A distributed optical fiber fence monitoring feature learning method

Abstract

The present application relates to the field of distributed optical fiber sensing-based intrusion event classification and identification, in particular to a distributed optical fiber fence monitoring feature learning method. The present application improves the learning efficiency and recognition performance of the network through feature learning, and proposes a vibration recognition method based on MTF-NMF and 2-D CNN. The one-dimensional phase time domain signal without denoising is converted into a Markov transition field (MTF) matrix, then the base matrix is extracted from the MTF through non-negative matrix factorization (NMF) and saved as an RGB image, and finally the feature learning and recognition are performed by the 2-D CNN network, so as to realize the classification of different types of intrusion events on the optical fiber fence. The present application effectively solves the problems of low learning efficiency of the intrusion event recognition model and high construction cost of the recognition network training data set in the current Φ-OTDR monitoring field.

Description

Technical Field

[0001] This invention relates to the field of intrusion event classification and identification based on distributed optical fiber sensing, and specifically to a representation learning method for distributed optical fiber fence monitoring. Background Technology

[0002] Distributed fiber optic vibration sensors based on phase-sensitive optical time-domain reflectometers (Φ-OTDR) can simultaneously detect multiple vibration events occurring on an optical fiber. Due to their advantages such as long measurement distance, high sensitivity, and simple configuration, they have attracted widespread attention in various security monitoring fields. However, in practical applications, significant noise interference can lead to a high false alarm rate for Φ-OTDRs. Therefore, researching methods to accurately identify threat events and reduce false alarms is crucial. Based on the linear relationship between the Φ-OTDR phase signal and external vibration signals, it is used to describe vibration events for classification and identification. Currently, most methods input a one-dimensional feature vector of the phase signal into the classifier. These features are limited to manually defined time-domain or frequency-domain calculations and cannot fully extract the signal characteristics.

[0003] To address this issue, end-to-end deep learning methods, such as convolutional neural networks (CNNs), have been proposed for event recognition in Φ-OTDR. These learning networks from the machine vision field perform well in image processing, but when applied to Φ-OTDR, accumulating data to obtain spatiotemporal images takes a long time. Furthermore, random noise in the optical fiber leads to a huge demand for training samples to ensure the network's accuracy. Therefore, when the operating environment of the Φ-OTDR monitoring system changes, significant time and cost are required to rebuild the dataset and retrain the classification network. While networks such as 1-D CNNs can directly process one-dimensional signals, effective noise reduction methods need to be designed for the target signal during preprocessing, making them difficult to apply universally. Summary of the Invention

[0004] To address the aforementioned issues, this invention employs the following technical solution: a representation learning method for distributed fiber optic fence monitoring, comprising: collecting backscattered Rayleigh scattering signal data returned by the fiber under test under different events; converting the phase time-domain signal at the corresponding location of each spatial acquisition point into a Markov transform field (MTF) matrix; extracting the basis matrix from the MTF through non-negative matrix factorization (NMF) and saving it as an RGB image; and finally performing feature learning and recognition using a 2-D CNN network.

[0005] The specific steps are as follows:

[0006] S1 collects fiber optic fence sensing signals;

[0007] S2 demodulation yields the phase time-domain sequence;

[0008] S3 normalizes the original sequence;

[0009] S4 divides the normalized sequence into Q quantiles according to the range and constructs the nearest neighbor matrix;

[0010] S5 constructs a Markov transition field M ;

[0011] S6 is processed using the nonnegative matrix factorization method. M Obtain the basis matrix;

[0012] S7 saves the basis matrix as an RGB image;

[0013] S8 adds labels to create a dataset, which is then used to train a 2D CNN network.

[0014] S1 includes: fixing the optical fiber under test to a fiber optic fence, and using a Φ-OTDR monitoring system to collect backscattered Rayleigh scattering signal data returned by the optical fiber under test under different events. The signal data has dual-domain information in time and space and is stored in a two-dimensional matrix form. D N×S , N The number of optical pulses injected into the fiber under test, corresponding to time domain information. S This represents the number of sampling points across the entire fiber optic link under test, corresponding to spatial domain information.

[0015] S2 includes: processing a two-dimensional data matrix. D N×S Demodulation is performed to obtain the phase signal at each point, and the position of each spatial sampling point is determined. i A phase-time sequence can be obtained at any point. x i ={x1, x2, …, x n}, n For the corresponding time;

[0016] S3 includes: normalizing the original sequence obtained in S2 using linear function normalization, expressed by the formula:

[0017] .

[0018] S4 includes: dividing the normalized sequence into Q quantiles according to the value range, and each normalized sequence value x nom Only one quantile is corresponding q i , i ∈[1, Q].

[0019] S5 includes: calculating the transition probabilities between different quantiles along the time axis to construct a proximity matrix.W Matrix elements w i,j quantiles q j A point and quantiles q i The frequency of a point adjacent to another point in the spectrum is expressed as:

[0020] ;

[0021] Constructing Markov Transformation Fields M Matrix elements M ij Indicates from quantiles q i Jump to quantile q j The probability is expressed as:

[0022] .

[0023] S6 includes: decomposing the matrix using a nonnegative matrix factorization method. M The matrix is ​​decomposed into two smaller nonnegative matrices, which are the basis matrices. V sum coefficient matrix H The approximate decomposition of the matrix is ​​achieved by solving the following optimization problem:

[0024] .

[0025] S7 includes: obtaining the basis matrix V Save as an RGB color image.

[0026] S8 includes: labeling the corresponding events, and after training and testing the CNN network, classifying the new data using the trained network.

[0027] The beneficial effects of this invention are as follows:

[0028] This invention effectively solves the problem of low learning efficiency of intrusion event recognition models in the field of Φ-OTDR monitoring. By further characterizing the data features before feeding them into the 2-D CNN network, the original phase time series is made more suitable for the network to learn, thereby improving learning efficiency and recognition accuracy.

[0029] This invention effectively solves the problem of high cost in constructing training datasets for recognition networks. In traditional Ф-OTDR recognition technology using images, the spatiotemporal data matrix obtained by the entire sensing link in one detection cycle can often only be used as a sample. However, the phase time series obtained by this invention at different spatial sampling points can be used as a separate sample. That is, in one detection cycle, the maximum number of samples can be the same as the number of spatial sampling points.

[0030] This invention converts one-dimensional time series into two-dimensional images for recognition, which will better align with the current development trend in the field of machine vision and promote the integration of Ф-OTDR recognition technology with future cutting-edge machine vision technologies. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the representation learning method for distributed fiber optic fence monitoring according to the present invention.

[0032] Figure 2 This is a schematic diagram of the fiber optic fence detection structure of the present invention.

[0033] Figure 3 The phase sequence time-domain plots and corresponding MTF and MTF-NMF images for the three events are shown.

[0034] Figure 4 This is a comparison chart of the training loss curves for the three recognition methods.

[0035] Figure 5 The diagram shows the confusion matrix results for the three recognition methods. Detailed Implementation

[0036] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description is provided below with reference to the accompanying drawings and specific embodiments.

[0037] Example 1:

[0038] The representation learning method for distributed fiber optic fence monitoring is as follows: collect backscattered Rayleigh scattering signal data returned by the fiber under test under different events, convert the phase time domain signal at the corresponding location of each spatial acquisition point into a Markov transform field (MTF) matrix, extract the basis matrix from the MTF through non-negative matrix factorization (NMF) and save it as an RGB image, and finally perform feature learning and recognition by a 2-D CNN network.

[0039] Specifically, the following steps are included:

[0040] Step 1: Fix the fiber under test to the fiber optic fence, and use the Φ-OTDR monitoring system to collect the backscattered Rayleigh scattering signal data returned by the fiber under test under different events. The data has dual-domain information in time and space and is stored in the form of a two-dimensional matrix. D N×S , N The number of optical pulses injected into the fiber under test, corresponding to time domain information. S The number of sampling points for the entire fiber optic link under test, corresponding to spatial domain information;

[0041] Step 2: For the two-dimensional data matrix D N×SDemodulation is performed to obtain the phase signal at each point, and the position of each spatial sampling point is determined. i A phase-time sequence can be obtained at any point. x i ={x1, x2, …, x n}, n For the corresponding time;

[0042] Step 3: Normalize the original sequence using a linear function, expressed by the formula:

[0043] ;

[0044] Step 4: Divide the normalized sequence into Q quantiles according to the value range, and each normalized sequence value... x nom Only one quantile is corresponding q i , i ∈[1, Q];

[0045] Step 5: Calculate the transition probabilities between different quantiles along the time axis to construct the proximity matrix. W Matrix elements w i,j quantiles q j A point and quantiles q i The frequency of a point adjacent to another point in the spectrum is expressed as:

[0046] ;

[0047] Constructing Markov Transformation Fields M Matrix elements M ij Indicates from quantiles q i Jump to quantile q j The probability is expressed as:

[0048] ;

[0049] Step 6: Use nonnegative matrix factorization to... M The matrix is ​​decomposed into two smaller nonnegative matrices, which are the basis matrices. V sum coefficient matrix H The approximate decomposition of the matrix is ​​achieved by solving the following optimization problem:

[0050] ;

[0051] Step 7: Obtain the basis matrix VSave the images as RGB color images, label them with corresponding event tags, and create a dataset for training and testing 2D CNN networks.

[0052] Example 2: As Figure 1 As shown, this invention provides a representation learning method suitable for distributed fiber optic fence monitoring, comprising the following steps:

[0053] Step 1: Fix the fiber under test to the fiber optic fence, and use the Φ-OTDR monitoring system to collect the backscattered Rayleigh scattering signal data returned by the fiber under test under different events. The data has dual-domain information in time and space and is stored in the form of a two-dimensional matrix. D N×S , N The number of optical pulses injected into the fiber under test, corresponding to time domain information. S The number of sampling points for the entire fiber optic link under test, corresponding to spatial domain information;

[0054] In this example, the 14.8 km long optical fiber under test was fixed to the fence in an S-shape. The acousto-optic modulator in the Φ-OTDR system had a frequency shift of 80 MHz, the optical pulse emission frequency was set to 1 kHz, and the data acquisition card had a data acquisition rate of 250 MS / s. The data matrix obtained in one test... D The size is 1000×35000;

[0055] In this example, there are four target events to be identified on the fiber optic fence: climbing, tapping, wind, and spurious interference. Climbing events are generated by people climbing the fiber optic fence; tapping events are generated by people tapping the fiber optic fence with tools; wind events are collected by placing the fiber optic fence in windy weather; and spurious interference includes various non-threat events, such as system background noise, pedestrian noise, and rain noise. The dataset structure is shown in Table 1.

[0056] Table 1

[0057]

[0058] Step 2: For the two-dimensional data matrix D N×S Demodulation is performed to obtain the phase signal at each point, and the position of each spatial sampling point is determined. i A phase-time sequence can be obtained at any point. x i ={x1, x2, …, x n}, n For the corresponding time;

[0059] Step 3: Normalize the original sequence using a linear function, expressed by the formula:

[0060] ;

[0061] In this example, a linear transformation is performed on the original data to map its range to [0, 1]. The time-domain plots of the normalized phase sequences of the three events—climbing, knocking, and wind blowing—are shown below. Figure 3 As shown in (a).

[0062] Step 4: Divide the normalized sequence into Q quantiles according to the value range, and each normalized sequence value... x nom Only one quantile is corresponding q i , i ∈[1, Q];

[0063] Step 5: Calculate the transition probabilities between different quantiles along the time axis to construct the proximity matrix. W Matrix elements w i,j quantiles q j A point and quantiles q i The frequency of a point adjacent to another point in the spectrum is expressed as:

[0064] ;

[0065] Constructing Markov Transformation Fields M Matrix elements M ij Indicates from quantiles q i Jump to quantile q j The probability is expressed as:

[0066] ;

[0067] In this example, the MTF matrix does not need to be saved as an image in the process of this invention; subsequent calculations only process the obtained matrix data. However, for comparison and to highlight the advantages of the MTF-NMF method, the MTF matrix is ​​saved as a PNG file with a resolution of 77×77 pixels for network classification testing. The MTF images for the three events of climbing, knocking, and wind blowing are shown below. Figure 3 As shown in (b).

[0068] Step 6: Use nonnegative matrix factorization to... M The matrix is ​​decomposed into two smaller nonnegative matrices, which are the basis matrices. V sum coefficient matrix H The approximate decomposition of the matrix is ​​achieved by solving the following optimization problem:

[0069] ;

[0070] Step 7: Obtain the basis matrix V Save the images as RGB color images, label them with corresponding event tags, and create a dataset for training and testing 2D CNN networks.

[0071] In this example, the saved basis matrix V The color image resolution is also 77×77 pixels. The MTF-NMF images corresponding to the three events of climbing, knocking, and wind blowing are as follows: Figure 3 As shown in (c).

[0072] In this example, the one-dimensional phase time series will be fed into a 1-D CNN and the images obtained by the two methods will be fed into a 2-D CNN for comparison. All three recognition methods use the same training and test sets. The learning rate is optimized by the Adam algorithm and the training error is evaluated by the cross-entropy cost function. The parameters of the two CNN network structures used are shown in Table 2.

[0073] In this example, the training loss curves and test confusion matrices of the three methods are shown in the following figures. Figure 4 and Figure 5 As shown, the training convergence speed of the method provided by this invention is much higher than that of the other two methods, and the test recognition rate for the four types of events reaches 100%, making it suitable for the recognition and detection of fiber optic fence events.

[0074] Table 2

[0075]

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A representation learning method for distributed fiber optic fence monitoring, characterized in that: Backscattered Rayleigh scattering signal data returned by the optical fiber under test under different events are collected. The phase time domain signal at the corresponding position of each spatial acquisition point is converted into a Markov transformed field MTF matrix. Then, the basis matrix is ​​extracted from the MTF by non-negative matrix factorization (NMF) and saved as an RGB image. Finally, a 2-D CNN network is used for feature learning and recognition. The specific steps include: S1 collects fiber optic fence sensing signals; S2 demodulation yields the phase time-domain sequence; S3 normalizes the original sequence; S4 divides the normalized sequence into Q quantiles according to the range and constructs the nearest neighbor matrix; S5 constructs a Markov transition field M ; S6 is processed using the nonnegative matrix factorization method. M Obtain the basis matrix; S7 saves the basis matrix as an RGB image; S8 adds labels to create a dataset, which is then input into the 2D CNN network for training. S1 further includes: fixing the optical fiber under test to the optical fiber fence, and using a Φ-OTDR monitoring system to collect backscattered Rayleigh scattering signal data returned by the optical fiber under test under different events. The signal data has dual-domain information in time and space and is stored in a two-dimensional matrix form. D N×S , N The number of optical pulses injected into the fiber under test, corresponding to time-domain information. S The number of sampling points for the entire fiber optic link under test, corresponding to spatial domain information; S2 further includes: processing the two-dimensional data matrix D N×S Demodulation is performed to obtain the phase signal at each point, and the position of each spatial sampling point is determined. i A phase-time sequence can be obtained at any point. x i ={x1, x2, …, x n }, n This corresponds to the time.

2. The representation learning method for distributed fiber optic fence monitoring according to claim 1, characterized in that, S3 further includes: normalizing the original sequence obtained in S2 using linear function normalization, expressed by the formula: 。 3. The representation learning method for distributed fiber optic fence monitoring according to claim 2, characterized in that, S4 further includes: dividing the normalized sequence into Q quantiles according to the value range, and each normalized sequence value x nom Only one quantile is corresponding q i , i ∈[1, Q].

4. The representation learning method for distributed fiber optic fence monitoring according to claim 3, characterized in that, S5 further includes: calculating the transition probabilities between different quantiles along the time axis to construct a proximity matrix. W Matrix elements w i,j quantiles q j A point and quantiles q i The frequency of a point adjacent to another point in the spectrum is expressed as: ； Constructing Markov Transformation Fields M Matrix elements M ij Indicates from quantiles q i Jump to quantile q j The probability is expressed as: 。 5. The representation learning method for distributed fiber optic fence monitoring according to claim 4, characterized in that, S6 further includes: decomposing the matrix using a nonnegative matrix factorization method. M The matrix is ​​decomposed into two smaller nonnegative matrices, which are the basis matrices. V sum coefficient matrix H The approximate decomposition of the matrix is ​​achieved by solving the following optimization problem: 。 6. The representation learning method for distributed fiber optic fence monitoring according to claim 1, characterized in that, The S8 also includes: four types of labels, namely climbing, knocking, wind blowing, and false interference. After training and testing the CNN network, the trained network is used to classify new data.

Images