1. A fiber optic early warning system event classification method based on transfer learning is characterized by comprising the following steps:

step 1: detecting a Rayleigh scattering light intensity signal obtained when a vibration event occurs through an optical fiber early warning system;

step 2: making the Rayleigh scattering light intensity signal into a vibration event data sample;

and step 3: and inputting the vibration event data sample into a deep convolution neural network obtained based on transfer learning to obtain a classification result.

2. The method for classifying events of a fiber optic early warning system based on transfer learning according to claim 1, wherein the fiber optic early warning system comprises: the device comprises a narrow line width light source, a first 50:50 coupler, a frequency shift acousto-optic modulator, a 90:10 coupler, an electro-optic modulator, a first circulator, a filter grating, a first optical amplifier, a second circulator, a sensing optical fiber, a pulse modulation acousto-optic modulator, a second optical amplifier, a third circulator, a second 50:50 coupler, an acquisition card and an upper computer;

the rayleigh scattering light intensity signal obtained when the vibration event is detected by the optical fiber early warning system comprises: the sensing optical fiber senses a vibration event, the narrow-linewidth light source generates continuous light, the continuous light is divided into probe light and pump light after passing through a first 50:50 coupler, the probe light is modulated into pulse light through the pulse modulation acousto-optic modulator, and a frequency shift f is generated₁Then, after light amplification is carried out by the second optical amplifier, the light is injected from the head end of the sensing optical fiber through the third circulator, and backward Rayleigh scattering light is generated in the sensing optical fiber; the back rayleigh scattered light reaches the second 50:50 coupler via the third circulator;

the pump light generates a frequency shift f through a frequency shift acousto-optic modulator₂Then, 10% of the pump light is separated by a 90:10 coupler to serve as local oscillation light, and the local oscillation light reaches a second 50:50 coupler;

the remaining 90% of the pump light reaches the electro-optical modulator and carries out double-sideband frequency shift, the pump light after the double-sideband frequency shift passes through the first circulator and the filter grating and then filters an upper sideband, then the pump light is injected from the tail end of the sensing optical fiber through the second circulator after being amplified by the second optical amplifier, and energy migration is generated between the pump light and pulse light in the sensing optical fiber, so that distributed optical amplification of the pulse light is completed;

the local oscillator light and the Rayleigh scattered light generate interference at a second 50:50 coupler to form carrier wave with beat frequency of | f₁-f₂Acquiring the interference signal by an acquisition card, and completing digital orthogonal demodulation in an upper computer to obtain a Rayleigh scattering light intensity signal;

the rayleigh scattered light intensity signal comprises a vibration signal to be detected.

3. The method for classifying the events of the optical fiber early warning system based on the transfer learning as claimed in claim 1, wherein the step of inputting the vibration event data samples into a deep convolutional neural network obtained based on the transfer learning to obtain a classification result comprises:

step S1: acquiring a trained deep convolutional neural network which is trained on an ImageNet data set and has the Top5 accuracy rate of more than 80%, wherein the deep convolutional neural network is marked as Pretrailing _ AlexNet;

step S2: constructing a training Set Train _ Set and a Test Set Test _ Set according to the vibration event data samples, wherein the vibration time data samples are divided into n event types;

step S3: freezing the first 15 layers of the predrained _ AlexNet in step S1, and changing the 23 th and 25 th layers of the predrained _ AlexNet to the full connection layer and the n-class output layer;

step S4: training the frozen Pretrained _ AlexNet in step S3 by using the training Set Train _ Set obtained in step S2, converging, and storing locally to obtain a trained deep convolutional neural network and all weight parameters thereof, where the deep convolutional neural network is recorded as: local _ Trained _ AlexNet;

step S5: inputting sample data in the Test Set Test _ Set into a locally Trained deep convolutional neural network Local _ Trained _ AlexNet to obtain a classification result and evaluation of the recognition capability of the deep convolutional neural network, if the evaluation effect reaches the standard, storing the deep convolutional neural network, if the evaluation effect does not reach the standard, repeating the step S3 and the step S4, and changing the freezing proportion in the step S3 or the training times in the step S4 until the optimal classification effect is obtained;

step S6: and (5) inputting the deep convolutional neural network obtained in the step S5 by using the method in the step S2 as a vibration event data sample to obtain a classification result.

4. The method for classifying events of an optical fiber early warning system based on transfer learning of claim 1, wherein the step of making the rayleigh scattering optical intensity signal into a vibration event data sample comprises:

step S21: obtaining Rayleigh scattered light intensity signals when a vibration event occurs by using an optical fiber early warning system, wherein the ith Rayleigh scattered light intensity signal is recorded as TR_i；

Step S22: rearranging a plurality of Rayleigh scattered light intensity signals obtained in 1s into a Matrix form to obtain a space-time Matrix signal, wherein the space-time Matrix signal is marked as TR _ Matrix;

step S23: positioning the vibration event by using a moving differential average method, taking the positioning position of the vibration event as a center and 25m as a radius, extracting data of a corresponding position in a space-time Matrix signal TR _ Matrix, and recording the data as TR _ Pick;

step S24: carrying out high-pass filtering on the column direction of the space-time matrix signal TR _ Pick, wherein the cut-off frequency is 5Hz, and filtering out direct-current components;

step S25: graying the whole space-time matrix signal TR _ Pick filtered in the step S24;

step S26: and adjusting the row number of the TR _ Pick by adopting a down-sampling method, adjusting the column number of the TR _ Pick by adopting an interpolation method, adjusting the channel number of the TR _ Pick by adopting a copying method, and synthesizing into a data type adaptive to the deep convolutional neural network to obtain a vibration event data sample.

5. The method for classifying events of an optical fiber early warning system based on transfer learning of claim 4, wherein the specific method for graying the whole spatio-temporal matrix signal TR _ Pick filtered in the step S24 in the step S25 includes: and (3) taking the maximum value and the minimum value of the space-time matrix signal TR _ Pick, respectively recording the maximum value and the minimum value as Max _ TR and Min _ TR, and recording the graying operation as:

TR_Pick_gray＝floor[255×(TR_Pick-Min_TR)/(Max_TR-Min_TR)]；

wherein, TR _ Pick _ gray is the space-time matrix signal after graying, and floor [. cndot ] is the operation of rounding down.

6. The method for classifying events of an optical fiber early warning system based on transfer learning of claim 3, wherein the method for deep convolutional neural network training in step S4 specifically comprises:

step S41, inputting the vibration event data sample j in the training Set Train _ Set into the partially frozen deep convolution neural network Pretrained _ AlexNet to complete the forward propagation process and obtain the classification outputProbability of occurrence and classification p_j；

Step S42, calculating a loss function value by a cross entropy loss function according to the classification output obtained in step S41 and the label of the input vibration event data sample, the formula is,

wherein L is a cross entropy loss function value, N is a total number of samples, M is a total number of categories, and p_jcIs the probability, y, that a vibration event data sample j belongs to the class c_jcWhen the category of the vibration event data sample j is the same as the category c, the value is 1, and the other values are 0;

step S43, updating the unfrozen weight of the deep convolutional neural network by using an Adam back propagation gradient descent algorithm according to the loss function value obtained in the step S42;

the weight updating method during the t-th training comprises the following steps:

m_t＝β₁m_t-1+(1-β₁)g_t；

where θ represents the unfrozen weight of the deep convolutional neural network, and its subscript is expressed asThe weight value after the t-th training,the partial derivative, beta, of the function over theta₁,β₂E [0,1) is set to be beta₁＝0.9，β₂＝0.1；

The weight value updating method is as follows:

where α is a learning rate, α is set to 0.001, ξ is a constant, and ξ is set to 10^-6。

Background

The distributed optical fiber sensing system is widely applied to the fields of long-distance oil and gas pipeline monitoring, perimeter security protection, building structure health monitoring and the like due to the characteristics of high sensitivity, high positioning accuracy, electromagnetic immunity and the like. The Phase Sensitive Optical Time Domain Reflectometer (phi-OTDR) technology detects the coherent result of Optical pulse return light by using a long coherent light source, and the interference method can effectively realize dynamic response, can simultaneously realize high positioning precision and high sensitivity detection, and is particularly suitable for pipeline early warning for the detection of weak disturbance signals.

The general phi-OTDR sensing distance is about 20km, and the sensing length requirement of a long-distance pipeline of dozens of kilometers is not met. The introduction of the distributed optical amplification technology greatly improves the sensing distance of the phi-OTDR. However, in order to ensure that the signal-to-noise ratio of the rayleigh scattered light does not drop significantly, a heterodyne detection structure is required before a/D sampling. The introduction of heterodyne detection results in a high beat frequency (hundreds of MHz) of the rayleigh scattered light signal, requiring the use of a spectrum analyzer or analog frequency conversion circuitry, making the a/D sampling process very complex and expensive.

On the other hand, due to the characteristic of phi-OTDR qualitative measurement, event identification is difficult to realize, and the false alarm rate is high. The traditional event classification technology is based on manual feature analysis, and is difficult to select universal features suitable for various environments. The introduction of deep learning technology, especially Convolutional Neural Networks (CNN), automatically obtains event features through machine learning, and solves the problem of feature selection. However, deep learning requires a large amount of training data and a high-performance graphics card as a hardware basis for training a neural network. In practical application, it is often difficult to obtain a large amount of sample data in a new environment, and the field generally does not have the neural network training condition of the high-performance display card. Therefore, there is a need for a method that can accomplish the task of event classification with a small data sample and on the basis of an insuffient computational resource.

Disclosure of Invention

The invention aims to provide a method for classifying events of an optical fiber early warning system based on transfer learning, which is used for solving one or more technical problems in the prior art and at least providing a beneficial selection or creation condition.

The solution of the invention for solving the technical problem is as follows: the event classification method of the optical fiber early warning system based on the transfer learning is provided, and comprises the following steps:

step 1: detecting a Rayleigh scattering light intensity signal obtained when a vibration event occurs through an optical fiber early warning system;

step 2: making the Rayleigh scattering light intensity signal into a vibration event data sample;

and step 3: and inputting the vibration event data sample into a deep convolution neural network obtained based on transfer learning to obtain a classification result.

Further, the optical fiber early warning system comprises: the device comprises a narrow line width light source, a first 50:50 coupler, a frequency shift acousto-optic modulator, a 90:10 coupler, an electro-optic modulator, a first circulator, a filter grating, a first optical amplifier, a second circulator, a sensing optical fiber, a pulse modulation acousto-optic modulator, a second optical amplifier, a third circulator, a second 50:50 coupler, an acquisition card and an upper computer;

the rayleigh scattering light intensity signal obtained when the vibration event is detected by the optical fiber early warning system comprises: the sensing optical fiber senses a vibration event, the narrow-linewidth light source generates continuous light, the continuous light is divided into probe light and pump light after passing through a first 50:50 coupler, the probe light is modulated into pulse light through the pulse modulation acousto-optic modulator, and a frequency shift f is generated₁Then, after light amplification is carried out by the second optical amplifier, the light is injected from the head end of the sensing optical fiber through the third circulator, and backward Rayleigh scattering light is generated in the sensing optical fiber; the back rayleigh scattered light reaches the second 50:50 coupler via the third circulator;

the pump light is acousto-optic modulated via frequency shiftA frequency shifter for generating a frequency shift f₂Then, 10% of the pump light is separated by a 90:10 coupler to serve as local oscillation light, and the local oscillation light reaches a second 50:50 coupler;

the remaining 90% of the pump light reaches the electro-optical modulator and carries out double-sideband frequency shift, the pump light after the double-sideband frequency shift passes through the first circulator and the filter grating and then filters an upper sideband, then the pump light is injected from the tail end of the sensing optical fiber through the second circulator after being amplified by the second optical amplifier, and energy migration is generated between the pump light and pulse light in the sensing optical fiber, so that distributed optical amplification of the pulse light is completed;

the local oscillator light and the Rayleigh scattered light generate interference at a second 50:50 coupler to form carrier wave with beat frequency of | f₁-f₂Acquiring the interference signal by an acquisition card, and completing digital orthogonal demodulation in an upper computer to obtain a Rayleigh scattering light intensity signal;

the rayleigh scattered light intensity signal comprises a vibration signal to be detected.

Further, the inputting the vibration event data samples into a deep convolutional neural network obtained based on transfer learning, and obtaining a classification result includes:

step S1: acquiring a trained deep convolutional neural network which is trained on an ImageNet data set and has the Top5 accuracy rate of more than 80%, wherein the deep convolutional neural network is marked as Pretrailing _ AlexNet;

step S2: constructing a training Set Train _ Set and a Test Set Test _ Set according to the vibration event data samples, wherein the vibration time data samples are divided into n event types;

step S3: freezing the first 15 layers of the predrained _ AlexNet in step S1, and changing the 23 th and 25 th layers of the predrained _ AlexNet to the full connection layer and the n-class output layer;

step S4: training the frozen Pretrained _ AlexNet in step S3 by using the training Set Train _ Set obtained in step S2, converging, and storing locally to obtain a trained deep convolutional neural network and all weight parameters thereof, where the deep convolutional neural network is recorded as: local _ Trained _ AlexNet;

step S5: inputting sample data in the Test Set Test _ Set into a locally Trained deep convolutional neural network Local _ Trained _ AlexNet to obtain a classification result and evaluation of the recognition capability of the deep convolutional neural network, if the evaluation effect reaches the standard, storing the deep convolutional neural network, if the evaluation effect does not reach the standard, repeating the step S3 and the step S4, and changing the freezing proportion in the step S3 or the training times in the step S4 until the optimal classification effect is obtained;

step S6: and (5) inputting the deep convolutional neural network obtained in the step S5 by using the method in the step S2 as a vibration event data sample to obtain a classification result.

Further, the making the rayleigh scattered light intensity signal into a vibration event data sample comprises:

step S21: obtaining Rayleigh scattered light intensity signals when a vibration event occurs by using an optical fiber early warning system, wherein the ith Rayleigh scattered light intensity signal is recorded as TR_i；

Step S22: rearranging a plurality of Rayleigh scattered light intensity signals obtained in 1s into a Matrix form to obtain a space-time Matrix signal, wherein the space-time Matrix signal is marked as TR _ Matrix;

step S23: positioning the vibration event by using a moving differential average method, taking the positioning position of the vibration event as a center and 25m as a radius, extracting data of a corresponding position in a space-time Matrix signal TR _ Matrix, and recording the data as TR _ Pick;

step S24: carrying out high-pass filtering on the column direction of the space-time matrix signal TR _ Pick, wherein the cut-off frequency is 5Hz, and filtering out direct-current components;

step S25: graying the whole space-time matrix signal TR _ Pick filtered in the step S24;

step S26: and adjusting the row number of the TR _ Pick by adopting a down-sampling method, adjusting the column number of the TR _ Pick by adopting an interpolation method, adjusting the channel number of the TR _ Pick by adopting a copying method, and synthesizing into a data type adaptive to the deep convolutional neural network to obtain a vibration event data sample.

Further, the specific method for graying the entire spatio-temporal matrix signal TR _ Pick filtered in step S24 in step S25 includes: and (3) taking the maximum value and the minimum value of the space-time matrix signal TR _ Pick, respectively recording the maximum value and the minimum value as Max _ TR and Min _ TR, and recording the graying operation as:

TR_Pick_gray＝floor[255×(TR_Pick-Min_TR)/(Max_TR-Min_TR)]；

wherein, TR _ Pick _ gray is the space-time matrix signal after graying, and floor [. cndot ] is the operation of rounding down.

Further, the method for deep convolutional neural network training in step S4 specifically includes:

step S41, inputting the vibration event data sample j in the training Set Train _ Set into the partially frozen deep convolution neural network Pretrailing _ AlexNet, completing the forward propagation process, and obtaining the classification output and the classification probability p_j；

Step S42, calculating a loss function value by a cross entropy loss function according to the classification output obtained in step S41 and the label of the input vibration event data sample, the formula is,

wherein L is a cross entropy loss function value, N is a total number of samples, M is a total number of categories, and p_jcIs the probability, y, that a vibration event data sample j belongs to the class c_jcWhen the category of the vibration event data sample j is the same as the category c, the value is 1, and the other values are 0;

step S43, updating the unfrozen weight of the deep convolutional neural network by using an Adam back propagation gradient descent algorithm according to the loss function value obtained in the step S42;

the weight updating method during the t-th training comprises the following steps:

m_t＝β₁m_t-1+(1-β₁)g_t；

wherein theta represents the unfrozen weight of the deep convolutional neural network, the subscript of theta represents the weight after the t training,the partial derivative, beta, of the function over theta₁,β₂E [0,1) is set to be beta₁＝0.9，β₂＝0.1；

The weight value updating method is as follows:

where α is a learning rate, α is set to 0.001, ξ is a constant, and ξ is set to 10^-6。

The invention has the beneficial effects that: the invention can finish the event classification task on the basis of smaller data samples and portable computer conditions, and is suitable for the event identification task of the optical fiber early warning system applied to a new installation environment. The invention is mainly used in the technical field of optical fiber early warning.

Drawings

In order to more clearly illustrate the technical solution in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly described below. It is clear that the described figures are only some embodiments of the invention, not all embodiments, and that a person skilled in the art can also derive other designs and figures from them without inventive effort.

FIG. 1 is a schematic diagram of a fiber optic warning system;

FIG. 2 is a flowchart of the general steps to obtain classification results through a deep convolutional neural network based on transfer learning;

FIG. 3 is a schematic diagram of the structure of a deep convolutional neural network in an embodiment of use of the invention;

FIG. 4 is a sample schematic of vibration event data obtained by the method of the present invention;

in fig. 4, I is a sunny data sample, II is a rainy data sample, III is a human walking data sample, IV is a human jumping data sample, V is a water-flushing data sample, VI is a shoveling land data sample, VII is a shoveling land data sample, and VIII is a bicycle passing data sample.

The components represented by the various reference numerals in fig. 1 are as follows:

1. a narrow line width light source; 2. a first 50:50 coupler; 3. a frequency-shift acousto-optic modulator; 4. a 90:10 coupler; 5. an electro-optic modulator; 6. a first circulator; 7. a filter grating; 8. a first optical amplifier; 9. a second circulator; 10. a sensing optical fiber; 11. a pulse modulated acousto-optic modulator; 12. a second optical amplifier; 13. a third circulator; 14. a second 50:50 coupler; 15. collecting cards; 16. and (4) an upper computer.

Detailed Description

The conception, the specific structure, and the technical effects produced by the present invention will be clearly and completely described below in conjunction with the embodiments and the accompanying drawings to fully understand the objects, the features, and the effects of the present invention. It is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all embodiments, and those skilled in the art can obtain other embodiments without inventive effort based on the embodiments of the present invention, and all embodiments are within the protection scope of the present invention. In addition, all the coupling/connection relationships mentioned herein do not mean that the components are directly connected, but mean that a better coupling structure can be formed by adding or reducing coupling accessories according to specific implementation conditions. All technical characteristics in the invention can be interactively combined on the premise of not conflicting with each other.

Embodiment 1, with reference to fig. 1 and fig. 2, provides a fiber optic early warning system event classification method based on transfer learning, including:

step 1: detecting a Rayleigh scattering light intensity signal obtained when a vibration event occurs through an optical fiber early warning system;

step 2: making the Rayleigh scattering light intensity signal into a vibration event data sample;

and step 3: and inputting the vibration event data sample into a deep convolution neural network obtained based on transfer learning to obtain a classification result.

Wherein, optic fibre early warning system includes: the device comprises a narrow line width light source 1, a first 50:50 coupler 2, a frequency shift acousto-optic modulator 3, a 90:10 coupler 4, an electro-optic modulator 5, a first circulator 6, a filter grating 7, a first optical amplifier 8, a second circulator 9, a sensing optical fiber 10, a pulse modulation acousto-optic modulator 11, a second optical amplifier 12, a third circulator 13, a second 50:50 coupler 14, an acquisition card 15 and an upper computer 16;

the rayleigh scattering light intensity signal obtained when the vibration event is detected by the optical fiber early warning system comprises: the sensing fiber 10 senses a vibration event, and the narrow linewidth light source 1 generates continuous light which is divided into probe light and pump light after passing through the first 50:50 coupler 2.

The probe light is modulated into pulse light of 200ns width by the pulse modulation acousto-optic modulator 11, and frequency shift f of MHz level is generated₁For example 200 MHz. And then is optically amplified by the second optical amplifier 12 and injected from the head end of the sensing fiber 10 through the third circulator 13. Wherein the frequency shift f₁The amount of frequency shift is determined by the pulse-modulated acousto-optic modulator 11, which is 200MHz in this embodiment.

The pulsed light generates back rayleigh scattered light in the sensing fiber 10; the back rayleigh scattered light reaches the second 50:50 coupler 14 via the third circulator 13;

the pump light is passed through a frequency-shifting acousto-optic modulator 3 to generate another frequency shift f of MHz level₂E.g., 180MHz, and then separates 10% of the pump light via the 90:10 coupler 4 as local oscillator light, which reaches the second 50:50 coupler 14. Wherein, frequencyMoving f₂The amount of frequency shift is determined by the frequency-shift acousto-optic modulator 3, which is 180MHz in this embodiment.

The remaining 90% of the pump light reaches the electro-optical modulator 5 and undergoes a double sideband frequency shift, resulting in a 10.8GHz double sideband frequency shift, the specific value of which is determined by the brillouin frequency shift value of the sensing fiber 10.

The pump light after the double-side band frequency shift passes through the first circulator 6 and the filter grating 7, then the upper side band is filtered out, the optical power of the pump light is amplified through the second optical amplifier 12, the pump light is injected from the tail end of the sensing optical fiber 10 through the second circulator 9, energy migration is generated between the pump light and the pulse light in the sensing optical fiber 10, and distributed optical amplification of the pulse light is completed. This optic fibre early warning system can promote effective sensing distance through like this. Wherein, the third port of the second circulator 9 is suspended to block the forward propagating detection light.

The local oscillator light and the Rayleigh scattered light interfere at a second 50:50 coupler 14 to form a carrier beat frequency of | f₁-f₂The carrier of this embodiment is 20 MHz.

The interference signals are acquired by an acquisition card 15, the sampling rate of the acquisition card 15 is 100MHz, then the data are sent to an upper computer 16, and digital quadrature demodulation is completed in the upper computer 16. In this embodiment, a sine wave and a cosine wave with a frequency of 20MHz are used to complete orthogonal demodulation, and a rayleigh scattered light intensity signal TR is obtained.

Wherein, the inputting the vibration event data sample into a deep convolutional neural network obtained based on transfer learning, and the obtaining of the classification result comprises:

step S1: acquiring a trained deep convolutional neural network which is trained on an ImageNet data set and has a Top5 accuracy rate of more than 80%, wherein if the trained AlexNet is used, the size of an input layer is 277 multiplied by 3, and the deep convolutional neural network is marked as Pretrailing _ AlexNet; the structure of which is shown in fig. 3.

Step S2: constructing a training Set Train _ Set and a Test Set Test _ Set according to the vibration event data samples, wherein the vibration time data samples are divided into n event types; the specific embodiment takes 8 event categories as examples, and specifically includes: the 8 event types are 1200 vibration event data samples in total, and the types are respectively sunny data samples, rainy data samples, human walking data samples, human jumping data samples, water flushing data samples, land shoveling data samples and bicycle passing data samples. Preprocessing the vibration event data samples, changing the size of the vibration event data samples to 277 multiplied by 3, randomly selecting 1000 vibration event data samples to construct a training Set Train _ Set, and constructing a Test Set Test _ Set by the remaining 200 vibration event data samples.

Step S3: freezing the first 15 layers of the predrained _ AlexNet in step S1, and changing the 23 th and 25 th layers of the predrained _ AlexNet to the full connection layer and the n-class output layer; in the present embodiment, n is equal to 8.

Step S4: training the frozen Pretrained _ AlexNet in step S3 by using the training Set Train _ Set obtained in step S2, converging, and storing locally to obtain a trained deep convolutional neural network and all weight parameters thereof, where the deep convolutional neural network is recorded as: local _ Trained _ AlexNet;

step S5: inputting sample data in the Test Set Test _ Set into a locally Trained deep convolutional neural network Local _ Trained _ AlexNet to obtain a classification result and evaluation of the recognition capability of the deep convolutional neural network, if the evaluation effect reaches the standard, storing the deep convolutional neural network, if the evaluation effect does not reach the standard, repeating the step S3 and the step S4, and changing the freezing proportion in the step S3 or the training times in the step S4 until the optimal classification effect is obtained;

step S6: and (5) inputting the deep convolutional neural network obtained in the step S5 by using the method in the step S2 as a vibration event data sample to obtain a classification result.

Wherein the making of the rayleigh scattered light intensity signal into a vibration event data sample comprises:

step S21: obtaining Rayleigh scattered light intensity signals when a vibration event occurs by using an optical fiber early warning system, wherein the ith Rayleigh scattered light intensity signal is recorded as TR_i；

Step S22: multiple Rayleigh scattered lights obtained within 1sRearranging the light intensity signals into a Matrix form to obtain a space-time Matrix signal, and recording the space-time Matrix signal as TR _ Matrix; assuming that the repetition frequency of the detection light pulse of the optical fiber early warning system is 1kHz, 1000 rayleigh scattered light intensity signals are obtained in 1s, and the formed space-time Matrix signal is recorded as TR _ Matrix [ TR ] Matrix ═ TR₁,TR₂,...,TR₁₀₀₀]^TAt this time, the row dimension of the space-time matrix signal represents the sensing space dimension, and the column dimension represents the time sequence of the rayleigh scattered light intensity signal.

Step S23: positioning the vibration event by using a moving differential average method, taking the positioning position of the vibration event as a center and 25m as a radius, extracting data of a corresponding position in a space-time Matrix signal TR _ Matrix, and recording the data as TR _ Pick; the size of TR _ Pick at this time is 1000 × 50.

Step S24: carrying out high-pass filtering on the column direction of the space-time matrix signal TR _ Pick, wherein the cut-off frequency is 5Hz, and filtering out direct-current components;

step S25: graying the whole space-time matrix signal TR _ Pick filtered in the step S24;

step S26: and adjusting the row number of the TR _ Pick by adopting a down-sampling method, adjusting the column number of the TR _ Pick by adopting an interpolation method, adjusting the channel number of the TR _ Pick by adopting a copying method, and synthesizing into a data type adaptive to the deep convolutional neural network to obtain a vibration event data sample. Specifically, this embodiment is: the number of rows of TR _ Pick is adjusted to 277 using a down-sampling method, the number of columns of TR _ Pick is adjusted to 277 using an interpolation method, TR _ Pick is copied three times using a copying method to form 3 channels, and vibration application data samples having a total size of 277 × 277 × 3 are synthesized while recording the type of the vibration event as a label of the vibration event data samples, and the formed vibration event data samples are as shown in fig. 4.

In some preferred embodiments, the specific method for graying the whole spatio-temporal matrix signal TR _ Pick filtered in step S24 in step S25 includes: and (3) taking the maximum value and the minimum value of the space-time matrix signal TR _ Pick, respectively recording the maximum value and the minimum value as Max _ TR and Min _ TR, and recording the graying operation as:

TR_Pick_gray＝floor[255×(TR_Pick-Min_TR)/(Max_TR-Min_TR)]；

wherein, TR _ Pick _ gray is the space-time matrix signal after graying, and floor [. cndot ] is the operation of rounding down.

In some preferred embodiments, the idea for deep convolutional neural network training in step S4 is: the feature extraction capability of the convolutional neural network is guaranteed by utilizing transfer learning, only 1000 vibration event data samples are used, only 2/5 adjustment parameters needed by the convolutional neural network are left, the operation amount is greatly reduced, and the training process can be carried out on a GPU (GTX1050Ti) of a common portable computer.

The method for deep convolutional neural network training in step S4 specifically includes:

step S41, inputting the vibration event data sample j in the training Set Train _ Set into the partially frozen deep convolution neural network Pretrailing _ AlexNet, completing the forward propagation process, and obtaining the classification output and the classification probability p_j；

Step S42, calculating a loss function value by using a cross entropy loss function according to the classification output obtained in step S41 and the label of the input vibration event data sample, where the formula is:

wherein L is a cross entropy loss function value, N is a total number of samples, M is a total number of categories, and p_jcIs the probability, y, that a vibration event data sample j belongs to the class c_jcWhen the category of the vibration event data sample j is the same as the category c, the value is 1, and the other values are 0;

step S43, updating the unfrozen weight of the deep convolutional neural network by using an Adam back propagation gradient descent algorithm according to the loss function value obtained in the step S42;

the weight updating method during the t-th training comprises the following steps:

m_t＝β₁m_t-1+(1-β₁)g_t；

wherein theta represents the unfrozen weight of the deep convolutional neural network, the subscript of theta represents the weight after the t training,the partial derivative, beta, of the function over theta₁,β₂E [0,1) is set to be beta₁＝0.9，β₂＝0.1；

The weight value updating method is as follows:

where α is a learning rate, α is set to 0.001, ξ is a constant, and ξ is set to 10^-6。

Further, the evaluation of the recognition capability of the deep convolutional neural network in step S5 specifically comprises the following steps: using the classification result, respectively calculating the classification Accuracy, Precision, Recall and F1_ score under each event category, taking event category 1 as an example, the formula is as follows:

Accuracy₁＝(TP₁+TN₁)/(TP₁+TN₁+FP₁+FN₁)

Precision₁＝TP₁/(TP₁+FP₁)

Recall₁＝TP₁/(TP₁+FN₁)

F1_score₁＝Precision₁·Recall₁/(Precision₁+Recall₁)

wherein, TP₁The number of samples representing event Category 1 classified as event Category 1, FP₁The number of samples, TN, representing non-event class 1 classified as event class 1₁The number of samples classified as non-event class 1, FN₁The samples representing non-event category 1 are classified as the number of event categories 1.

In this example, when the Accuracy of all categories is set to be greater than 95%, the evaluation result reaches the standard.

While the preferred embodiments of the present invention have been illustrated and described, it will be understood by those skilled in the art that the present invention is not limited to the details of the embodiments shown and described, but is capable of numerous equivalents and substitutions without departing from the spirit of the invention and its scope is defined by the claims appended hereto.