Deep learning-based multimode fiber distributed temperature and position prediction method and system

Abstract

The invention discloses a deep learning-based multimode fiber distributed temperature and position prediction method and system. The method comprises the steps of constructing a convolutional neural network dual-task prediction model for predicting temperature and position based on a speckle image of a multimode fiber; speckle images, collected at different temperatures corresponding to different positions, of the multimode optical fiber are obtained; the collected speckle images are classified based on the temperature and the position, and a training set and a test set are constructed in proportion; training the constructed dual-task prediction model by using the training set to obtain a trained dual-task prediction model; and inputting the test set into the trained double-task prediction model, and outputting temperature and position prediction results. The dual-task prediction model is used for analyzing the speckle image of the multimode optical fiber so as to predict the temperature and the heating position of different positions along the optical fiber at the same time, and temperature information of a target area is remotely acquired through the optical fiber under the condition that a measured object does not need to be directly contacted by means of optical fiber transmission and speckle image acquisition.

Description

Technical Field

[0001] The present invention relates to the technical field of optical fibers, and particularly relates to a method and system for predicting the distributed temperature and position of multimode optical fibers based on deep learning. Background Art

[0002] With the rapid development of multimode fiber (MMF) sensing technology, fiber-based monitoring has become a research hotspot in fields such as industry, healthcare, and environmental monitoring, and is widely used in various scenarios such as bend detection, temperature monitoring, hydrodynamic measurement, steel bar corrosion monitoring, fire warning, and weak magnetic field detection. Traditional fiber optic temperature sensing technologies mainly rely on the reflection signals or temperature change characteristics of optical fibers, but these methods usually face problems such as high cost, complex installation, and insufficient accuracy, which limit their widespread application. In recent years, the combination of multimode optical fibers and deep learning technologies has gradually attracted attention. For example, researchers have used CNN technology to perform intensity classification detection on the speckle patterns of multimode optical fibers. High-fidelity imaging of bent multimode fibers (MMFs) is achieved through CNN, and the input-output relationship is learned in a 0.75m long MMF. There is also research that has proposed a method based on the short-time Fourier transform (STFT) and Resnet152 neural network for identifying multiple sensing event patterns (such as nine events including climbing, collision, cutting, etc.). There is also research that uses deep learning methods to interpret strain distributions in real time to detect spatially distributed cracks. However, none of these studies have involved temperature measurement. In terms of temperature measurement, there is research that uses CNN to measure the temperature of an earth-rock dam, but the fiber optic temperature measurement signal of a Raman optical time domain reflectometer (ROTDR) is used. In addition, deep learning has also been used in the distributed optical fiber temperature measurement system RDTS (Raman Distributed Temperature Sensing) to accurately identify small-scale abnormal temperature warnings. However, none of these methods use the speckle pattern as the basis for signal processing.

[0003] In recent years, with the rise of computer vision and deep learning technologies, fiber optic temperature and position prediction methods based on speckle imaging have gradually attracted attention. Multimode optical fibers, due to their support for multiple propagation modes, can generate complex speckle patterns that are extremely sensitive to environmental changes (such as temperature, pressure, and bending) of the optical fiber. Therefore, using the speckle patterns of multimode optical fibers for distributed sensing has great potential. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for predicting the distributed temperature and position of multimode optical fibers based on deep learning. By analyzing the speckle images of multimode optical fibers through a convolutional neural network (CNN) dual-task prediction model, high-precision temperature and position prediction can be achieved.

[0005] The present invention adopts the following technical solutions:

[0006] On the one hand, a multi-mode fiber distributed temperature and position prediction method based on deep learning includes:

[0007] Construct a convolutional neural network dual-task prediction model for predicting temperature and position from speckle images based on multi-mode fiber; the convolutional neural network dual-task prediction model includes a feature extraction module, a temperature prediction branch, and a position prediction branch; the feature extraction module is stacked by a number of sequentially connected convolutional layers, batch normalization layers, ReLU activation functions, and max pooling layers, and outputs a feature map corresponding to the speckle image; the temperature prediction branch reduces the feature map to a size of 1×1 through a first global average pooling layer, then flattens it into a one-dimensional vector through a first Flatten layer, and outputs to a first output layer after classification by a first fully connected layer, outputting the predicted temperature; the position prediction branch reduces the feature map to a size of 1×1 through a second global average pooling layer, then flattens it into a one-dimensional vector through a second Flatten layer, and outputs to a second output layer after classification by a second fully connected layer, outputting the predicted position; wherein, the number of nodes in the first output layer is the same as the number of temperature categories; the number of nodes in the second output layer is the same as the number of position categories;

[0008] Obtain the speckle images of the multi-mode fiber collected at different temperatures corresponding to different positions;

[0009] Classify the collected speckle images based on temperature and position, and construct a training set and a test set according to a ratio;

[0010] Use the training set to train the constructed convolutional neural network dual-task prediction model to obtain a trained convolutional neural network dual-task prediction model;

[0011] Input the test set into the trained convolutional neural network dual-task prediction model, and output the prediction results of temperature and position.

[0012] Preferably, before using the training set to train the constructed convolutional neural network dual-task prediction model, it further includes:

[0013] Adjust the size and normalize the collected speckle images.

[0014] Preferably, after inputting the test set into the trained convolutional neural network dual-task prediction model and outputting the prediction results of temperature and position, it further includes:

[0015] Evaluate the temperature prediction accuracy through the tolerance accuracy and the strict accuracy, and evaluate the position prediction accuracy through the confusion matrix.

[0016] Preferably, the convolutional kernel size of each convolutional layer is 3×3, the stride is 1, and the padding is 1, ensuring that the size of the feature map is not affected by the convolution operation.

[0017] Preferably, the convolution kernel size of each max-pooling layer is 2×2, and the stride is 2, so as to gradually reduce the spatial dimension of the feature map while expanding the receptive field.

[0018] Preferably, the speckle image of the multimode fiber is collected by a fiber speckle image acquisition device, and the fiber speckle image acquisition device includes: a laser, a multimode fiber, an industrial camera, a reflector, an objective lens, a fiber holder, and a neutral density filter; a fiber holder is placed at each end of the multimode fiber; the neutral density filter is placed at the front end of the industrial camera; the reflector and the objective lens couple the laser emitted by the laser into the multimode fiber; the multimode fiber has at least two groups of fibers with different cladding diameters and lengths; the speckle images of each group of fibers at different positions and temperatures are collected in sequence by the industrial camera.

[0019] On the other hand, a multimode fiber distributed temperature and position prediction system based on deep learning includes:

[0020] A convolutional neural network dual-task prediction model construction module for constructing a convolutional neural network dual-task prediction model for predicting temperature and position based on the speckle image of the multimode fiber; the convolutional neural network dual-task prediction model includes a feature extraction module, a temperature prediction branch, and a position prediction branch; the feature extraction module is stacked by several sequentially connected convolutional layers, batch normalization layers, ReLU activation functions, and max-pooling layers, and outputs a feature map corresponding to the speckle image; the temperature prediction branch reduces the feature map to a size of 1×1 through a first global average pooling layer, and then flattens it into a one-dimensional vector through a first Flatten layer, and outputs to a first output layer after classification by a first fully connected layer, outputting the predicted temperature; the position prediction branch reduces the feature map to a size of 1×1 through a second global average pooling layer, and then flattens it into a one-dimensional vector through a second Flatten layer, and outputs to a second output layer after classification by a second fully connected layer, outputting the predicted position; wherein, the number of nodes in the first output layer is the same as the number of temperature categories; the number of nodes in the second output layer is the same as the number of position categories;

[0021] A speckle image acquisition module for acquiring the speckle images of the multimode fiber collected at different temperatures corresponding to different positions;

[0022] A speckle image classification module for classifying the collected speckle images based on temperature and position, and constructing a training set and a test set according to a ratio;

[0023] A convolutional neural network dual-task prediction model training module for training the constructed convolutional neural network dual-task prediction model using the training set to obtain a trained convolutional neural network dual-task prediction model;

[0024] A prediction module, which is used to input a test set into a trained convolutional neural network dual-task prediction model and output prediction results of temperature and position.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] The present invention successfully realizes the prediction of the temperature and position of a fiber optic speckle image by using a convolutional neural network (CNN) dual-task prediction model. Under different experimental settings, the dual-task prediction model demonstrates stable and efficient prediction performance. Whether in the temperature prediction or position prediction task, a high accuracy rate can be achieved. Especially in the dual-task learning framework, the prediction accuracy of temperature and position remains very high. The experimental results show that the present invention can efficiently and accurately predict the temperature change and heating position of an optical fiber on optical fibers with different diameters, and has good practicability and popularization potential. Description of the Drawings

[0027] Figure 1 It is a flowchart of a multi-mode fiber distributed temperature and position prediction method based on deep learning according to an embodiment of the present invention;

[0028] Figure 2 It is a schematic structural diagram of a convolutional neural network dual-task prediction model according to an embodiment of the present invention;

[0029] Figure 3 It is a schematic experimental structure diagram of a fiber optic speckle image acquisition device according to an embodiment of the present invention;

[0030] Figure 4 It is a schematic diagram of a data set established in the first group of experiments according to an embodiment of the present invention;

[0031] Figure 5 It is a similarity matrix diagram of point B in the first group of experiments according to an embodiment of the present invention;

[0032] Figure 6 It is a prediction result diagram of temperature and position in the first group of experiments according to an embodiment of the present invention;

[0033] Figure 7 It is a temperature curve during dual-task simultaneous prediction in the first group of experiments according to an embodiment of the present invention;

[0034] Figure 8 It is a schematic diagram of six speckle images and their corresponding predicted temperatures and positions in the first group of experiments according to an embodiment of the present invention;

[0035] Figure 9 It is a temperature and position curve during dual-task simultaneous prediction in the second group of experiments according to an embodiment of the present invention;

[0036] Figure 10Six speckle patterns of the second group of experiments in the embodiments of the present invention and schematic diagrams of their corresponding predicted temperatures and positions;

[0037] Figure 11 The structural block diagram of a multimode fiber distributed temperature and position prediction system based on deep learning provided by the embodiments of the present invention. Specific embodiments

[0038] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of this application.

[0039] See Figure 1 As shown, a multimode fiber distributed temperature and position prediction method based on deep learning in this embodiment includes the following steps.

[0040] S101. Construct a convolutional neural network dual-task prediction model for predicting temperature and position based on speckle images of multimode fibers; the convolutional neural network dual-task prediction model includes a feature extraction module, a temperature prediction branch, and a position prediction branch; the feature extraction module is stacked by a number of sequentially connected convolutional layers, batch normalization layers, ReLU activation functions, and max pooling layers, and outputs a feature map corresponding to the speckle image; the temperature prediction branch reduces the feature map to a size of 1×1 through a first global average pooling layer, and then flattens it into a one-dimensional vector through a first Flatten layer, and outputs to a first output layer after classification by a first fully connected layer, outputting the predicted temperature; the position prediction branch reduces the feature map to a size of 1×1 through a second global average pooling layer, and then flattens it into a one-dimensional vector through a second Flatten layer, and outputs to a second output layer after classification by a second fully connected layer, outputting the predicted position; wherein, the number of nodes in the first output layer is the same as the number of temperature categories; the number of nodes in the second output layer is the same as the number of position categories.

[0041] See Figure 2 As shown, the convolutional neural network dual-task prediction model of this embodiment includes structures such as multiple convolutional layers, pooling layers, batch normalization, and fully connected layers. The feature extraction module uses multiple layers of convolution, and the number of channels is gradually increased layer by layer, from 1 channel to 2048 channels, to effectively extract complex features in the speckle image.

[0042] The model adopts a dual-task learning framework to predict temperature and position respectively. Under this framework, the feature extraction module is shared, which not only improves the performance of the model on the two tasks but also reduces the computational cost. Specifically, the temperature prediction task outputs through a classifier, such as with 80 classes (corresponding to 80 temperature levels); the position prediction task varies according to the settings of different experimental points. For example, in the first group of experiments, 3 position classes are output, and in the second group of experiments, 5 position classes are output.

[0043] In this model, the feature extraction module is stacked by several convolutional layers, batch normalization layers, ReLU activation functions, and max pooling layers. After each convolutional layer, a batch normalization layer and a ReLU activation function are followed to improve the network's non-linear expression ability and training stability. The convolutional kernel size is 3×3, the stride is 1, and the padding is 1 to ensure that the size of the feature map is not affected by the convolution operation. The pooling operation uses 2×2 max pooling with a stride of 2 to gradually reduce the spatial dimension of the feature map while expanding the receptive field.

[0044] The temperature prediction branch reduces the feature map to a size of 1×1 through global average pooling (AdaptiveAvgPool2d), then flattens it into a one-dimensional vector and inputs it into a fully connected layer for classification. The number of nodes in the output layer is the same as the number of temperature classes, such as set to 80 temperature levels for temperature prediction.

[0045] The position prediction branch has a similar structure to the temperature prediction branch and makes predictions through global average pooling and fully connected layers. The number of output nodes of this branch is adjusted according to the experimental design. For example, it is 3 position classes in the first group of experiments and 5 position classes in the second group of experiments.

[0046] S102. Obtain the speckle images of multimode optical fibers collected at different temperatures corresponding to different positions.

[0047] In this embodiment, a gas helium-neon laser is used as the laser source, and the laser is coupled into the optical fiber through a mirror and an objective lens. The optical fiber is in the form of a bare fiber and a high-temperature-resistant material is selected. Its outer layer is coated with polyimide resin, which can withstand high-temperature environments. To ensure that the temperature changes outside the heating area are not interfered by the heating zone, in the experiments of this embodiment, an FR-4 glass fiber board with good heat insulation performance is used to insulate the optical fiber in the non-heating area. To ensure that the heating point heats up quickly and evenly, taking advantage of the excellent thermal conductivity of brass, two 8-mm-wide brass blocks are placed at both ends of the optical fiber to clamp the optical fiber to heat the specified sampling points.

[0048] During the experiment, both ends of the optical fiber were fixed by clamping plates to eliminate the interference of mechanical vibration on the speckle image, ensuring that the image change was only caused by temperature or position change, thus guaranteeing the stability of image acquisition. The speckle image was captured by an industrial camera (CCD), which was equipped with a neutral density attenuation filter to prevent damage to the CCD caused by excessive light intensity. The exposure time was set to 8000 ms to ensure the clarity and contrast of the speckles, and thus effectively capture their characteristic changes. For the specific optical fiber speckle image acquisition device, refer to Figure 3 as shown. Among them, Figure 3 (a) shows the configuration of the acquisition device under the condition of turning on the light, Figure 3 (b) is the experimental device diagram for acquiring speckles under the condition of turning off the light, Figure 3 (c) is the top view schematic diagram of the acquisition device. In the experimental grouping, the first group of experimental devices was marked with a dashed box of "Group 1", while for the second group of experiments, the dashed box of "Group 1" was replaced with "Group 2", and the rest of the devices remained unchanged, so as to distinguish the specific configurations of the two groups of experiments.

[0049] The Laser used in the experiment was a 633 nm He-Ne laser; M1 and M2 were plane mirrors; OBJ was an objective lens with a magnification of 20 and a numerical aperture (NA) of 0.4; C1 and C2 were fiber clamps (Fiber Clip); ND was a neutral density filter (Neutral Density Filter); the industrial camera was of the GT2050 model with a GigE interface.

[0050] In this embodiment, the experiment was divided into two groups. The first group used a bare fiber with a cladding diameter of 600 μm and a total fiber length of 65 cm. In this group of experiments, a point (Point B) in the middle of the fiber was selected for heating, and heating points were set 5 cm to the left (Point A) and 6 cm to the right (Point C) of Point B respectively. The second group used a bare fiber with a cladding diameter of 400 μm and a total fiber length of 100 cm. The sampling points included five points D, E, F, G, and H, and the intervals between them were 1 cm, 2 cm, 4 cm, and 5 cm respectively.

[0051] During the experiment, speckle images were acquired for each sampling point. For example, at Point A, first, 10 speckle images were continuously acquired using the camera driver of MATLAB at room temperature of 20 °C, then another 10 were acquired at 21 °C, and so on. Acquisition was carried out every 1 °C until 99 °C. 10 speckle images were acquired at each temperature, and a total of 80 temperature points were acquired, finally obtaining 800 speckle images. Similarly, at Point B, Point C, Point D, Point E, Point F, Point G, and Point H, the corresponding speckle images were also acquired in the temperature range from 20 °C to 99 °C.

[0052] S103. Classify the collected speckle images based on temperature and position, and construct a training set and a test set proportionally.

[0053] When constructing the data set, among the 10 speckle images at each temperature point, 8 are used for the training set and 2 are used for the test set. For the first group of experiments (optical fiber with a cladding diameter of 600 μm), the data set includes three points A, B, and C. Among them, the training set contains 1920 speckle images, the test set contains 480 speckle images, and a total of 2400 speckle images. The second group of experiments (optical fiber with a cladding diameter of 400 μm) includes five points D, E, F, G, and H. Among them, the training set contains 3200 speckle images, the test set contains 800 speckle images, and a total of 4000 speckle images. See the specific data set structure in Figure 4 as shown.

[0054] To explore the differences between data sets, similarity analysis was performed on 10 speckle images at the same temperature (20 °C, point B), and a similarity matrix diagram was plotted. The similarity was calculated using the Root Mean Squared Error (RMSE) formula, and the formula is as follows:

[0055]

[0056] where x i and y i respectively represent the gray values of the two images at the i-th pixel point, and n is the total number of pixel points in the image. Through this formula, the difference degree between images can be quantified.

[0057] Among them, Figure 5 (a) shows the similarity matrix diagram plotted based on the analysis of 8 training images under the condition of 20 °C. Figure 5 (b) randomly selects 8 temperature points from the training set (a total of 80 groups of temperature data), compares and analyzes the corresponding speckle images, and then plots the similarity matrix diagram.

[0058] Before calculating the RMSE, the speckle images are normalized. The maximum RMSE value of 1 represents the difference between pure black and pure white images. Deeper colors in the matrix indicate smaller differences, while lighter colors indicate larger differences.

[0059] Figure 5(a) shows the similarity matrix of the images captured at the same temperature. The maximum difference at the same temperature is 0.008, indicating a high similarity between the images captured at the same temperature. 5(b) also compares the images taken at different temperatures. Compared with the values at the same temperature, the similarity differences between different temperatures are more significant. The greater the temperature difference, the greater the similarity difference value. However, the maximum value of 0.16 is still relatively small, indicating that the differences between the speckle images in the dataset are not significant.

[0060] Although the speckle images may visually appear similar, the differences between the images can still be identified by a specially designed algorithm at the pixel level. Specifically, the gray values and texture features of the speckle images vary slightly, reflecting different temperature or position information. Therefore, deep learning methods, especially convolutional neural networks (CNNs), can be used to learn these subtle differences and provide highly accurate predictions. The analysis of the similarity matrix further validates the feasibility of using deep learning methods for temperature and position prediction.

[0061] S104, Use the training set to train the constructed convolutional neural network dual-task prediction model to obtain a trained convolutional neural network dual-task prediction model.

[0062] In this embodiment, before using the training set to train the constructed convolutional neural network dual-task prediction model, it also includes data preprocessing of the collected speckle images.

[0063] Specifically, first, all input images are subjected to size unification and normalization processing. The 2048×2048-sized images collected by the CCD are reshaped into 256×256-sized images to ensure that the images can meet the input requirements of the neural network. All images are adjusted to the same size, and their gray values are normalized to the range [0,1] to improve the training efficiency and convergence speed of the model.

[0064] In this embodiment, the learning rate can be set to 0.00002 during the training process, and the batch size is 4, which can be specifically optimized according to the training process.

[0065] S105, Input the test set into the trained convolutional neural network dual-task prediction model and output the prediction results of temperature and position.

[0066] The prediction results of the above first and second groups of experiments will be evaluated as follows. Specifically, the temperature prediction accuracy is evaluated by the tolerance accuracy and the strict accuracy, and the position prediction accuracy is evaluated by the confusion matrix.

[0067] In this embodiment, the experimental results of the first group and the second group are both presented in three steps. First, the temperature is predicted separately; second, the position is predicted separately; finally, the predictions of the temperature and the position are combined to achieve the simultaneous prediction of the dual tasks.

[0068] The experimental results of the first group will be evaluated as follows.

[0069] In the case of only predicting the temperature, the temperature prediction curves are respectively plotted for the prediction results at three points (Point A, Point B, and Point C). In the temperature prediction task, two different accuracy evaluation methods are adopted: strict accuracy and tolerance accuracy. Strict accuracy requires that the predicted value is exactly the same as the actual value. For example, when the actual temperature is 98 °C, only the prediction result of 98 °C is considered a correct prediction. Tolerance accuracy, on the other hand, considers the prediction correct within a certain range. For example, if the actual temperature is 98 °C, then the predicted temperature of 97 °C or 99 °C is considered a correct prediction.

[0070] Figure 6 The results of the three figures (a), (b), and (c) show that the tolerance accuracies at the three points are 100%, 100%, and 100% respectively; the strict accuracies are 98.12%, 100%, and 96.25% respectively. These results indicate that when predicting the temperature, the model can relatively accurately predict the temperature change.

[0071] In the case of only predicting the position, a confusion matrix diagram is plotted based on the prediction results, as Figure 6 shown in (d). The results show that the prediction position accuracies at Point A, Point B, and Point C are all 100%. Specifically, for the test sets at the three positions A, B, and C in the figure, the positions of the 160 speckle patterns included in each are accurately predicted. This indicates that in the position prediction task, the model shows perfect classification ability and can accurately identify the position of each point.

[0072] In the task of simultaneously predicting the temperature and the position, the result of the temperature prediction is as Figure 7 shown. The average value of the same temperature prediction correct rates at the three points (Point A, Point B, and Point C) is calculated. The result shows that the tolerance accuracy is 100% and the strict accuracy is 97.92%. At the same time, the result of the position prediction is the same as that of Figure 6 (d), and the accuracy is still 100%.

[0073] Such prediction data indicates that by inputting any speckle pattern in the test set into the deep learning model, the accurate prediction of its temperature and position can be achieved. Figure 8 Six speckle patterns and their corresponding predicted temperatures and positions are shown, and the actual temperature values (51 °C and 80 °C respectively) at the three position points A, B, and C are marked. The results show that the temperatures and positions predicted by the model are highly consistent with the actual ones.

[0074] The experimental results of the second group will be evaluated as follows.

[0075] In the second group of experiments, the temperature prediction task involved five different points (Point D, Point E, Point F, Point G, and Point H). By plotting the temperature prediction curve, the tolerance accuracy and strict accuracy of each point were obtained. The specific results are shown in Table 1 below.

[0076] Table 1 Temperature Prediction Accuracy of Each Point in the Second Group of Experiments

[0077]

[0078] It can be seen from the results that as the fiber length increases, the cladding diameter changes, and the number of modes in the multimode fiber decreases, the speckle features that the model can extract decrease, resulting in a slight decrease in the temperature prediction accuracy compared to the first group of experiments. The decrease in the cladding diameter will lead to a decrease in the number of propagation modes in the fiber because the number of modes is closely related to factors such as the core and cladding diameters of the fiber and the wavelength. As the number of modes decreases, the extractable speckle features also decrease, thereby affecting the accuracy of temperature prediction. Especially at Point H, the accuracy of the prediction result is lower than that of other points. Nevertheless, the model can still maintain a high prediction accuracy at different points, especially within the temperature tolerance accuracy range, and the prediction results show a high accuracy. When only predicting the position, the confusion matrix diagram was also plotted, and the results show that the prediction position accuracy of Point D, Point E, Point F, Point G, and Point H is all 100%. This indicates that the model still performs excellently in the multi-point position prediction task and can accurately predict the position of each point.

[0079] In the task of simultaneously predicting temperature and position, the average value of the correct temperature prediction rates of the five points (Point D, Point E, Point F, Point G, and Point H) was calculated, as Figure 9 (a), the results show that the tolerance accuracy is 95.12% and the strict accuracy is 88.12%, but after 70 °C, the deviation of the temperature prediction increases. At the same time, the confusion matrix of the predicted position Figure 9 (b) shows that the prediction position accuracy of all points is still 100%.

[0080] Similarly, the above prediction data indicate that in the second group of experiments, by inputting any speckle pattern in the test set into the deep learning model, the temperature and position can also be accurately predicted. Next Figure 10 Six speckle patterns and their corresponding predicted temperatures and positions are shown, and the actual temperature values (51 °C and 80 °C respectively) of the five position points D, E, F, G, and H are marked. The results show that except that 51 °C at Point H is predicted as 52 °C (still within a reasonable error range), the remaining predicted temperatures and positions are consistent with the actual ones.

[0081] Through the comparative analysis of the experimental results of the first group and the second group, the following conclusions can be drawn: regardless of the fiber diameter (400μm and 600μm) or different fiber lengths (65cm and 100cm), the convolutional neural network dual-task prediction model of this embodiment can accurately predict temperature and position. This indicates that the distributed temperature and position prediction based on multimode fiber can maintain good prediction effects under different experimental conditions.

[0082] Under the conditions of different fiber diameters and different distances between measurement points, the temperature prediction accuracy of the model is not significantly affected. Especially in the case of a 400μm fiber diameter, the model can accurately distinguish and predict points separated by 1cm, demonstrating its robustness under different experimental conditions. Although the features extractable from the speckle image decrease with the decrease of the fiber diameter, which leads to a slight decrease in the prediction accuracy, the overall accuracy still remains at a very high level.

[0083] In the task of simultaneously predicting temperature and position, both the tolerance accuracy and the strict accuracy show relatively high levels, indicating that the multi-task learning has a good synergistic effect in temperature and position prediction. Especially in the temperature prediction task, the tolerance accuracy is close to 100%, and the strict accuracy also remains at a relatively high level, showing that the convolutional neural network dual-task prediction model can effectively capture the distribution changes of the fiber temperature.

[0084] It should be noted that the above only takes two groups of experiments as examples to predict the temperature and the corresponding position. In actual application, only the actually collected speckle images need to be input into the trained model to complete the temperature and the corresponding position prediction.

[0085] Based on the above method, as shown in Figure 11 This invention also discloses a multimode fiber distributed temperature and position prediction system based on deep learning, including:

[0086] The convolutional neural network dual-task prediction model construction module 1101 is used to construct a convolutional neural network dual-task prediction model for predicting temperature and position based on the speckle images of multimode optical fibers. The convolutional neural network dual-task prediction model includes a feature extraction module, a temperature prediction branch, and a position prediction branch. The feature extraction module is stacked by a number of sequentially connected convolutional layers, batch normalization layers, ReLU activation functions, and max pooling layers, and outputs a feature map corresponding to the speckle image. The temperature prediction branch reduces the feature map to a size of 1×1 through a first global average pooling layer, then flattens it into a one-dimensional vector through a first Flatten layer, and outputs the predicted temperature to a first output layer after classification by a first fully connected layer. The position prediction branch reduces the feature map to a size of 1×1 through a second global average pooling layer, then flattens it into a one-dimensional vector through a second Flatten layer, and outputs the predicted position to a second output layer after classification by a second fully connected layer. Among them, the number of nodes in the first output layer is the same as the number of temperature categories; the number of nodes in the second output layer is the same as the number of position categories.

[0087] The speckle image acquisition module 1102 is used to acquire the speckle images of multimode optical fibers collected at different temperatures corresponding to different positions.

[0088] The speckle image classification module 1103 is used to classify the collected speckle images based on temperature and position, and construct a training set and a test set according to a ratio.

[0089] The convolutional neural network dual-task prediction model training module 1104 is used to train the constructed convolutional neural network dual-task prediction model using the training set to obtain a trained convolutional neural network dual-task prediction model.

[0090] The prediction module 1105 is used to input the test set into the trained convolutional neural network dual-task prediction model and output the prediction results of temperature and position.

[0091] The specific implementation of each module of a multimode optical fiber distributed temperature and position prediction system based on deep learning is the same as that of a multimode optical fiber distributed temperature and position prediction method based on deep learning, and will not be repeated in this embodiment.

[0092] Although the specific implementation manners of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are illustrative rather than used to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered by the scope protected by the claims of the present invention.

Claims

1. A method for predicting multi-mode fiber distributed temperature and position based on deep learning, characterized in that, Including: Construct a convolutional neural network dual-task prediction model for predicting temperature and position based on speckle images of multimode optical fibers; the convolutional neural network dual-task prediction model includes a feature extraction module, a temperature prediction branch, and a position prediction branch; the feature extraction module is stacked by a number of sequentially connected convolutional layers, batch normalization layers, ReLU activation functions, and max pooling layers, and outputs a feature map corresponding to the speckle image; the temperature prediction branch reduces the feature map to a size of 1×1 through a first global average pooling layer, and then flattens it into a one-dimensional vector through a first Flatten layer, and after classification by a first fully connected layer, it outputs to a first output layer to output the predicted temperature; the position prediction branch reduces the feature map to a size of 1×1 through a second global average pooling layer, and then flattens it into a one-dimensional vector through a second Flatten layer, and after classification by a second fully connected layer, it outputs to a second output layer to output the predicted position; wherein, the number of nodes in the first output layer is the same as the number of temperature categories; the number of nodes in the second output layer is the same as the number of position categories; Obtain the speckle images of multimode optical fibers collected at different temperatures corresponding to different positions. Classify the collected speckle images based on temperature and position, and construct a training set and a test set in proportion. Use the training set to train the constructed convolutional neural network dual-task prediction model to obtain a trained convolutional neural network dual-task prediction model. Input the test set into the trained convolutional neural network dual-task prediction model to output the predicted results of temperature and position.

2. The method for predicting the temperature and position of a multimode optical fiber distributed based on deep learning according to claim 1, wherein Before using the training set to train the constructed convolutional neural network dual-task prediction model, it further includes: Adjust the size and normalize the collected speckle images.

3. The multimode fiber distributed temperature and position prediction method based on deep learning according to claim 1, wherein After inputting the test set into the trained convolutional neural network dual-task prediction model and outputting the predicted results of temperature and position, it further includes: Evaluate the temperature prediction accuracy through the tolerance accuracy and the strict accuracy, and evaluate the position prediction accuracy through the confusion matrix.

4. The multimode fiber distributed temperature and position prediction method based on deep learning according to claim 1, characterized in that The convolution kernel size of each convolutional layer is 3×3, the stride is 1, and the padding is 1 to ensure that the size of the feature map is not affected by the convolution operation.

5. The multimode fiber distributed temperature and position prediction method based on deep learning according to claim 1, wherein The convolution kernel size of each max pooling layer is 2×2, and the stride is 2 to gradually reduce the spatial dimension of the feature map while expanding the receptive field.

6. The multimode fiber distributed temperature and position prediction method based on deep learning according to claim 1, characterized in that The speckle images of multimode optical fibers are collected by a fiber speckle image acquisition device, and the fiber speckle image acquisition device includes: a laser, a multimode optical fiber, an industrial camera, a reflector, an objective lens, a fiber holder, and a neutral density filter; a fiber holder is placed at each end of the multimode optical fiber; the neutral density filter is placed at the front end of the industrial camera; the reflector and the objective lens couple the laser emitted by the laser into the multimode optical fiber; the multimode optical fiber has at least two groups of optical fibers with different cladding diameters and lengths; the speckle images of each group of optical fibers at different positions and temperatures are collected in sequence by the industrial camera.

7. A multi-mode fiber distributed temperature and position prediction system based on deep learning, characterized in that, Including: A convolutional neural network dual-task prediction model construction module is used to construct a convolutional neural network dual-task prediction model for predicting temperature and position based on speckle images of multimode fibers. The convolutional neural network dual-task prediction model includes a feature extraction module, a temperature prediction branch, and a position prediction branch. The feature extraction module is stacked by several consecutive convolutional layers, batch normalization layers, ReLU activation functions, and max pooling layers, and outputs a feature map corresponding to the speckle image. The temperature prediction branch reduces the feature map to a size of 1×1 through a first global average pooling layer, then flattens it into a one-dimensional vector through a first Flatten layer, and outputs the predicted temperature to the first output layer after classification by a first fully connected layer. The position prediction branch reduces the feature map to a size of 1×1 through a second global average pooling layer, then flattens it into a one-dimensional vector through a second Flatten layer, and outputs the predicted position to the second output layer after classification by a second fully connected layer. Among them, the number of nodes in the first output layer is the same as the number of temperature categories; the number of nodes in the second output layer is the same as the number of position categories; A speckle image acquisition module is used to acquire speckle images of multimode fibers collected at different temperatures corresponding to different positions; A speckle image classification module is used to classify the collected speckle images based on temperature and position, and construct a training set and a test set according to a ratio; A convolutional neural network dual-task prediction model training module is used to train the constructed convolutional neural network dual-task prediction model using the training set to obtain a trained convolutional neural network dual-task prediction model; A prediction module is used to input the test set into the trained convolutional neural network dual-task prediction model and output the prediction results of temperature and position.

Images