阅读说明：本技术 一种基于卷积神经网络集成学习的Sagnac分布光纤传感系统的定位方法 (Positioning method of Sagnac distributed optical fiber sensing system based on convolutional neural network integrated learning ) 是由 方捻 吕继东 王陆唐 王春华 于 2020-11-02 设计创作，主要内容包括：本发明公开了一种基于卷积神经网络集成学习的Sagnac分布光纤传感系统的定位方法,以传感光纤上固定间隔点为扰动点,由传感系统分别获取在各点模拟扰动产生的干涉信号,对其做预处理后作为训练集和验证集；针对不同的损失函数训练两个卷积神经网络CNN模型,使两者分别准确定位近端和远端扰动；通过集成学习方法组合两个模型的训练结果,得到基于CNN集成学习的扰动位置预测模型；通过验证集优化每个模型的参数。将待定位的干涉信号进行预处理后作为测试样本,利用训练好的预测模型对其测试,得到其扰动位置。本发明无需信号解调,系统复杂度低,且对噪声不敏感,数据处理方法简单,定位结果稳定、准确,可用于环形或直线式Sagnac分布光纤传感系统的扰动定位。(The invention discloses a positioning method of a Sagnac distributed optical fiber sensing system based on convolutional neural network integrated learning, which takes fixed interval points on a sensing optical fiber as disturbance points, respectively obtains interference signals generated by simulated disturbance at each point by the sensing system, and takes the interference signals as a training set and a verification set after preprocessing; training two convolutional neural network CNN models aiming at different loss functions to accurately position near-end disturbance and far-end disturbance respectively; combining the training results of the two models through an ensemble learning method to obtain a disturbance position prediction model based on CNN ensemble learning; the parameters of each model are optimized through the validation set. And preprocessing the interference signal to be positioned to be used as a test sample, and testing the interference signal by using a trained prediction model to obtain the disturbance position of the interference signal. The method has the advantages of no need of signal demodulation, low system complexity, insensitivity to noise, simple data processing method and stable and accurate positioning result, and can be used for disturbance positioning of an annular or linear Sagnac distribution optical fiber sensing system.)

1. A positioning method of a Sagnac distributed optical fiber sensing system based on convolutional neural network integrated learning is characterized by comprising the following steps:

1) taking fixed interval points on the sensing optical fiber as disturbance points, respectively acquiring interference signals generated by simulating disturbance at each point by a sensing system, preprocessing the interference signals, and selecting one part as a training set and the other part as a verification set;

2) training two convolutional neural network CNN models aiming at different loss functions, so that the two CNN models can accurately position near-end disturbance and far-end disturbance respectively; performing parameter optimization through the verification set to obtain the optimal training effect;

3) the combination of the training results of the two CNN models is realized through an ensemble learning method, and a final disturbance position prediction model based on CNN ensemble learning is obtained; similarly, the verification set is utilized to carry out parameter optimization to obtain the optimal training effect;

4) carrying out the same pretreatment on an interference signal to be positioned to be used as a test sample;

5) inputting the test sample into a trained disturbance position prediction model based on CNN ensemble learning, predicting a disturbance position corresponding to the test sample, and realizing disturbance positioning.

2. The positioning method of the Sagnac distributed optical fiber sensing system based on the convolutional neural network ensemble learning of claim 1, wherein the preprocessing in the steps 1) and 4) comprises: and acquiring the frequency spectrum of the interference signal or the frequency spectrum of the interference signal, carrying out normalization processing, and determining the appropriate frequency spectrum data length.

3. The positioning method of the Sagnac distributed optical fiber sensing system based on convolutional neural network ensemble learning of claim 1, wherein the ensemble learning method in the step 3) adopts a Stacking, Bagging or Boosting ensemble learning method.

Technical Field

The invention relates to a positioning method of a distributed optical fiber sensing system, in particular to a positioning method of a Sagnac distributed optical fiber sensing system based on Convolutional Neural Network (CNN) integrated learning.

Background

The distributed optical fiber sensing technology has a wide application prospect in high-pressure pipeline leakage detection and positioning, wherein the Sagnac distributed optical fiber sensing system has the advantages of strong anti-interference performance, low requirement on a light source and the like, and is one of the hot spots of the current research. The zero frequency method is the main positioning method of the Sagnac distributed optical fiber sensing system, however, the zero frequency is often submerged in noise, and the positioning accuracy is affected. Positioning accuracy can be improved by improving the structure of a sensing system and using phase generated carrier homodyne demodulation, and positioning accuracy can also be improved by a series of algorithms such as wavelet soft threshold denoising, twice Fourier transform and the like, but the methods increase the complexity of the system to a certain extent and are sensitive to system noise.

In recent years, machine learning methods have begun to be used in the distributed fiber sensing field. The research is more on the identification of optical fiber disturbance signals, and the research on disturbance positioning is gradually appeared. Some learners use a Support Vector Machine (SVM) regression model to enable the measured position to approach the real position, but the zero frequency method is needed to be used for preliminary positioning in advance, and the positioning process is complex. The applicant has proposed a discretization positioning method based on a machine learning classification model for a Sagnac distributed optical fiber sensing system, which converts the positioning problem of disturbance into a multi-classification problem of interference signals caused by disturbance of different sensing optical fiber positions, however, the classification method needs to collect interference signals of all disturbance positions, and loses the advantage that distributed optical fiber sensing can be continuously monitored, which becomes a technical problem to be solved urgently.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a positioning method of a Sagnac distributed optical fiber sensing system based on convolutional neural network integrated learning, which simplifies the signal processing process, reduces the acquisition amount of training samples and predicts any disturbance position.

In order to achieve the purpose of the invention, the invention adopts the following technical scheme:

a positioning method of a Sagnac distributed optical fiber sensing system based on convolutional neural network integrated learning comprises the following steps:

1) taking fixed interval points on the sensing optical fiber as disturbance points, respectively acquiring interference signals generated by simulating disturbance at each point by a sensing system, preprocessing the interference signals, and selecting one part as a training set and the other part as a verification set;

2) training two convolutional neural network CNN models aiming at different loss functions, so that the two CNN models can accurately position near-end disturbance and far-end disturbance respectively; performing parameter optimization through the verification set to obtain the optimal training effect;

3) the combination of the training results of the two CNN models is realized through an ensemble learning method, and a final disturbance position prediction model based on CNN ensemble learning is obtained; similarly, the verification set is utilized to carry out parameter optimization to obtain the optimal training effect;

4) carrying out the same pretreatment on an interference signal to be positioned to be used as a test sample;

5) inputting the test sample into a trained disturbance position prediction model based on CNN ensemble learning, predicting a disturbance position corresponding to the test sample, and realizing disturbance positioning.

Preferably, the pre-treatment in the step 1) and the step 4) comprises: and acquiring the frequency spectrum of the interference signal or the frequency spectrum of the interference signal, carrying out normalization processing, and determining the appropriate frequency spectrum data length.

Preferably, the ensemble learning method in the step 3) adopts a Stacking, Bagging or Boosting ensemble learning method.

Working principle of the invention

Taking a ring-shaped Sagnac interference distribution optical fiber sensing system as an example, when a broadband disturbance signal is at a distance R from a Sagnac interference ring₁The frequency difference between adjacent zeros of the frequency spectrum of the phase change signal caused by the interference on the sensing fiber is called zero frequency difference, and is denoted as f, f and the disturbance position R₁The relationship of (1) is:

in the above formula, c represents the speed of light in vacuum, n represents the refractive index of the fiber core, and L represents the length of the fiber ring; as can be seen from equation (1), the perturbation positions are different, the magnitude of the zero frequency difference is also different, and R is₁The smaller f is, namely the disturbance closer to the coupler, which is called near-end disturbance for short, and the zero frequency difference of interference signal frequency spectrum caused by the disturbance is smaller; therefore, the frequency spectrum of the interference signal can well reflect the disturbance position and can be used as input data of a disturbance position prediction model.

R₁The rate of change of relative f is:

according to formula (2), R₁The rate of change of relative f is inversely related to the square. Since the near-end perturbation causes a smaller f, the rate of change of the near-end perturbation position from the zero frequency difference is larger, i.e. a small change in f will result in a larger change in the perturbation position. Therefore, two CNN models are used to solve the problem of near-end and far-end disturbance positioning respectively, the output of model 1 is the normalized disturbance position, and the output of model 2 is the normalized zero frequency difference. The convolutional neural network has strong feature extraction capability, so that feature extraction is not needed, and the frequency spectrum data of interference signals caused by disturbance can be directly used as input signals of the two CNN models.

Due to the fact that the positioning performance of the Sagnac distribution optical fiber sensing system is in non-uniform distribution along the length direction of the sensing optical fiber, the mean square error loss function is used, and a good output result is difficult to obtain in the whole sensing optical fiber range. Therefore, the model 1 outputting as the normalized disturbance position focuses on solving the near-end disturbance problem which is difficult to accurately locate. The prediction error of the position of the near-end disturbance is made smaller by increasing the weight of the prediction error of the position of the near-end disturbance in the loss function, i.e. minimizing the loss function. For the model 2 with the output of the normalized zero frequency difference, the zero frequency difference of the near-end disturbance can cause a larger positioning error as long as a smaller prediction error exists, and the larger prediction error of the far-end disturbance has a smaller influence on the positioning error, so that the model is focused on solving the far-end disturbance positioning. By adding proper weight values into the loss function, the model can balance the overall positioning performance of the far-end disturbance to a certain extent.

The ensemble learning method is good at solving the problem of insufficient learning ability of a single learner, so that a disturbance position prediction model based on CNN ensemble learning can be obtained by combining the training results of two CNN models by using the ensemble learning method. The model can be used to predict the position of any unknown disturbance on the sensing fiber.

Compared with the prior art, the invention has the following obvious and prominent substantive characteristics and remarkable advantages:

1. the positioning method of the Sagnac distributed optical fiber sensing system based on convolutional neural network integrated learning does not need signal demodulation, has low system complexity, is insensitive to noise, has a simple data processing method, has stable and accurate positioning result, and can be used for disturbance positioning of an annular or linear Sagnac distributed optical fiber sensing system;

2. the invention does not need to collect interference signals of all disturbance positions, can ensure the continuous monitoring of the distributed optical fiber sensing system, can predict the position of any unknown disturbance on the sensing optical fiber, and has simple positioning process and high efficiency.

Drawings

Fig. 1 is a schematic structural diagram of a system adopted in the embodiment of the present invention.

Fig. 2 is a flow chart of a positioning method of the present invention.

FIG. 3 is a normalized spectrum of interference signals acquired in the absence and presence of noise.

Fig. 4 shows the structure of two convolutional neural networks used in the embodiment.

Fig. 5 shows the structure of a single hidden layer neural network in the Stacking ensemble learning method used in the embodiment.

Detailed Description

The above-described scheme is further illustrated below with reference to specific embodiments, which are detailed below:

the first embodiment is as follows:

in this embodiment, referring to fig. 1-2, a positioning method for a Sagnac distributed optical fiber sensing system based on convolutional neural network integrated learning includes the following steps:

1) taking fixed interval points on the sensing optical fiber as disturbance points, respectively acquiring interference signals generated by simulating disturbance at each point by a sensing system, preprocessing the interference signals, and selecting one part as a training set and the other part as a verification set;

2) training two convolutional neural network CNN models aiming at different loss functions, so that the two CNN models can accurately position near-end disturbance and far-end disturbance respectively; performing parameter optimization through the verification set to obtain the optimal training effect;

3) the combination of the training results of the two CNN models is realized through an ensemble learning method, and a final disturbance position prediction model based on CNN ensemble learning is obtained; similarly, the verification set is utilized to carry out parameter optimization to obtain the optimal training effect;

4) carrying out the same pretreatment on an interference signal to be positioned to be used as a test sample;

5) inputting the test sample into a trained disturbance position prediction model based on CNN ensemble learning, predicting a disturbance position corresponding to the test sample, and realizing disturbance positioning.

The positioning method of the Sagnac distributed optical fiber sensing system based on convolutional neural network integrated learning does not need signal demodulation, is low in system complexity, insensitive to noise, simple in data processing method and stable and accurate in positioning result, and can be used for disturbance positioning of an annular or linear Sagnac distributed optical fiber sensing system.

Example two:

this embodiment is substantially the same as the first embodiment, and is characterized in that:

in this embodiment, the preprocessing in step 1) and step 4) includes: and acquiring the frequency spectrum of the interference signal or the frequency spectrum of the interference signal, carrying out normalization processing, and determining the appropriate frequency spectrum data length.

The ensemble learning method in the step 3) adopts a Stacking, Bagging or Boosting ensemble learning method.

The embodiment can ensure the continuous monitoring of the distributed optical fiber sensing system, can predict the position of any unknown disturbance on the sensing optical fiber, and has simple positioning process and high efficiency.

Example three:

this embodiment is substantially the same as the above embodiment, and is characterized in that:

in this embodiment, since there is no condition temporarily, and various data required by the disturbance position prediction model cannot be actually acquired, monitoring of the pipeline leakage by using the OptiSystem software simulation annular Sagnac distributed optical fiber sensing system is used to verify the feasibility of the positioning method based on the convolutional neural network ensemble learning in this embodiment.

As shown in fig. 1, the simulated ring-shaped Sagnac distributed optical fiber sensing system includes a continuous laser 1, a 2 × 2 bidirectional 3dB optical coupler 2, a sensing optical fiber 3, a sensing optical fiber 4, a delay optical fiber 5, a phase modulator 6, a photodetector 7, and a data acquisition and processing unit 8. Sensing fiber 3, length R of sensing fiber 4₁、R₂The sum being equal to the length R of the delay fiber₃. Since the bandwidth of the pipeline leakage signal is about 60kHz, the phase modulator 6 is driven by a Sinc function with the bandwidth of 60kHz to simulate the disturbance of the leakage at different positions of the pipeline to the sensing optical fiber laid on the pipeline. By varying R₁、R₂Simulating different disturbance positions.

In this embodiment, the delay fiber length R₃10km, the total length of the sensing fiber is also 10 km. And taking a sensing optical fiber within 8.5km from the coupler as an effective sensing area.

In order to better evaluate the positioning result of each disturbance on the sensing optical fiber, a multi-parameter evaluation index is adopted, wherein the multi-parameter evaluation index comprises average absolute error, standard deviation of error, maximum error, minimum error and the number of samples with the error larger than a preset value. The average absolute error reflects the actual situation of the prediction error of the disturbance position, the standard error is used for measuring the discrete degree of the prediction error of each sample position, the maximum error shows the worst situation of the model prediction error, and the smaller the value is, the smaller the error positioning risk of the model is. The minimum error shows the best predictive potential of the model. Presetting acceptable maximum positioning error, and measuring the overall position prediction effect of the model by using the number of samples with the error larger than a preset value, wherein the less the number of samples with the error larger than the preset value is, the stronger the practicability of the model is.

The pipeline leak is located according to the flow shown in fig. 2. An acceptable error of 50m is set, and the sensing optical fiber is disturbed at intervals of 50 m. Thus, a total of 171 interference signals generated by the fixed interval position disturbances are acquired over an effective detection length of 8.5 km. Because of the large number of CNN parameters, a large sample set is required to prevent the network model from being over-fitted. The interference signal data obtained under the noise with different intensities is simulated by adding Gaussian white noise with different intensities to each noise-free interference signal obtained by a simulation system, and obtaining 200 interference signals at each position, wherein the signal-to-noise ratio of the interference signals is increased from 1dB to 20.9dB, and the amplification is 0.1 dB.

The frequency spectrum of all interference signals is obtained by fast fourier transform. Since at least three zero frequency points, namely two zero frequency differences, need to appear in the frequency spectrum to resolve the position characteristics, and the farther the disturbance position is, the larger the zero frequency difference is, and the two zero frequency differences are about 134kHz at the farthest disturbance position of 8.5km, therefore, the input data length is set to 1024, that is, the frequency spectrum of 0-160kHz is reserved as the input data of the CNN model. In order to avoid large module value difference of different interference signal frequency spectrums, normalization processing is carried out on each interference signal frequency spectrum according to the module value range of the self frequency spectrum. Fig. 3 shows the normalized spectrum of the resulting noise-free and noise-containing interference signal, with the perturbation located at 4 km.

The interference signal spectra obtained at the 171 fixed interval positions are grouped and numbered in the order of their perturbation positions. Interference signal spectrums with different signal-to-noise ratios of 200 under each odd group number are selected as a training set, and interference signal spectrums with different signal-to-noise ratios of 30 under each even group number are randomly selected as a verification set.

The CNN model used is a 1-dimensional CNN, and its structure is shown in fig. 4, where the convolutional layer and the pooling layer have three layers, and the fully-connected layer has four layers, where the last layer is a network output layer. The output of the model 1 is the normalized disturbance position, the output of the model 2 is the normalized zero frequency difference, and the min-max normalization method is used. The activation function and partial parameters of the two models are the SAME, the convolution layer uses a ReLU activation function and a 'SAME' filling mode, the size of a convolution kernel is 15, and the step length is 1; the pooling layer adopts a maximum pooling method, the size of a pooling window is 12, and the step length is 2; the first three layers of the fully connected layer also use the ReLU activation function, with the output layer using the tanh activation function. In contrast, in model 1 and model 2, the number of convolution kernels in each layer is 1, 8, 16 and 1, 16, 32, respectively, and the number of neurons in the full connection layer is 2048, 512, 128, 1 and 4096, 1024, 256, 1, respectively.

Setting the loss function of model 1 training to

The loss function of model 2 training is

In formulae (3) and (4), y_pPredict output for model, y_iFor model target output, m represents the number of samples, k is a constant, and k is fine-tuned by the model training result in order to balance the weight.

The CNN model is based on a TensorFlow neural network framework, a TensorFlow version 1.6.0 and a Python version 3.6.3, and network training and parameter optimization are carried out by using an Invitta GTX 1080Ti video card. And optimizing the CNN model parameters by an adaptive moment estimation (Adma) optimization algorithm to minimize a loss function value. Then, according to the test effect of the CNN model in the verification set, the batch size, the learning rate, the dropout proportion, the loss function parameter k and other super parameters of the training data are finely adjusted, so that the model training effect is optimal. Finally, the batch size of the two networks is determined to be 32, and the learning rate is 110^-4The dropout ratio is 20%. Set 0-2.4km as the near endAnd 2.4km-8.5km is taken as a far end. Two CNN models are trained by selecting different parameters k, so that the average absolute error of the disturbance position prediction of the model 1 at the near end and the disturbance position prediction of the model 2 at the far end and the number of samples with the error larger than 50m are the lowest. Through multiple times of training, when the model is iterated for 60 times, the average absolute errors of the model 1 at the near end and the model 2 at the far end respectively reach the minimum values of 13.93m and 16.76m, the number of samples with the errors larger than 50m respectively reach the minimum values of 3 and 38, and the k values 1 and 1.5 at the moment are selected as the optimal parameters of the two loss functions.

And respectively normalizing the training outputs of the two models or performing normalization on the training outputs, and then converting the training outputs by the formula (1) to obtain a disturbance position for subsequent ensemble learning.

And combining two different CNN models by using a Stacking ensemble learning method, namely combining 2 CNN models by using the CNN as a base learner and using a single hidden layer neural network as a meta-learner. The single hidden layer neural network structure is shown in fig. 5. The input layer comprises 2 neurons, namely 2 base learners, the hidden layer comprises 16 neurons, and the output layer comprises 1 neuron, which is the combined output of the base learners.

And combining the disturbance positions of the single sample, which are predicted by the CNN model 1 and the CNN model 2, into a feature vector, wherein the feature vector is used as the input of the single hidden layer neural network, and the real disturbance position corresponding to the sample is used as the target output of the network. And taking the feature vectors obtained by training the two CNN models as a training set of the single hidden layer neural network. Similarly, the feature vectors obtained from the verification set of the CNN model are used as the verification set of the single hidden layer neural network.

Using the mean square error as a loss function, an Adma optimization algorithm determines the optimal model parameters of the single hidden layer neural network. The learning rate of the Adma optimization algorithm is also 110^-4. When the model iterates 20 ten thousand times, the mean square error loss function value of the training set reaches 5.1710^-6And the decline is slow, the model can be considered to be trained.

In the testing stage, disturbing the sensing optical fiber at random positions in each 100m interval to obtain interference signals generated by disturbance at 85 random interval positions; interference signals generated by disturbance at 24 small fixed intervals in the near area, the middle area and the far area of the sensing optical fiber are obtained in the areas of 1.0-1.1km, 4.0-4.1km and 7.0-7.1km at intervals of 10m, wherein 171 disturbance positions with 50m large fixed intervals are planed out. Similar to the training process, white gaussian noise of different intensities is also added to the noise-free interference signal obtained at each random interval and small fixed interval position to test the position prediction effect of the interference signal at different intensities. And acquiring frequency spectrums of the interference signals, similarly selecting 1024 frequency spectrum data as input data, and performing the same normalization processing on the frequency spectrums.

In interference signal frequency spectrums obtained from 85 random interval positions, 30 interference signal frequency spectrums with different signal-to-noise ratios are randomly selected at each position to serve as a test set 1, namely, each section of the sensing optical fiber is tested once to check the positioning effect of the CNN ensemble learning-based disturbance positioning method in the whole range of the sensing optical fiber. In interference signal frequency spectrums obtained from 24 small fixed interval positions, 30 interference signal frequency spectrums with different signal-to-noise ratios are randomly selected at each position to serve as a test set 2, so that the contingency is eliminated, the positioning feasibility of any disturbance position is further verified, and the positioning resolution is explored.

And testing the samples of the two test sets by using the two CNN models obtained by training to respectively obtain prediction outputs. And converting by using the same method as the training stage to obtain the predicted disturbance positions of the two CNN models for subsequent meta-learner combination and test.

Similarly, the feature vectors obtained from the test set of the CNN model are used as the test set of the single hidden layer neural network. Inputting the trained single hidden layer neural network for combination to obtain a prediction result of the disturbance position, namely realizing disturbance positioning.

The average absolute error, standard error deviation, maximum error, minimum error and the ratio of the number of samples with the error larger than a preset value to the total number of the test samples of the disturbance position prediction model based on CNN ensemble learning are respectively 14.6m, 9.8m, 41.6m, 1.2m and 1.25%; for test set 2, 11.4m, 5.9m, 27.0m, 1.9m and 0.28% respectively. It can be seen that the test results of the two test sets are not obviously different, which indicates that the disturbance position prediction model based on the CNN ensemble learning really has the feasibility of positioning any disturbance position of the sensing optical fiber, the average positioning error is only tens of meters, the resolution can reach 10m, and the positioning result is stable and reliable.

To sum up, in the above embodiment, the positioning method of the Sagnac distributed optical fiber sensing system based on convolutional neural network ensemble learning takes the fixed interval points on the sensing optical fiber as disturbance points, and the sensing system respectively acquires interference signals generated by simulated disturbance at each point, and after preprocessing the interference signals, the interference signals are used as a training set and a verification set; training two convolution neural network models aiming at different loss functions to accurately position near-end disturbance and far-end disturbance respectively; combining the training results of the two models through an ensemble learning method to obtain a disturbance position prediction model based on CNN ensemble learning; the parameters of each model are optimized through the validation set. And preprocessing the interference signal to be positioned to be used as a test sample, and testing the interference signal by using a trained prediction model to obtain the disturbance position of the interference signal. The method has the advantages of no need of signal demodulation, low system complexity, insensitivity to noise, simple data processing method and stable and accurate positioning result, and can be used for disturbance positioning of an annular or linear Sagnac distribution optical fiber sensing system.

The embodiments of the present invention have been described with reference to the accompanying drawings, but the present invention is not limited to the embodiments, and various changes and modifications can be made according to the purpose of the invention, and any changes, modifications, substitutions, combinations or simplifications made according to the spirit and principle of the technical solution of the present invention shall be equivalent substitutions, as long as the purpose of the present invention is met, and the present invention shall fall within the protection scope of the present invention without departing from the technical principle and inventive concept of the present invention.