A method and system for Brillouin hetero-profile spectrum recognition based on machine learning

Abstract

The application discloses a Brillouin heteromorphic spectrum recognition method and system based on machine learning, relates to the field of Brillouin sensing data demodulation technology, and is used for accurately recognizing Brillouin heteromorphic spectrum. Technical points of the application include: simulating to generate pictures containing multiple types of Brillouin gain spectrum to form a training set; inputting the training set into a classification model based on machine learning for training to obtain a trained classification model; inputting an unknown type of Brillouin gain spectrum into the trained classification model to obtain the type of the Brillouin gain spectrum. The application proposes that a convolutional neural network algorithm in a deep learning framework is applied to optimize data processing of a distributed optical fiber sensing system, and the deep learning framework is used for heteromorphic spectrum line image classification and recognition of the distributed optical fiber sensing system, so that the application has very important practical significance for improving the spatial resolution, data processing speed and accuracy of the distributed optical fiber sensing system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Brillouin data demodulation, and particularly relates to a Brillouin heteromorphic spectrum recognition method and system based on machine learning. BACKGROUND

[0002] Distributed Brillouin fiber sensing technology has important application value in the fields of large infrastructure structure health monitoring, geophysical research, etc. due to its advantages of continuous non-blind area, large number of sensing points, long measurement distance, etc. For example, for monitoring activities of large buildings, many positions usually need to be monitored, at this time, if point sensors are used, the construction cost will be greatly increased due to the need to set a large number of monitoring points, and with the increase of monitoring points, the complexity of the whole system will also be greatly increased, which brings many inconveniences to the specific implementation. The use of optical fiber sensing can greatly reduce the complexity of the whole monitoring system and the construction cost of the monitoring system. At the same time, the monitoring objects are mainly but not limited to tunnel fire monitoring, pipeline temperature monitoring, dam damage and aging monitoring, in these specific application scenarios, the temperature and strain distribution of the monitored object can be monitored in real time and at high speed through the optical fiber sensing system, and the abnormal temperature or abnormal strain is collected to the data processing system at the first time when the monitored object has an abnormal condition, so that various abnormal conditions can be effectively handled. However, due to the limitation of sensing mechanism, the traditional distributed Brillouin fiber sensing technology has the problem of low spatial resolution, which seriously restricts its application field and practical value. Therefore, how to improve the spatial resolution is a research difficulty and hot issue of distributed Brillouin sensing technology.

[0003] In a Brillouin optical time domain reflectometer or Brillouin optical time domain analysis system, it is usually necessary to obtain a Brillouin gain spectrum (BGS) and obtain a Brillouin frequency shift (BFS) therefrom, the temperature and stress changes in the external environment have a certain linear relationship with the Brillouin frequency shift, so the temperature or stress change in the external environment can be indirectly calculated by demodulating the Brillouin frequency shift. In the past BOTDA or BOTDR system, since the line type of the Brillouin gain spectrum is Lorentz line type, the demodulation method usually adopted is Lorentz curve fitting (LCF) fitting, and the frequency corresponding to the peak value of the fitting result is the Brillouin frequency shift. However, the Lorentz fitting method is seriously affected by the initial value, and the selection of the initial value directly affects the accuracy of the fitting result. Moreover, in the case of low signal-to-noise ratio, the number of iterations of the fitting will also increase, which will greatly prolong the time used for curve fitting, which is not conducive to real-time monitoring. SUMMARY

[0004] In view of the above problems, the present invention proposes a Brillouin shaped spectrum recognition method and system based on machine learning to accurately identify Brillouin shaped spectra, thereby improving the spatial resolution of distributed optical fiber sensing systems.

[0005] According to one aspect of the present invention, a method for identifying Brillouin profiles based on machine learning is proposed, which is characterized by comprising the following steps:

[0006] Step 1: Obtain images containing multiple types of Brillouin gain spectra to form a training set;

[0007] Step 2: Input the training set into the classification model based on machine learning for training to obtain the trained classification model;

[0008] Step 3: Input the unknown type of Brillouin gain spectrum into the trained classification model to obtain the type of the Brillouin gain spectrum.

[0009] Furthermore, the multiple types of Brillouin gain spectra in step 1 include Brillouin singlet spectra, Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra.

[0010] Furthermore, in step 1, images of various types of Brillouin gain spectra are generated by simulation; wherein the Brillouin single peak spectrum is the ideal Brillouin gain spectrum plus random Gaussian white noise, and the generation formula is as follows:

[0011]

[0012] Where g(ν) is the ideal Brillouin gain, n(ν) is the random Gaussian white noise, and g B is the peak value of the Brillouin gain spectrum, ν is the frequency independent variable, i.e. the sweep frequency range, B is the Brillouin frequency shift corresponding to the spectral line, Δν B is the spectral line width; random(N) is the standard normal distribution with length N, N is the number of sweeps, SNR is the signal-to-noise ratio; g i (v) represents the Brillouin gain value corresponding to the i-th frequency sweep point;

[0013] Generate multiple Brillouin single peak spectrum images corresponding to two signal-to-noise ratios: high signal-to-noise ratio 100dB and low signal-to-noise ratio 15dB.

[0014] Furthermore, the Brillouin-like singlet spectrum, Brillouin short-range doublet spectrum, and Brillouin long-range doublet spectrum in step 1 are all linear superpositions of two different ideal Brillouin singlet spectra, and two random Gaussian white noises are added. The generation formula is as follows:

[0015] G(ν)=∑g(ν)+∑n(ν)=∑g(ν)+(n1(ν)+n2(ν))

[0016]

[0017]

[0018]

[0019] Where g1(ν) and g2(ν) represent two different ideal Brillouin single peak spectra, the difference between them is the relative distance between the peak points of the two Brillouin frequency shifts. The Brillouin frequency shift corresponding to g1(ν) is ν B , the Brillouin frequency shift corresponding to g2(ν) is ν' B =ν B +a·Δν B , a is a constant coefficient; Γ1 and Γ2 represent the gain coefficients of two ideal Brillouin singlet spectra respectively; n1(ν) and n2(ν) represent two random Gaussian white noises;

[0020] When the constant coefficient a is set between 0.5 and 1.0, a Brillouin-like single-peak spectrum is generated; when the constant coefficient a is set between 1.0 and 1.5, a Brillouin short-range double-peak spectrum is generated; when the constant coefficient a is set between 1.5 and 2.0, a Brillouin long-range double-peak spectrum is generated.

[0021] Generate multiple Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra corresponding to two signal-to-noise ratios: high signal-to-noise ratio (100dB) and low signal-to-noise ratio (15dB).

[0022] Furthermore, the machine learning-based classification model described in step 2 adopts a fully connected neural network classification model based on the DenseNet-121 convolutional base.

[0023] According to another aspect of the present invention, a Brillouin profile recognition system based on machine learning is proposed, the system comprising:

[0024] A data acquisition module configured to acquire images containing multiple types of Brillouin gain spectra to form a training set; the multiple types of Brillouin gain spectra include Brillouin singlet spectra, Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra;

[0025] A classification model training module is configured to input a training set into a classification model based on machine learning for training, and obtain a trained classification model;

[0026] The classification module is configured to input the Brillouin gain spectrum of unknown type into the trained classification model to obtain the type of the Brillouin gain spectrum.

[0027] Furthermore, the data acquisition module simulates and generates images of various types of Brillouin gain spectra, wherein the Brillouin single peak spectrum is an ideal Brillouin gain spectrum plus random Gaussian white noise, and the generation formula is as follows:

[0028]

[0029] Where g(ν) is the ideal Brillouin gain, n(ν) is the random Gaussian white noise, and g B is the peak value of the Brillouin gain spectrum, ν is the frequency independent variable, i.e. the sweep frequency range, B is the Brillouin frequency shift corresponding to the spectral line, Δν B is the spectral line width; random(N) is the standard normal distribution with length N, N is the number of sweeps, SNR is the signal-to-noise ratio; g i (v) represents the Brillouin gain value corresponding to the i-th frequency sweep point;

[0030] Generate multiple Brillouin single peak spectrum images corresponding to two signal-to-noise ratios: high signal-to-noise ratio 100dB and low signal-to-noise ratio 15dB.

[0031] Furthermore, the Brillouin-like singlet spectrum, Brillouin short-range doublet spectrum, and Brillouin long-range doublet spectrum in the data acquisition module are all linear superpositions of two different Brillouin ideal singlet spectra, and two random Gaussian white noises are added. The generation formula is as follows:

[0032] G(ν)=∑g(ν)+∑n(ν)=∑g(ν)+(n1(ν)+n2(ν))

[0033]

[0034]

[0035]

[0036] Where g1(ν) and g2(ν) represent two different ideal Brillouin single peak spectra, the difference between them is the relative distance between the peak points of the two Brillouin frequency shifts. The Brillouin frequency shift corresponding to g1(ν) is ν B , the Brillouin frequency shift corresponding to g2(ν) is ν' B =ν B +a·Δν B , a is a constant coefficient; Γ1 and Γ2 represent the gain coefficients of two ideal Brillouin singlet spectra respectively; n1(ν) and n2(ν) represent two random Gaussian white noises;

[0037] Setting the constant coefficient a between 0.5 and 1.0 generates a Brillouin-like single-peak spectrum; setting the constant coefficient a between 1.0 and 1.5 generates a Brillouin short-range double-peak spectrum; setting the constant coefficient a between 1.5 and 2.0 generates a Brillouin long-range double-peak spectrum;

[0038] Generate multiple Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra corresponding to two signal-to-noise ratios: high signal-to-noise ratio (100dB) and low signal-to-noise ratio (15dB).

[0039] Furthermore, the machine learning-based classification model in the classification model training module adopts a fully connected neural network classification model based on the DenseNet-121 convolutional base.

[0040] The beneficial technical effects of the present invention are:

[0041] The present invention proposes to apply the convolutional neural network algorithm in the deep learning framework to optimize the data processing of the distributed fiber optic sensing system, and uses the deep learning framework to perform the classification and recognition of the special-shaped spectrum line images of the distributed fiber optic sensing system, which has very important practical significance for improving the spatial resolution, data processing speed and accuracy of the distributed fiber optic sensing system. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0043] Figure 1 A flow chart of a method for identifying Brillouin profiles based on machine learning according to an embodiment of the present invention is shown;

[0044] Figure 2 Four types of Brillouin gain spectrum line diagrams are shown; among them, (a) is a single-peak gain spectrum; (b) is a quasi-single-peak gain spectrum; (c) is a short-distance double-peak gain spectrum; (d) is a long-distance double-peak gain spectrum;

[0045] Figure 3 A schematic diagram showing the measured data of Brillouin frequency shift and OPGW cable transmission distance is shown;

[0046] Figure 4 The following diagram shows the patterns of four types of Brillouin gain spectrum lines corresponding to the actual demodulation results;

[0047] Figure 5 Schematic diagrams of single-peak spectra under two different signal-to-noise ratios are shown; (a) corresponds to SNR = 100 dB; (b) corresponds to SNR = 15 dB;

[0048] Figure 6 Schematic diagrams of three types of spectra under two different signal-to-noise ratios are shown; among them, (a) a quasi-single-peak spectrum corresponding to SNR=100dB; (b) a quasi-single-peak spectrum corresponding to SNR=15dB; (c) a short-distance double-peak spectrum corresponding to SNR=100dB; (d) a short-distance double-peak spectrum corresponding to SNR=15dB; (e) a long-distance double-peak spectrum corresponding to SNR=100dB; (f) a long-distance double-peak spectrum corresponding to SNR=15dB;

[0049] Figure 7 Figure 2 shows a schematic diagram of a three-hidden-layer fully connected structure;

[0050] Figure 8 Figure 1 shows the simulation results of the DenseNet-121 convolutional base under 100dB conditions; (a) corresponds to the change in the loss function value between training and testing; (b) corresponds to the change in accuracy between training and testing;

[0051] Figure 9 Figure 1 shows the simulation results of the DenseNet-121 convolutional base under 15dB conditions; (a) corresponds to the change in the loss function value between training and testing; (b) corresponds to the change in accuracy between training and testing;

[0052] Figure 10 The following figure shows the simulation results of the ResNet-101 convolutional base under 100dB conditions; (a) corresponds to the change in the loss function value between training and testing; (b) corresponds to the change in the accuracy between training and testing;

[0053] Figure 11 Figure 1 shows the simulation results of the ResNet-101 convolutional base under 15dB conditions; (a) corresponds to the change in the loss function value between training and testing; (b) corresponds to the change in accuracy between training and testing;

[0054] Figure 12 Figure 1 shows the simulation results of the VGG-16 convolutional base under 100dB conditions; (a) corresponds to the change in the loss function value between training and testing; (b) corresponds to the change in accuracy between training and testing;

[0055] Figure 13 The following figure shows the simulation results of the VGG-16 convolutional base under 15dB conditions; among them, (a) corresponds to the change in the loss function value between training and testing; (b) corresponds to the change in accuracy between training and testing. DETAILED DESCRIPTION

[0056] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0057] It is known to those skilled in the art that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. It should be understood that any number of elements in the accompanying drawings is for illustration and not limitation, and any nomenclature is for distinction only and does not have any limiting meaning.

[0058] Distributed fiber optic sensing typically calculates temperature and strain by sampling the Brillouin gain at various locations during fiber transmission, demodulating the Brillouin gain to derive the Brillouin frequency shift, and then calculating the temperature and strain at the abnormal location from the Brillouin frequency shift. Because the standard Brillouin gain spectrum conforms to the Lorentzian lineshape, the most common method for deriving the Brillouin frequency shift is to use Lorentz curve fitting.

[0059] However, in actual measurements, noise often exists, which makes the Brillouin gain spectrum not conform to the Lorentz line shape very well, resulting in a large number of various special-shaped spectra. When encountering Brillouin special-shaped spectra, the traditional LCF algorithm usually fits the double peaks separately or fits the double peaks into one peak, which affects the application of LCF. If the Brillouin frequency shift is demodulated by increasing the number of fitting iterations, it will consume a lot of time and the demodulation accuracy will also be affected.

[0060] Therefore, the present invention proposes a Brillouin anomaly spectrum recognition method and system based on machine learning, which adopts a data processing method based on machine learning algorithm to accurately identify Brillouin anomaly spectrum, thereby improving the demodulation speed and accuracy of Brillouin frequency shift.

[0061] The embodiment of the present invention proposes a Brillouin profile recognition method based on machine learning, such as Figure 1 As shown, the method includes the following steps:

[0062] Step 1: Obtain images containing multiple types of Brillouin gain spectra to form a training set;

[0063] Step 2: Input the training set into the classification model based on machine learning for training to obtain the trained classification model;

[0064] Step 3: Input the unknown type of Brillouin gain spectrum into the trained classification model to obtain the type of the Brillouin gain spectrum.

[0065] In step 1, images containing multiple types of Brillouin gain spectra are obtained to form a training set; according to the linear structure characteristics of the Brillouin gain shaped spectrum or double peak spectrum, the multiple types of Brillouin gain spectra include the following four types: Brillouin single peak spectrum (Single Peak), Brillouin similar single peak spectrum (Simi), Brillouin short distance double peak spectrum (Short Interval Double Peak, SI), Brillouin long distance double peak spectrum (Long Interval Double Peak, LI), Figure 2 Four types of Brillouin gain spectra are shown.

[0066] Figure 3 A schematic diagram of the measured data of Brillouin frequency shift and OPGW cable transmission distance is shown, which also provides experimental data verification for the rationality of the four types of classification. Figure 4 The graph shows the patterns of four types of Brillouin gain spectrum lines corresponding to the actual demodulation results. Types I to IV correspond to the Brillouin single-peak spectrum (type I), Brillouin-like single-peak spectrum (type II), Brillouin short-distance double-peak spectrum (type III), and Brillouin long-distance double-peak spectrum (type IV) mentioned above.

[0067] As an example, the specific process of obtaining an image containing various types of Brillouin gain spectra is as follows:

[0068] For each type of spectral line, the simulation generates a certain amount of image data, and the image content is the corresponding Brillouin gain spectrum of different types.

[0069] 1) For the Brillouin single peak spectrum, since the Brillouin gain spectrum line is a Lorentz line type, the Lorentz line type theory formula is used to generate the ideal Brillouin gain spectrum BGS formula:

[0070]

[0071] Where g(ν) is the Brillouin gain, g B is the peak value of the Brillouin gain spectrum, ν is the frequency independent variable, i.e. the sweep frequency range, B is the Brillouin frequency shift corresponding to the spectral line, Δν B is the spectral line width (Sweep Width, SW).

[0072] In order to make the spectrum closer to the real data, random Gaussian white noise is added to the ideal BGS to simulate the real measured BGS:

[0073]

[0074] Where n(ν) is Gaussian noise, random(N) is the standard normal distribution with length N, N is the number of sweeps, SNR is the signal-to-noise ratio, g i (v) represents the Brillouin gain value corresponding to the i-th sweep frequency point. So the final spectrum line generation formula is:

[0075]

[0076] The entire data set consists of two situations: high signal-to-noise ratio (SNR=100dB) and low signal-to-noise ratio (SNR=15dB). The high signal-to-noise ratio facilitates better training of the designed model; the low signal-to-noise ratio allows the model to maintain good prediction results even in the presence of large noise interference, thereby expanding the application range of the designed model.

[0077] It should be noted that, in an actual data acquisition device, the sweep frequency range can be set to 0 to 200 MHz, the sweep frequency step size is 1, a total of 200 sweep frequency points are generated, and a total of 300 images for each signal-to-noise ratio are generated as a training data set for the Brillouin singlet spectrum. Figure 5 The training set for each signal-to-noise ratio is shown. The frequency sweep range 0 to 200 MHz represents the frequency ν in the above formula.

[0078] 2) For the three types of spectral line shapes except the single-peak spectrum, they are essentially the superposition of two single-peak spectra (the quasi-single-peak spectrum is actually also a double-peak spectrum, but it is called a quasi-single-peak spectrum to simplify the name because the distance between the two peaks is extremely small). The difference lies in the relative distance between the peak points of the two Brillouin frequency shifts. Due to the randomness of the positions and peak sizes of the two peaks, let the Brillouin frequency shift at the relatively left position of the spectrum be ν B , then the Brillouin frequency shift formula on the right side of the spectrum line is:

[0079] ν' B =ν B +a·Δν B

[0080] Where a is a constant coefficient. Its product with the line width is the distance between the two peaks in the double-peak spectrum. Therefore, the ideal BGS line equation of the double-peak spectrum is the linear superposition of two different Brillouin single-peak spectra, which is expressed by the following formula:

[0081]

[0082] Among them, Γ1 and Γ2 are the peak values ​​of the two peaks. Their size relationship is also intuitively reflected in the relative height of the two peaks, and also represents the gain coefficients of the two ideal Brillouin single-peak spectra.

[0083] Afterwards, the ideal bimodal BGS curve is processed with Gaussian random white noise in the same way as the unimodal spectrum:

[0084]

[0085] The Gaussian white noise applied to the second peak of the Brillouin anomaly spectrum is also expressed by the following formula:

[0086]

[0087] Therefore, the actual BGS double-peak spectrum equation with Gaussian random noise is:

[0088] G(ν)=∑g(ν)+∑n(ν)=∑g(ν)+(n1(ν)+n2(ν))

[0089] On this basis, different types of double-peak spectra can be generated by controlling the numerical range of a. That is, for type II (single-peak spectrum), the range of a is set between 0.5 and 1.0; for type III (short-distance double-peak spectrum), the range of a is set between 1.0 and 1.5; for type IV (long-distance double-peak spectrum), the range of a is set between 1.5 and 2.0.

[0090] It should be noted that, in the actual data acquisition device, the sweep frequency range can also be set to 0 to 200MHz, the sweep frequency step size is 1, and a total of 200 sweep frequency points are generated, generating a total of 9000 images with 3000 images for each signal-to-noise ratio as the training data set of the three bimodal spectra. Similarly, two cases of high signal-to-noise ratio 100dB and low signal-to-noise ratio 15dB are prepared. Figure 6 The effects of three types of images with two different signal-to-noise ratios are shown.

[0091] At this point, the dataset has been created, with a total of 18,600 images serving as the training and test datasets for the entire system.

[0092] In step 2, the training set is input into a classification model based on machine learning for training to obtain a trained classification model. As an example, a fully connected neural network classification model based on the DenseNet-121 (Densely Connected Convolutional Networks) convolutional base is designed.

[0093] Compared to common single-branch convolutional networks, the DenseNet network boasts advantages such as dense connections, parameter sharing, and regularized overfitting resistance. This allows it to use fewer training parameters and offers stronger protection against vanishing and exploding gradients in backpropagation optimization. The key difference between it and the common residual ResNet connection is that the input to each layer in a residual connection is the sum of the output of the corresponding previous layer and the total raw input of the entire model, while the input to each layer of a DenseNet is the sum of the output of each previous layer and the total raw input, ensuring maximum reuse of each layer's data and the ability to achieve the same prediction results with fewer training parameters than ResNet. Furthermore, its built-in regularization effectively prevents test errors caused by overfitting the model to the training data, making the DenseNet convolutional architecture a very broad application in current computer vision, image recognition, and target detection scenarios.

[0094] After the DenseNet convolutional base, we designed a fully connected deep network architecture. Because the DenseNet convolutional base already has a relatively powerful feature extraction function, in order to achieve a fast response of the system, a three-hidden layer structure is used in addition to the input layer and output layer. Figure 7 The fully connected deep network architecture is shown.

[0095] For the input layer, the multidimensional tensor (Tensor) after passing through the DenseNet network is flattened into a one-dimensional vector according to a certain rule. The length of the one-dimensional vector is used as the input layer and input into the hidden layer. The number of hidden layer neurons is set to 1024→256→64 respectively. Finally, it enters the output layer and outputs the labels corresponding to the four types of spectral lines. The torch.utils module method is then used to load the set training and test datasets into the training loader (Train DataLoader) and the test loader (Test DataLoader) to facilitate batch data processing and increase the speed of model training and prediction. The batch size of each small batch is set to 32 to complete the entire training process. After multiple forward and backpropagation gradient optimization updates, the model weight parameters are retained to complete the model training.

[0096] During the model training process, the loss function is selected to use the cross entropy loss function (Cross Entropy LossFunction) and the classification accuracy for evaluation. The optimizer function is the Adam function, the learning rate is set to 0.001, and the number of training epochs is 100.

[0097] In step 3, the Brillouin gain spectrum of unknown type is input into the trained classification model to obtain the type of the Brillouin gain spectrum. Alternatively, a test data set can be input into the trained classification model to test the model accuracy.

[0098] The training data set and the test data set are divided in a ratio of 8:2. Figure 8 The graph shows the overall training and testing results. The graph selects the last 30 training and testing rounds, under a high signal-to-noise ratio of 100dB. The blue graph shows the change in the loss function and prediction accuracy during training, while the orange graph shows the change in the loss function and prediction accuracy during testing. The curves show that during training, the loss value remains around 0.0004, with an accuracy of 100%. During testing, the loss value remains around 0.001, with an average accuracy of 99.62%. This indicates that this model construction fully meets the expected accuracy requirements.

[0099] Figure 9 The graph shows the changes in the cross-entropy loss function and accuracy during training and testing under a low signal-to-noise ratio of 15dB. During training, the loss rate can be maintained at around 0.06%, with an accuracy of 99.98%. During testing, the loss rate is maintained at 0.37%, with an accuracy of up to 98.87%, achieving the accuracy requirements required in real-world applications.

[0100] The machine learning-based Brillouin shaped spectrum identification method proposed in the embodiment of the present invention focuses on the establishment of a classification and identification algorithm model for Brillouin gain spectra based on an image recognition algorithm and the model prediction effect. At the same time, using the measured demodulation effect of OPGW optical cable as an example, the classification basis of four spectral shapes, including Brillouin gain single-peak spectrum and three types of Brillouin gain double-peak shaped spectra, is explained, and the preparation and training process of the classification and identification data training set for transfer learning is explained.

[0101] The technical effects of the present invention are further verified through experiments.

[0102] Two other types of convolutional base models in modern neural network architectures that are currently widely used in image recognition research are selected, namely, fully connected neural network architectures using ResNet-101 and VGG-16 as convolutional bases for feature extraction, and compared with the prediction accuracy of the designed DenseNet-121 convolutional base to observe the prediction effects of the three convolutional base structures.

[0103] The main feature of the ResNet residual network is that the input of each layer carries the original input data of the entire model, which effectively prevents the problem of data gradient disappearing after processing too many layers. Therefore, it removes the limitation that the number of layers of deep neural networks cannot be increased indefinitely, and the number of layers of the training model can be continuously increased in this way.

[0104] The simulation results select the last 30 training and testing rounds, such as Figure 10 、 Figure 11 As shown in the figure, the blue plot shows the change in loss function and prediction accuracy during training, while the orange plot shows the change in loss function and prediction accuracy during testing. Again, two conditions, a high signal-to-noise ratio (SNR) of 100dB and a low SNR of 15dB, are used. For the 100dB condition, the curve shows that during training, the loss remains around 0.0011, with an accuracy of 99.92%. During testing, the loss remains around 0.0018, with an average accuracy of 99.24%. This performance is slightly lower than that of the DenseNet-121 convolutional architecture. For the 15dB condition, the curve shows that during training, the loss remains around 0.0037, with an accuracy of 99.43%. During testing, the loss remains around 0.0018, with an average accuracy of only 91.56%. This indicates that under low SNR conditions, the classification prediction performance of the ResNet-101 convolutional base is significantly lower than that of the DenseNet-121 convolutional base.

[0105] Then we select the VGG-16 convolutional architecture and compare its performance with the classification prediction model composed of two types of convolutional base structures, ResNet and DenseNet. The simulation results are selected from the last 30 training and testing rounds, such as Figure 12 、 Figure 13 As shown in the figure, the blue shows the change in loss function and prediction accuracy during the training process, and the orange shows the change in loss function and prediction accuracy during the testing process. For 100dB, the curve shows that during the training process, the cross entropy loss function value eventually drops to around 0.0038, and the classification prediction accuracy increases to 99.13%. During the testing process, the loss function loss value drops to around 0.0098, and the average prediction accuracy is 90.15%. This situation is significantly lower than the DenseNet-121 convolutional base architecture. For 15dB, the curve shows that during the training process, the loss value remains at around 0.0091, and the accuracy is 97.22%. During the testing process, the loss value remains at around 0.0329, and the average prediction accuracy is 82.65%. This situation is significantly lower than the DenseNet-121 convolutional base architecture.

[0106] The present invention uses a fully connected neural network classification prediction model based on the DenseNet convolutional base architecture, and compares the effects with two other widely used modern neural network convolutional architectures, ResNet and VGG-16. When the signal-to-noise ratio is 100dB, the loss function value and classification prediction accuracy of the neural network architecture based on the DenseNet-121 convolutional base are higher than those of the neural network architectures based on the VGG-16 convolutional base and the ResNet-101 convolutional base, both in the training process and the test process; when the signal-to-noise ratio is 15dB, the loss function value and classification prediction accuracy of the DenseNet-121 convolutional base architecture during the training process, as well as the prediction accuracy during the test process, are also higher than those of the VGG-16 and ResNet-101 architectures. Only the loss function value in the test phase is slightly higher than that of ResNet-101.

[0107] Another embodiment of the present invention provides a Brillouin profile recognition system based on machine learning, the system comprising:

[0108] A data acquisition module configured to acquire images containing multiple types of Brillouin gain spectra to form a training set; the multiple types of Brillouin gain spectra include Brillouin singlet spectra, Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra;

[0109] A classification model training module is configured to input a training set into a classification model based on machine learning for training, and obtain a trained classification model;

[0110] The classification module is configured to input the Brillouin gain spectrum of unknown type into the trained classification model to obtain the type of the Brillouin gain spectrum.

[0111] In this embodiment, preferably, the data acquisition module simulates and generates images of various types of Brillouin gain spectra, wherein the Brillouin singlet spectrum is an ideal Brillouin gain spectrum plus random Gaussian white noise, and the generation formula is as follows:

[0112]

[0113] Where g(ν) is the ideal Brillouin gain, n(ν) is the random Gaussian white noise, and g B is the peak value of the Brillouin gain spectrum, ν is the frequency independent variable, i.e. the sweep frequency range, B is the Brillouin frequency shift corresponding to the spectral line, Δν B is the spectral line width; random(N) is the standard normal distribution with length N, N is the number of sweeps, SNR is the signal-to-noise ratio; g i (v) represents the Brillouin gain value corresponding to the i-th frequency sweep point;

[0114] Generate multiple Brillouin single peak spectrum images corresponding to two signal-to-noise ratios: high signal-to-noise ratio 100dB and low signal-to-noise ratio 15dB.

[0115] In this embodiment, preferably, the Brillouin-like singlet spectrum, the Brillouin short-range doublet spectrum, and the Brillouin long-range doublet spectrum in the data acquisition module are all linear superpositions of two different Brillouin ideal singlet spectra, and two random Gaussian white noises are added, and the generation formula is as follows:

[0116] G(ν)=∑g(ν)+∑n(ν)=∑g(ν)+(n1(ν)+n2(ν))

[0117]

[0118]

[0119]

[0120] Where g1(ν) and g2(ν) represent two different ideal Brillouin single peak spectra, the difference between them is the relative distance between the peak points of the two Brillouin frequency shifts. The Brillouin frequency shift corresponding to g1(ν) is ν B , the Brillouin frequency shift corresponding to g2(ν) is ν' B =ν B +a·Δν B , a is a constant coefficient; Γ1 and Γ2 represent the gain coefficients of two ideal Brillouin singlet spectra respectively; n1(ν) and n2(ν) represent two random Gaussian white noises;

[0121] Setting the constant coefficient a between 0.5 and 1.0 generates a Brillouin-like single-peak spectrum; setting the constant coefficient a between 1.0 and 1.5 generates a Brillouin short-range double-peak spectrum; setting the constant coefficient a between 1.5 and 2.0 generates a Brillouin long-range double-peak spectrum;

[0122] Generate multiple Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra corresponding to two signal-to-noise ratios: high signal-to-noise ratio (100dB) and low signal-to-noise ratio (15dB).

[0123] In this embodiment, preferably, the machine learning-based classification model in the classification model training module adopts a fully connected neural network classification model based on the DenseNet-121 convolutional base.

[0124] It should be noted that although several units, modules, or submodules are mentioned in the detailed description above, such division is merely exemplary and not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above may be embodied in one module. Conversely, the features and functions of one module described above may be further divided and embodied by multiple modules.

[0125] Furthermore, although the operations of the method of the present invention are described in a particular order in the accompanying drawings, this does not require or imply that these operations must be performed in this particular order, or that all illustrated operations must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0126] Although the spirit and principles of the present invention have been described with reference to several specific embodiments, it should be understood that the present invention is not limited to the specific embodiments disclosed, and the division into various aspects does not mean that the features of these aspects cannot be combined to benefit. Such division is only for the convenience of expression. The present invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A Brillouin profile recognition method based on machine learning, characterized in that: The following steps are involved: Step 1: Acquire images containing multiple types of Brillouin gain spectra to form a training set; the multiple types of Brillouin gain spectra include Brillouin singlet spectra, Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra; simulate and generate images of multiple types of Brillouin gain spectra, where the Brillouin singlet spectra are ideal Brillouin gain spectra plus random Gaussian white noise, and the generation formula is as follows: ； Where, is the ideal Brillouin gain, is random Gaussian white noise, is the peak value of the Brillouin gain spectrum, is the frequency independent variable, i.e. the frequency sweep range, is the Brillouin frequency shift corresponding to the spectral line, is the spectral line width; The length is The standard normal distribution of is the sweep frequency number, is the signal-to-noise ratio; Indicates the Brillouin gain value corresponding to the i-th sweep frequency point; The Brillouin-like singlet spectrum, Brillouin short-range doublet spectrum, and Brillouin long-range doublet spectrum are all linear superpositions of two different ideal Brillouin singlet spectra, and two random Gaussian white noises are added. The generation formula is as follows: ； ； ； ； Where, 、 They represent two different ideal Brillouin single peak spectra, the difference between which is the relative distance between the peak points of the two Brillouin frequency shifts. The corresponding Brillouin frequency shift is , The corresponding Brillouin frequency shift is , is a constant coefficient; and They represent the gain coefficients of the two ideal Brillouin singlet spectra respectively; 、 represents two random Gaussian white noises; Let constant coefficient The value of is between 0.5 and 1.0, generating a Brillouin-like single-peak spectrum; let the constant coefficient The value of is between 1.0 and 1.5, generating a Brillouin short-range double-peak spectrum; let the constant coefficient The value of is between 1.5 and 2.0, generating a Brillouin long-distance double-peak spectrum; Step 2: Input the training set into the classification model based on machine learning for training to obtain the trained classification model; Step 3: Input the unknown type of Brillouin gain spectrum into the trained classification model to obtain the type of the Brillouin gain spectrum.

2. The method for identifying Brillouin profiles based on machine learning according to claim 1, wherein: In step 1, multiple Brillouin singlet spectrum images corresponding to two signal-to-noise ratios (SNR) of 100 dB and 15 dB are generated.

3. The method for identifying Brillouin profiles based on machine learning according to claim 1, wherein: In step 1, multiple Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra images corresponding to two signal-to-noise ratios: a high signal-to-noise ratio of 100 dB and a low signal-to-noise ratio of 15 dB are generated.

4. The method for identifying Brillouin profiles based on machine learning according to claim 1, wherein: The machine learning-based classification model described in step 2 adopts a fully connected neural network classification model based on the DenseNet-121 convolutional base.

5. A Brillouin profile recognition system based on machine learning, characterized in that: include: The data acquisition module is configured to acquire images containing multiple types of Brillouin gain spectra to form a training set; the multiple types of Brillouin gain spectra include Brillouin singlet spectra, Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra; and simulate and generate images of the multiple types of Brillouin gain spectra, wherein the Brillouin singlet spectra are ideal Brillouin gain spectra plus random Gaussian white noise, and the generation formula is as follows: ； Where, is the ideal Brillouin gain, is random Gaussian white noise, is the peak value of the Brillouin gain spectrum, is the frequency independent variable, i.e. the frequency sweep range, is the Brillouin frequency shift corresponding to the spectral line, is the spectral line width; The length is The standard normal distribution of is the sweep frequency number, is the signal-to-noise ratio; Indicates the Brillouin gain value corresponding to the i-th sweep frequency point; The Brillouin-like single-peak spectrum, Brillouin short-range double-peak spectrum, and Brillouin long-range double-peak spectrum are all linear superpositions of two different Brillouin ideal single-peak spectra, and two random Gaussian white noises are added. The generation formula is as follows: ； ； ； ； Where, 、 They represent two different ideal Brillouin single peak spectra, the difference between which is the relative distance between the peak points of the two Brillouin frequency shifts. The corresponding Brillouin frequency shift is , The corresponding Brillouin frequency shift is , is a constant coefficient; and They represent the gain coefficients of the two ideal Brillouin singlet spectra respectively; 、 represents two random Gaussian white noises; Let constant coefficient The value of is between 0.5 and 1.0, generating a Brillouin-like single-peak spectrum; let the constant coefficient The value of is between 1.0 and 1.5, generating a Brillouin short-range double-peak spectrum; let the constant coefficient The value of is between 1.5 and 2.0, generating a Brillouin long-distance double-peak spectrum; A classification model training module is configured to input a training set into a classification model based on machine learning for training, and obtain a trained classification model; The classification module is configured to input the Brillouin gain spectrum of unknown type into the trained classification model to obtain the type of the Brillouin gain spectrum.

6. The Brillouin profile recognition system based on machine learning according to claim 5, characterized in that: The data acquisition module generates multiple Brillouin single peak spectrum images corresponding to two signal-to-noise ratios: a high signal-to-noise ratio of 100 dB and a low signal-to-noise ratio of 15 dB.

7. The Brillouin profile recognition system based on machine learning according to claim 5, characterized in that: The data acquisition module generates multiple Brillouin-like singlet spectra, Brillouin short-range doublet spectra, and Brillouin long-range doublet spectra corresponding to two signal-to-noise ratios: a high signal-to-noise ratio of 100 dB and a low signal-to-noise ratio of 15 dB.

8. The Brillouin profile recognition system based on machine learning according to claim 5, characterized in that: The machine learning-based classification model in the classification model training module adopts a fully connected neural network classification model based on the DenseNet-121 convolutional base.

Images