts of data, further reducing the computational resource consumption and redeployment time of the speckle wavelength meter in actual operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram comparing the principles of conventional data-driven and pattern-embedded speckle wavelength measurements according to the present invention;

[0028] Figure 2 Spectral correlation functions when using different lengths of optical fiber for the speckle wavelength measurement device;

[0029] Figure 3 Schematic diagram of the speckle wavelength measurement device and workflow based on pattern-embedded machine learning of the present invention;

[0030] Figure 4 Schematic diagram of the evolution of some modal coefficients in the optical fiber with the input laser wavelength;

[0031] Figure 5 is the estimated wavelength distribution within the measurement range and the corresponding absolute error;

[0032] Figure 6 Schematic diagram of the evolution of speckle along with system noise;

[0033] Figure 7Comparison results of modal coefficients and the first principal component of speckle pattern;

[0034] Figure 8 The error distribution statistics of wavelength estimation using few-mode fibers of different lengths are shown;

[0035] Figure 9 The error distribution statistics of wavelength estimation at different laser tuning rates are shown;

[0036] Figure 10 Schematic diagram of the measurement performance and estimated error distribution of rapidly changing wavelengths;

[0037] Figure 11 The results are a comparative evaluation of the performance of the traditional data-driven and pattern-embedded speckle wavelength measurements of the present invention. DETAILED DESCRIPTION

[0038] To address the shortcomings of existing speckle wavelength measurement technology, the present invention addresses this issue by leveraging the anti-speckle noise properties exhibited by modal coefficients, as well as their continuity and predictability with wavelength changes, to invert wavelength using a pattern-embedded machine learning approach. This approach not only enables high-precision, large-scale measurement of rapidly changing wavelengths, but also improves the noise robustness of the wavelength meter, reduces the model's training data and time, and avoids the over-reliance of speckle wavelength measurement on complex networks and large amounts of data. This reduces the consumption of computing resources and redeployment time, thus providing a new approach to enhancing speckle measurement technology.

[0039] The present invention proposes a method for measuring speckle wavelength based on pattern embedding machine learning, comprising the following steps:

[0040] S1, obtaining a speckle pattern of the light signal to be measured after passing through a fixed scattering element;

[0041] S2. extracting a modal coefficient vector from the speckle pattern;

[0042] S3. Inputting the modal coefficient vector into a pattern-embedded machine learning model to invert the wavelength of the optical signal to be measured; the pattern-embedded machine learning model is pre-trained by: collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting the corresponding modal coefficient vectors; and training a machine learning regression model using the obtained modal coefficient vectors and the corresponding wavelengths as input and output, respectively.

[0043] The speckle wavelength measurement device based on pattern embedding machine learning proposed in the present invention includes:

[0044] A speckle pattern acquisition module is used to acquire the speckle pattern of the light signal to be measured after passing through the fixed scattering element;

[0045] a mode decomposition module for extracting a modal coefficient vector from the speckle pattern;

[0046] A mode embedding machine learning model is used to invert the wavelength of the optical signal to be measured using the modal coefficient vector as input. The mode embedding machine learning model is pre-trained by collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting the corresponding modal coefficient vectors; and training a machine learning regression model using the obtained modal coefficient vectors and the corresponding wavelengths as input and output, respectively.

[0047] The present invention can extract modal coefficient vectors from speckle patterns based on a variety of existing mode decomposition methods, such as experimental measurement methods such as frequency-domain cross-correlation imaging, spatially and spectrally resolved imaging, wavefront measurement, and low-coherence interferometry techniques, or numerical analysis methods such as randomized parallel gradient descent, inverse matrix solution, and deep learning. These mode decomposition methods can partially or completely obtain the optical fiber modal coefficients. The inventors have found that using mode decomposition methods based on beam propagation models, such as randomized parallel gradient descent and inverse matrix solution, only a single intensity distribution measurement is required to accurately calculate the complete modal coefficients of the modes present in the optical fiber. At the same time, it does not rely on additional reference beams and can achieve high-precision and rapid decomposition using simple experimental equipment. Therefore, preferably, a mode decomposition method based on a beam propagation model is used to extract modal coefficient vectors from the speckle pattern.

[0048] The present invention can be implemented on the basis of various existing speckle wavelength meters, thereby significantly improving wavelength measurement accuracy. However, most existing speckle wavelength meters use large-core multimode optical fibers or expensive scattering chips as measurement elements. On the one hand, the increased number of modes makes mode decomposition difficult, and the modal coefficient vectors obtained by decomposition still contain multiple characteristic dimensions, requiring deep models to undergo tedious training to achieve high-resolution and wide-bandwidth wavelength measurement. On the other hand, complex mode coupling and dispersion may introduce more environmental interference caused by temperature changes and vibrations, reducing the stability of the speckle pattern and, in turn, the actual wavelength resolution. Therefore, preferably, the fixed scattering element is an optical fiber (multimode fiber or few-mode fiber) containing no more than 15 linear polarization modes.

[0049] The accuracy of the speckle wavelength measurement device of the present invention is significantly related to the fiber length and the laser wavelength tuning rate during training. Mode coupling and intermodal dispersion accumulate and become complex in long optical fibers, making speckle more sensitive to wavelength changes, thereby improving spectral resolution. However, excessively long optical fibers may introduce more environmental noise and instability. High-speed wavelength tuning improves data acquisition efficiency, but the nonlinear effects of intermodal transients may form unstable and fuzzy speckle, reducing wavelength resolution and model generalization. Low-speed tuning, on the other hand, collects more and denser speckle-wavelength data. Although the trained model can achieve higher regression accuracy and robustness, it increases the time cost and training complexity. As acquisition time increases, wavelength drift and speckle distortion may limit the accuracy of wavelength estimation. Therefore, it is necessary to optimize these two parameters to achieve the best wavelength measurement effect. After extensive analytical experiments, it was found that the best wavelength measurement effect is achieved when the fiber length is 5m and the laser wavelength tuning rate during training is 6.5pm / s (that is, when collecting the speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting the corresponding modal coefficient vectors, the wavelength tuning rate of the series of optical signals of known wavelengths is 6.5pm / s).

[0050] To facilitate public understanding, the technical solution and technical effects of the present invention are further described in detail below with reference to the accompanying drawings:

[0051] Figure 1 The following figure shows a comparison of the principles of the traditional data-driven speckle wavelength meter and the present invention (hereinafter referred to as the pattern-embedded type). The traditional data-driven speckle wavelength meter directly uses the speckle pattern of the input light signal after passing through the scattering element as the data input of the neural network model to invert the wavelength of the incident light signal; unlike the traditional data-driven method, the speckle wavelength measurement device based on pattern-embedded machine learning combines physical models with machine learning technology, such as Figure 1 As shown in Figure 2, the random speckle formed on the end face of the scattering element (multimode / few-mode fiber) is not directly input into the machine learning model for wavelength estimation, but the modal coefficients contained therein (including the mode amplitude ρ j and relative phase θ j ), machine learning estimates wavelength based on multiple orthogonal modal information, so it is called mode embedding, where the modal information and speckle follow the beam propagation model (BPM) of the optical fiber. It describes the light field as N fiber eigenmodes Ψ j The linear superposition of (x, y), the light intensity I(x, y) can be expressed as:

[0052]

[0053] Traditional speckle wavemeters infer the wavelength of incident light by analyzing the evolving characteristics of the speckle pattern formed by coherent light passing through a scattering medium. The wavelength information is encoded as microstructural changes (intensity distribution) in the speckle pattern. However, speckle exhibits decorrelation with small wavelength shifts Δλ, and this correlation decay results in the system being unable to resolve smaller wavelength differences. Numerous studies have demonstrated that the half-width at half-maximum (HWHM) of the spectral correlation function of speckle as it evolves with incident wavelength can characterize the theoretical resolution of the wavemeter, also known as the speckle correlation limit. The spectral correlation function of speckle can be expressed as:

[0054]

[0055] where I(λ,x) is the speckle intensity at spatial position x when the input wavelength is λ, λ and x denote the average value with respect to wavelength and position, respectively. Figure 2 The spectral correlation function of the wavelength meter when using different lengths of few-mode fibers in the experiment is demonstrated. The half-width at half maximum of the correlation curves of the 3m and 5m long fibers are 0.0241nm and 0.0072nm, respectively, representing the theoretical accuracy of the two.

[0056] Figure 3 The present invention shows a speckle wavelength measurement device and workflow based on mode embedded machine learning. The narrow linewidth laser emitted by the tunable laser is coupled into the few-mode fiber through a mode multiplexer (six-mode photon lantern). The laser linewidth is 200kHz, which can provide fine wavelength resolution. The adjustable optical attenuator is used to adjust the system optical power, and the polarization controller selects the polarization component of the light beam to effectively suppress polarization crosstalk. The few-mode fiber in this embodiment is a standard six-mode commercial step-index fiber with a core diameter of 18.5μm and a cladding diameter of 125μm. It supports LP at an operating wavelength of 1550nm. 01 LP 02 LP 11e LP 11o LP 21e and LP 21o Six linear polarization modes. The speckle pattern in the optical fiber, which encodes wavelength information, is captured by a CCD camera with a pixel pitch of 20 μm.

[0057] like Figure 3 As shown in the figure, the speckle wavelength measurement device of the present invention obtains each modal coefficient from the speckle pattern through mode decomposition, achieving accurate characterization of the light field. This process simplifies the cumbersome two-dimensional image matrix into modal vectors describing each orthogonal mode without losing wavelength information. In this embodiment, the mode decomposition module can use mode decomposition methods based on the beam propagation model, such as random parallel gradient descent and inverse matrix solution, and the mode amplitude is normalized. To ensure the efficiency and stability of the algorithm, a machine learning model is used to invert the corresponding wavelength from the modal information (the LightGBM model is used in this embodiment), and the mean square error (MSE) is selected as the loss function of the optimization model, which can be expressed as:

[0058]

[0059] where λ i and are the actual wavelength associated with the speckle and the estimated wavelength of the model, respectively, and N is the number of samples. Machine learning based on LightGBM performs well in memory usage and training speed, and is particularly suitable for processing large-scale high-dimensional feature data in fiber optic measurement or sensing applications. Compared with traditional data-driven speckle wavelength meters, the multiple modes in the pattern embedding machine learning model can be regarded as different feature dimensions, and the continuous variation of several modal coefficients with wavelength is easier to model analysis than the randomly evolving speckle, thereby reducing the dependence on complex deep models and large training data sets. Therefore, speckle patterns at different wavelengths are collected and mode decomposition is performed. The speckle wavelength meter with pattern embedding machine learning performs regression tasks based on the modal coefficient vector, which helps to achieve high-precision wavelength measurement.

[0060] By varying the wavelength of the narrow-linewidth laser light emitted by a tunable laser, a series of corresponding modal coefficient vectors can be obtained. The normalized modal coefficient vectors and the corresponding wavelengths are combined into modal vector-wavelength data pairs, which are used as the input and output of a machine learning regression model for training, thereby obtaining a mode embedding machine learning model. Once the mode embedding machine learning model is trained, the measurement device can be used to measure the wavelength of the optical signal to be measured.

[0061] To verify the technical effectiveness of the present invention's technical solution, a verification experiment was conducted using the aforementioned measurement device. In the experiment, a tunable laser was set to swept mode, and the laser wavelength was finely tuned within the range of 1548.73 nm to 1549.51 nm. A 3-meter-long few-mode fiber was initially selected. The speckle pattern at each wavelength was collected and subjected to mode decomposition to obtain modal vectors. For speckle formed by the interference of n linear polarization modes, the modal vector size after mode decomposition is 2n-1 (where θ1 = 0).

[0062] Figure 4 Shows the partial mode amplitude (LP) in the six-mode fiber 01 , LP 02 and LP 11e ) and relative phase (LP 02 , LP 11e and LP 11o) evolves with the input laser wavelength, where the mode field information changes continuously with wavelength. The differences in the correlation between different modes and wavelength reflect the diversity of mode responses. The obtained modal coefficient vector-wavelength data pairs are divided into training set, validation set, and test set in an 8:1:1 ratio to train, adjust, and evaluate the model. After the machine learning model was trained for 2000 iterations with an initial learning rate of 0.06, the MSE of the validation set continued to decrease to 2.69×10 -6 , which is close to the training set, indicating that the model has been effectively trained and has good generalization ability. The performance of the model in estimating wavelength on the untrained modal dataset is as follows Figure 5 As shown, the scattered dots represent the estimated wavelengths, the dashed line represents the theoretical value, and the dotted line represents the absolute error in the estimated wavelength. The estimated values ​​are concentrated around the theoretical value, with a standard deviation of 1.64 pm. Therefore, preliminary results indicate that the speckle wavelength meter based on pattern embedding machine learning can accurately estimate wavelengths within a continuous range of 0.78 nm, exceeding the theoretical resolution by more than 14 times and surpassing the speckle correlation limit.

[0063] The speckle wavelength meter based on pattern embedding machine learning achieves higher accuracy in a continuous range. In fact, it uses the mode field information that changes continuously with wavelength rather than directly using the sparse and random speckle pattern. In order to confirm and evaluate the characteristic differences between the modal field and the speckle pattern, it is necessary to compare the noise suppression capabilities of the two in an experimental environment, such as Figure 6 Figure 2 shows the evolution of the speckle pattern as system noise changes while the input wavelength remains constant. A total of 3,000 speckle images were captured by the CCD at a frame rate of 50 fps and an exposure time of 10 ms. Principal component analysis was used to reduce the dimensionality of the standardized speckle pattern and modal coefficient vectors to assess differences in the principal component changes between the two. Figure 7 The analysis results of the modal coefficient vectors and speckle patterns are presented. The standard deviation of the first principal component is 2.59 for the modal coefficient vectors and 51.54 for the speckle pattern, an improvement of nearly 20 times. This indicates that the mode field data is less sensitive to environmental disturbances and noise, while speckle is easily affected by environmental and instrument noise, especially CCD noise.

[0064] Next, we evaluate the performance of the speckle wavelength meter based on pattern embedding machine learning when using different lengths of few-mode fibers to confirm its impact on wavelength estimation. On the test set, the error distribution of wavelength estimation using four different fiber lengths from 3m to 9m is shown below. Figure 8As shown. The estimation errors on the continuous interval are distinguished by different line shapes, which obey the Gaussian distribution. The performance of wavelength meters using different lengths of optical fiber varies significantly. The short dotted line (5m) is sharper and steeper than the others, and has higher accuracy. The accuracy of the wavelength measurement is evaluated by standard deviation. The measurement accuracy of the speckle wavelength meter based on pattern embedding machine learning is always better than the theoretical resolution. Among them, the wavelength meter using 5m long optical fiber achieves an accuracy of 0.62pm, which exceeds the theoretical limit δλ2 by nearly 12 times. In addition, as the optical fiber length increases from 3m to 5m, the wavelength accuracy improves significantly. If it continues to increase to 9m, the accuracy deteriorates due to environmental interference and system noise, which is consistent with expectations.

[0065] Next, we keep the linewidth constant and the fiber length at 5m and further evaluate the performance of the speckle wavelength meter based on pattern embedding machine learning when changing the laser tuning rate. Training data is collected at different tuning periods T. The absolute error of the model's estimated wavelength on the test set is as follows: Figure 9 As shown, the solid line indicates the wavelength change within the range of 0.78nm at a tuning rate of 19.5pm / s. Within the continuous bandwidth, the wavelength estimation error at different tuning periods is smaller when the optimal fiber length is achieved, and the maximum absolute error always does not exceed 2pm. The distribution of the statistically corresponding estimation deviations is close to a Gaussian distribution. As the tuning rate slows down, adjacent speckles carry more subtle wavelength change information for model training. The wavelength meter estimation deviation gradually decreases, and the highest accuracy reaches 0.28pm (when T = 120s), which exceeds the speckle correlation limit by 25.7 times. In addition, when T increases to 160s, the environmental noise limits the further improvement of accuracy by slowing down the tuning rate.

[0066] High-precision measurement of rapidly changing wavelengths is an important capability of a wavelength meter. Its value lies in the real-time capture and precise analysis of dynamic optical signals, especially in the fields of precision wavelength manipulation and astronomical spectral observation. The following measurement of the wavelength of the frequency sweep is used to evaluate the performance of the speckle wavelength meter based on pattern embedding machine learning. The wavelength changes periodically from 1548.73nm to 1549.51nm at intervals of 78pm, at a rate of 50Hz. The estimated wavelength within the entire working bandwidth is as follows: Figure 10 As shown, the rapidly changing wavelength is accurately estimated with a standard deviation of 0.277 pm and less than ±1 pm. Although faster rates may be limited by the CCD frame rate, the present invention can still maintain high accuracy regardless of the wavelength change speed.

[0067] A wavelength meter based on pattern embedding machine learning estimates wavelength using mode fields extracted from speckle patterns, rejecting system noise and thus exceeding the speckle correlation limit. It is important to compare this with a speckle wavelength meter based on a data-driven model. We built an additional data-driven model based on a classic residual network architecture on the same computing platform. This model consists of 12 convolutional layers, along with corresponding normalization layers and activation functions (ReLU) layers. The input speckle pattern has a size of 256×256, and the output of the fully connected layer is mapped to wavelength.

[0068] The performance of the two models is compared on the same speckle dataset. The evaluation results of the two models, data-driven and modal embedding, are shown in Figure 2. Figure 11 As shown in Figure 2, including estimation accuracy, validation set loss, number of speckle patterns used for training, training time, and inference time. The estimation accuracy is 7.16pm and 0.28pm, respectively, which is an improvement of 25.6 times. The loss on the validation set is 5.12×10 -5 and 8.01×10 -8 . This shows that the wavelength meter based on pattern embedding machine learning is able to surpass the speckle correlation limitation by rejecting environmental and speckle noise, which is not achievable for traditional speckle wavelength meters. At the same time, the number of speckle patterns required for training is 40,000 and 6,000, respectively, which is a 6.7-fold reduction. It benefits from the model's efficient use of spatial modal features, which does not rely too much on large data sets. In addition, our method trains the model with modal coefficient vectors instead of large pixel-sized speckle patterns. The training and inference times are 493 minutes and 0.627 milliseconds for the data-driven model, and 0.5 minutes and 0.031 milliseconds for the modal embedding model, respectively, which are shortened by 986 times and 20 times, respectively. The reduction in training time and memory consumption makes the speckle wavelength measurement device of the present invention more suitable for rapid deployment in changing environments.

Claims

1. A speckle wavelength measurement method based on pattern embedding machine learning, characterized in that: The following steps are involved: S1, obtaining a speckle pattern of the light signal to be measured after passing through a fixed scattering element; S2. extracting a modal coefficient vector from the speckle pattern; S3. Inputting the modal coefficient vector into a pattern-embedded machine learning model to invert the wavelength of the optical signal to be measured; the pattern-embedded machine learning model is pre-trained by: collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting the corresponding modal coefficient vectors; and training a machine learning regression model using the obtained modal coefficient vectors and the corresponding wavelengths as input and output, respectively.

2. The speckle wavelength measurement method based on pattern embedding machine learning according to claim 1, characterized in that: The modal coefficient vectors are extracted from the speckle pattern using a mode decomposition method based on a beam propagation model.

3. The speckle wavelength measurement method based on pattern embedding machine learning according to claim 1, characterized in that: The fixed scattering element is an optical fiber containing no more than 15 linear polarization modes.

4. The speckle wavelength measurement method based on pattern embedding machine learning as claimed in claim 3, characterized in that: The length of the optical fiber is 5 m.

5. The speckle wavelength measurement method based on pattern embedding machine learning according to claim 1, characterized in that: When collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting corresponding modal coefficient vectors, the wavelength tuning rate of the series of optical signals of known wavelengths is 6.5 pm / s.

6. A speckle wavelength measurement device based on pattern embedding machine learning, characterized in that: include: A speckle pattern acquisition module is used to acquire the speckle pattern of the light signal to be measured after passing through the fixed scattering element; a mode decomposition module for extracting a modal coefficient vector from the speckle pattern; A mode embedding machine learning model is used to invert the wavelength of the optical signal to be measured using the modal coefficient vector as input. The mode embedding machine learning model is pre-trained by collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting the corresponding modal coefficient vectors; and training a machine learning regression model using the obtained modal coefficient vectors and the corresponding wavelengths as input and output, respectively.

7. The speckle wavelength measurement device based on pattern embedding machine learning according to claim 6, characterized in that: The mode decomposition module extracts the modal coefficient vectors from the speckle pattern using a mode decomposition method based on a beam propagation model.

8. The speckle wavelength measurement device based on pattern embedding machine learning according to claim 6, characterized in that: The fixed scattering element is an optical fiber containing no more than 15 linear polarization modes.

9. The speckle wavelength measurement device based on pattern embedding machine learning according to claim 8, characterized in that: The length of the optical fiber is 5 m.

10. The speckle wavelength measurement device based on pattern embedding machine learning according to claim 6, characterized in that: When collecting speckle patterns of a series of optical signals of known wavelengths after passing through the fixed scattering element and extracting corresponding modal coefficient vectors, the wavelength tuning rate of the series of optical signals of known wavelengths is 6.5 pm / s.

Images