Small sample optical fiber spectrum temperature prediction method based on deep learning

Abstract

The invention provides a small sample optical fiber spectrum temperature prediction method based on deep learning. Comprising the following steps: collecting optical fiber spectrum data covering a range of 30 DEG C to 130 DEG C, and totally obtaining 105 groups of high-quality data samples; an advanced signal processing technology is adopted to remove noise and abnormal values, and normalization processing is carried out on data; training and learning the spectral data by using multiple models of deep learning; and optimizing the structure and parameters of the model, and constructing a nonlinear mapping relation between the spectrum and the temperature. According to the invention, a novel effective method is provided for spectral temperature prediction under a small sample condition, and the application range of deep learning in the field of spectral analysis is expanded. The method has wide application prospects in the fields of industrial production, environmental monitoring, biomedicine and the like. The invention belongs to the technical field of temperature detection.

Description

(I) Technical field

[0001] The present invention relates to a small sample optical fiber spectral temperature prediction method based on deep learning, which provides a new and effective response measure for spectral temperature prediction under small sample conditions and expands the application scope of deep learning in the field of spectral analysis. The present invention belongs to the field of temperature detection technology. (II) Background technology

[0002] Fiber optic sensing technology, as a rapidly advancing field in the field of modern science and technology, has been widely used in many key fields such as medical diagnosis, environmental monitoring and signal transmission, demonstrating its excellent application potential and broad development prospects. Among them, fiber optic temperature sensing technology occupies a pivotal position in industrial applications and scientific research. Thanks to its excellent anti-electromagnetic interference performance, high precision and high sensitivity, easy integration and expansion, it has become an indispensable tool in scientific research and engineering practice. However, it also faces multiple challenges such as improving sensitivity, suppressing environmental interference, and managing complex data processing.

[0003] Patent CN 113030012A provides a spectral analysis method based on a multi-level partial least squares algorithm, which is designed to collect spectra of samples at multiple different temperatures, and filter pre-process the collected spectra; the spectra after filtering pre-processing are combined with a multi-level partial least squares algorithm to build an analysis model, and the analysis model is used to predict the spectral temperature and sample concentration to achieve spectral temperature correction. However, the sampling sample capacity is not sufficient in the process of collecting spectra, and it is more inclined to a single specific measurement, the accuracy and efficiency are not high enough, and when faced with abnormal data, the model may be unstable.

[0004] This paper proposes a method for predicting temperature of optical fiber spectrum based on deep learning for small samples. After extracting data from repeated samples, multiple models of deep learning are used to train and learn the spectrum data to construct a nonlinear mapping relationship between spectrum and temperature. This method is expected to improve the accuracy, stability and cost-effectiveness of optical fiber temperature sensors, while optimizing the structure and data processing levels. Compared with traditional methods, deep learning models show higher accuracy and stronger robustness in spectral temperature prediction. (III) Summary of the invention

[0005] The purpose of the present invention is to provide a small sample optical fiber spectral temperature prediction method based on deep learning, which provides a novel and effective technology for spectral temperature prediction under small sample conditions, and at the same time expands the application scope of deep learning in the field of spectral analysis.

[0006] The object of the present invention is achieved in that:

[0007] A method for predicting optical fiber spectrum temperature using a small sample size based on deep learning, comprising the following steps: S1. We collected fiber spectral data covering the range of 30°C to 130°C, and obtained a total of 105 sets of high-quality data samples; S2, using advanced signal processing techniques to remove noise and outliers and normalize the data; S3, using multiple deep learning models to train and learn spectral data; S4. Optimize the model structure and parameters to construct a nonlinear mapping relationship between spectrum and temperature.

[0008] In order to realize the above steps, the present invention is designed as follows:

[0009] The optical fiber sensor structure used for data acquisition uses a 3 mm single-mode optical fiber as the main part of the sensor structure. Multimode optical fibers with a length of 1 mm are welded on both sides of this single-mode optical fiber. Subsequently, a six-hole optical fiber with a length of 1 mm is welded again on the outside of the multimode optical fiber. The diameter of each circular hole of the six-hole optical fiber is about 7 μm, which is evenly distributed at a distance of about 12 μm from the midpoint of the optical fiber cross section. The distance between two adjacent circular holes is about 10 μm. In addition, the center of the six-hole optical fiber has a single-core optical fiber positioned along its axis, and its diameter is about 10 μm.

[0010] The high-precision spectrometer ICP-6800 was used to collect data from the optical fiber sensor. The collection range of each spectral data covered the spectral range from 1390 nanometers to 1530 nanometers. The data collection density was set to collect data every 0.01 nanometer, so that detailed spectral data containing 14,000 data points were obtained at each temperature point.

[0011] The spectral data collection interval of the spectrometer ICP-6800 is within the clear temperature range of 30 to 130 degrees. The sensor structure is heated by a heating device, and the spectral data is collected every 5 degrees. Each group collects spectral data at 21 different temperature points, such as Figure 2 Repeat the above data collection process five times to ensure the persuasiveness of the data, and finally obtain a total of 105 sets of spectral data, as shown in Figure 3 As shown;

[0012] The 105 sets of spectral data were integrated. In the data processing stage, the collected spectral data were first cleaned up by removing noise, processing outliers, and filling missing values. Then the data were normalized using the minimum-maximum standardization method to ensure the comparability of the data.

[0013] The processed data is marked with the corresponding temperature label, and the data set is divided into a training set and a test set in a ratio of 8:2 to ensure the effectiveness and accuracy of the model training, tuning and evaluation process, so as to carry out model training and verification work in the future.

[0014] Three deep learning models, namely convolutional neural network (CNN), recurrent neural network (RNN) and deep neural network (DNN), were selected for comparative analysis and comparison to intuitively demonstrate the comparative performance of different deep learning models and their potential in improving the accuracy and stability of temperature prediction.

[0015] During the model training process, the Adam optimizer is used to adaptively adjust the learning rate according to the gradient of the loss function to optimize the training process;

[0016] The batch size in the training setting is set to 16 and the number of iterations is set to 100 to ensure that the model has enough learning cycles to converge, thereby balancing the computational efficiency and the model's adaptability to data changes. The learning rate is adjusted to 0.001 to find the best balance between training speed and convergence stability. The change in loss value is shown in the figure below. Figure 4 shown.

[0017] The prediction accuracy, generalization ability and possible overfitting risk of each deep learning model are further verified on the test set to more comprehensively evaluate the model performance. The mean absolute error is selected as the loss function to effectively measure the difference between the model prediction value and the true value, and to evaluate the regression problem. (IV) Description of the drawings

[0018] Figure 1 It is a flow chart of the present invention.

[0019] Figure 2 This is a data acquisition experiment diagram for optical fiber spectrum, where 1 is the spectrum analyzer, 2 is the optical fiber heating device, and 3 is the light source.

[0020] Figure 3 This is a graph showing the spectral data after collection.

[0021] Figure 4 Schematic diagram of the change in loss function value during model training.

[0022] Figure 5 This is a comparison chart of the loss values ​​of the overall prediction results. (V) Specific implementation methods

[0023] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the method of the present invention is explained in detail below in conjunction with the accompanying drawings. It should be understood that the specific examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0024] The purpose of the present invention is to provide a small sample optical fiber spectral temperature prediction method based on deep learning, which provides a novel and effective technology for spectral temperature prediction under small sample conditions, and can effectively solve the problem of limited prediction accuracy faced by optical fiber spectroscopy technology in temperature prediction.

[0025] Reference Figure 1 , which shows the implementation process of the present invention, and provides a small sample optical fiber spectral temperature prediction method based on deep learning, including the following steps: S1. We collected fiber spectral data covering the range of 30°C to 130°C, and obtained a total of 105 sets of high-quality data samples; S2, using advanced signal processing techniques to remove noise and outliers and normalize the data; S3, using multiple deep learning models to train and learn spectral data; S4. Optimize the model structure and parameters to construct a nonlinear mapping relationship between spectrum and temperature.

[0026] First, in terms of the design of the sensor composite structure, a single-mode optical fiber with a length of 3 mm is used as the main part of the sensor. Its excellent optical properties ensure the high fidelity of the optical signal during transmission. On both sides of this single-mode optical fiber, a multimode optical fiber with a length of 1 mm is welded, and then a six-hole optical fiber with a length of 1 mm is welded again on the outside of the multimode optical fiber. The diameter of each circular hole of the six-hole optical fiber is about 7μm, evenly distributed at a distance of about 12μm from the midpoint of the optical fiber cross section, and the distance between two adjacent circular holes is about 10μm. In addition, the center of the six-hole optical fiber has a single-core optical fiber positioned along its axis, and its diameter is about 10μm. Such a composite structure ensures close contact between optical fibers, thereby optimizing the transmission efficiency and stability of optical signals, fully considering the physical properties and optical principles of optical fiber materials, and laying a solid foundation for subsequent data acquisition and temperature prediction.

[0027] In a laboratory environment, by systematically changing the temperature conditions, spectral data at different temperatures are collected. In step 1, a high-precision spectrometer ICP-6800 is used for data collection to ensure the accuracy and reliability of data collection at each temperature point. The sensor structure is heated using a heating device, and the temperature range is set to 30 to 130 degrees. Spectral data is collected every 5 degrees, and a total of spectral data at 21 different temperature points are collected. The collection range of each spectral data covers the spectral range from 1390 nanometers to 1530 nanometers, and the density of data collection is set to collect every 0.01 nanometers, thereby obtaining detailed spectral data containing 14,000 data points at each temperature point. To ensure the stability and reliability of the experimental results, the above data collection process was repeated five times, and a total of 105 sets of spectral data were finally obtained.

[0028] In the data processing process, the first step is to comprehensively process the collected spectral data, including noise elimination, correction of outliers, and filling of missing values. Then, in order to ensure the comparability of the data, step 2 uses the minimum-maximum standardization method to normalize the data. After the above processing, the data is assigned corresponding temperature labels, and the data set is divided into training sets and test sets for use in model training and verification in subsequent stages.

[0029] Three representative deep learning models, namely convolutional neural network, recurrent neural network and deep neural network, were selected to extract the features of spectral data from different angles.

[0030] During the model training process, the Adam optimizer is selected, which can adaptively adjust the learning rate according to the gradient of the loss function to optimize the training process.

[0031] In order to balance the computational efficiency and the adaptability of the model to data changes, the batch size in the training setting is set to 16 and the number of iterations is set to 100 to ensure that the model has enough learning cycles to converge, thereby balancing the computational efficiency and the adaptability of the model to data changes; the learning rate is adjusted to 0.001 to find the best balance between training speed and convergence stability. The change in loss value is shown in the figure below. Figure 4 shown.

[0032] Figure 4 The change of loss function during model training is fed back. The figure shows that the loss function values ​​of each model gradually decrease and stabilize after about 80 iterations. The final loss value is maintained at about 0.06. This shows that all three models can learn effective feature representations on a given dataset and achieve similar prediction accuracy. In particular, the convolutional neural network model Figure 4 It shows faster convergence speed and lower final loss value, which reflects its high efficiency in extracting features of spectral data.

[0033] Next, the prediction accuracy, generalization ability and possible overfitting risk of each model were further verified on the test set.

[0034] The data set is divided into a training set and a test set in a ratio of 8:2 to ensure the effectiveness and accuracy of the model training, tuning, and evaluation process. The mean absolute error (MAE) is selected as the loss function because it can effectively measure the difference between the model prediction value and the true value, focus on describing the average degree of prediction error and the performance of the model on each prediction value, and can more intuitively reflect the average level of prediction error. It is suitable for the evaluation of regression problems. The calculation formula of MAE is as follows:

[0035] Among them, n is the total amount, y i is the true value of the i-th sample, is the model's predicted value for the i-th sample.

[0036] In order to reduce overfitting, the L2 regularization term is introduced to penalize the L2 norm of the model parameters to avoid the model from being overly dependent on the training data and improve the generalization ability of the model. L2 regularization constrains the complexity of the model by adding the L2 norm penalty term of the model parameters to the loss function. To add the L2 regularization term to the MAE loss function, we need to first define the L2 norm of the model parameters. Assume that the model parameter set is θ = Pw 1 ,w 2 ... m}, where m is the sum of parameters (weights of all layers in the neural network). The L2 norm formula is as follows:

[0037] Among them, λ is the regularization strength, which is used to control the influence of the regularization term on the total loss. The optimal parameter obtained by grid search is 0.5, w j is the jth parameter (weight or bias) in the model. The mechanism of L2 regularization is that it encourages model parameters to tend to zero but not completely zero, which helps to reduce the complexity of the model and prevent the model from learning noise or special patterns in the training data. At the same time, L2 regularization can also improve the stability of the model, so that the model can make more reasonable predictions when facing new and unseen data. In order to specifically demonstrate the advantages of deep learning models in spectral temperature prediction, we conducted experiments. Plotted Figure 5 , which is a comparison chart of the loss values ​​of the overall prediction results.

[0038] like Figure 5As shown in the figure, the prediction errors of three deep learning models (CNN, RNN, DNN) and the original method on the test set are compared. The results show that the deep learning model exhibits significantly lower prediction errors, highlighting its superiority in spectral temperature prediction. Compared with traditional methods, deep learning technology effectively captures and utilizes complex features in spectral data through its powerful feature extraction and nonlinear modeling capabilities, thereby significantly improving the accuracy and stability of temperature prediction. In addition, deep learning also shows advantages in small sample conditions and is more adaptable, which provides important technical support for practical applications.

Claims

1. A method for predicting optical fiber spectrum temperature based on small sample size based on deep learning, characterized in that: The following steps are involved: S1. We collected fiber spectral data covering the range of 30°C to 130°C, and obtained a total of 105 sets of high-quality data samples; S2, using advanced signal processing techniques to remove noise and outliers and normalize the data; S3, using multiple deep learning models to train and learn spectral data; S4. Optimize the model structure and parameters to construct a nonlinear mapping relationship between spectrum and temperature.

2. According to the method for predicting temperature of small sample optical fiber spectrum based on deep learning in claim 1, it is characterized by: In step S1, the structure used uses a 3 mm single-mode optical fiber as the main part of the sensor structure. Multi-mode optical fibers with a length of 1 mm are welded on both sides of the single-mode optical fiber, and then a six-hole optical fiber with a length of 1 mm is welded again on the outside of the multi-mode optical fiber to form a composite optical fiber structure.

3. The method for predicting temperature of optical fiber spectrum using a small sample based on deep learning according to claim 1, characterized in that: In step S2, a high-precision spectrometer ICP-6800 is used for data acquisition. The acquisition range of each spectral data covers the spectral interval from 1390 nanometers to 1530 nanometers. The density of data acquisition is set to collect once every 0.01 nanometer, thereby obtaining detailed spectral data containing 14,000 data points at each temperature point.

4. The method for predicting temperature of optical fiber spectrum using a small sample based on deep learning according to claim 1, characterized in that: In step S2, the spectral data are collected by the spectrometer ICP-6800 within a clear temperature range of 30 to 130 degrees, and spectral data are collected every 5 degrees. A total of 21 spectral data at different temperature points are collected for each group. The above data collection process is repeated five times, and a total of 105 groups of spectral data are finally obtained.

5. The method for predicting temperature of optical fiber spectrum using a small sample based on deep learning according to claim 1, characterized in that: In step S3, three representative deep learning models, namely, convolutional neural network (CNN), recurrent neural network (RNN) and deep neural network (DNN), are selected for comparative analysis.

6. The method for predicting temperature of optical fiber spectrum using a small sample based on deep learning according to claim 1, characterized in that: In step S3, the Adam optimizer was selected in the model training process, the batch size was set to 16, the number of iterations was set to 100, and the learning rate was adjusted to 0.001.

Images