Fiber vibration event classification method based on double data amplification and storage medium

Abstract

The application discloses a fiber vibration event classification method based on double data augmentation and a storage medium, and the method comprises the following steps: acquiring vibration event data samples through a fiber data acquisition system; dividing the data samples into few-sample category data samples and multi-sample category data samples according to event categories; sequentially performing sample augmentation on the few-sample data samples through a time domain transfer data augmentation method and a CycleGAN data augmentation method; the augmented data samples and the original data samples constitute a total training set, a fiber classification model is obtained in combination with a deep convolutional network, and the classification of the current vibration event is realized through the fiber classification model. The application reduces the difficulty of obtaining the fiber vibration event data, increases the number of data samples through the double data augmentation method, has high fidelity of the augmented data and the real data, and can complete the training of the fiber vibration event classification model by using a small amount of data samples, thereby ensuring the model accuracy and reducing the training difficulty.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical fiber early warning, and particularly relates to an optical fiber vibration event classification method based on double data amplification and a storage medium. BACKGROUND

[0002] Distributed optical fiber sensing system is widely used in long-distance oil and gas pipeline monitoring, perimeter security, building structure health monitoring and other fields due to its high sensitivity, high positioning accuracy, electromagnetic immunity and other characteristics. The phase-sensitive optical time domain reflectometer technology detects the coherence result of the returned light of the optical pulse by using a long coherence light source, and the interference method can effectively realize dynamic response and simultaneously realize high positioning accuracy and high sensitivity detection, especially for the detection of weak disturbance signals, and is very suitable for pipeline early warning.

[0003] The sensing distance of a general phase-sensitive optical time domain reflectometer is about 20 kilometers, which does not meet the sensing length requirement of tens of kilometers of long-distance pipeline. At present, distributed amplification technology is introduced to improve the sensing distance of the phase-sensitive optical time domain reflectometer. However, in order to ensure that the signal-to-noise ratio of Rayleigh scattering light is within a normal range, an external detection structure is needed to improve the beat frequency of the Rayleigh scattering light signal before A / D sampling. The cost of the external detection structure is high, which makes the entire A / D sampling process complex and expensive.

[0004] Furthermore, due to the qualitative measurement characteristics of the phase-sensitive optical time domain reflectometer, the phase-sensitive optical time domain reflectometer is difficult to realize the identification of the optical fiber vibration event, resulting in a high false alarm rate of the optical fiber vibration event. The traditional vibration event classification method is usually manual feature analysis, which is difficult to select event general features suitable for various environments, increasing the difficulty of optical fiber event classification. At present, deep learning technology is introduced to obtain event features through deep learning technology. Although the deep learning technology solves the feature selection problem to some extent, the data-driven deep learning method often needs a large number of training samples to train the model to reach the best state.

[0005] Therefore, there is a need in the art for an identification method that reduces the difficulty of data collection and achieves high accuracy of optical fiber event classification using only a small number of training samples. SUMMARY

[0006] The purpose of the present application is to provide an optical fiber vibration event classification method based on double data amplification and a storage medium to solve one or more technical problems existing in the prior art and at least provide a beneficial choice or create conditions.

[0007] The solution to the technical problem of the present application is to provide an optical fiber vibration event classification method based on double data amplification, comprising the following steps:

[0008] A fiber data acquisition system is constructed, and a plurality of Rayleigh scattering light intensity signals of a historical sensing optical fiber when a vibration event occurs are detected through the fiber data acquisition system;

[0009] A plurality of Rayleigh scattering light intensity signals are made into vibration event data samples, the vibration event data samples include RGB images and TS time domain data, and the vibration event data samples are divided into few-sample category data samples and many-sample category data samples;

[0010] Wherein, the collection difficulty of the few-sample category data samples is higher than that of the many-sample category data samples, and the many-sample category data samples include vibration event data samples belonging to a background noise class;

[0011] According to the vibration event data samples of the background noise class, and in combination with a time domain transfer-based data augmentation method, TS time domain data in the few-sample category data samples is augmented to obtain augmented TS time domain data;

[0012] According to a specific data sample, the augmented TS time domain data is augmented through a cyclic generative adversarial network to obtain augmented data samples, the augmented data samples are converted into an RGB format and added to the vibration event data samples to obtain a total training set;

[0013] Wherein, the specific data sample is a vibration event data sample not belonging to the background noise class in the many-sample category data samples;

[0014] A pre-trained deep convolutional network is trained, the total training set is input into the pre-trained deep convolutional network and trained to obtain an optical fiber classification model;

[0015] A current Rayleigh scattering light intensity signal of a sensing optical fiber currently occurring a vibration event is detected and input into the optical fiber classification model to obtain a classification result of the vibration event.

[0016] In another aspect, the application also provides a storage medium having processor-executable instructions stored therein, the processor-executable instructions being used to execute the fiber vibration event classification method based on double data augmentation when executed by a processor.

[0017] The application has the beneficial effects that a fiber vibration event classification method and a storage medium based on double data augmentation are provided, the method includes reducing the difficulty of obtaining fiber vibration event data samples by a fiber data acquisition system, and the number of data samples is increased by a double data augmentation method for filling and augmenting the relatively small number of data samples, the number of samples for training the classification model is increased, and finally the augmented data samples and the original data samples are used to train the classification model, and then the trained classification model is used to realize the classification of the fiber vibration event. The application reduces the difficulty of obtaining fiber vibration event data samples, increases the number of data samples by a double data augmentation method, the augmented data has a high degree of simulation with the real data, and only a small number of data samples are needed to complete the training of the fiber vibration event classification model, which ensures the accuracy of the model while reducing the training difficulty, and is suitable for small sample learning and some event recognition scenarios where training samples are difficult to collect. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A structural schematic diagram of a fiber data acquisition system provided by an embodiment of the application is provided.

[0019] Figure 2 A method flowchart of a fiber vibration event classification method based on double data augmentation provided by an embodiment of the application is provided.

[0020] Figure 3 A structural schematic diagram of a data augmentation method based on a cyclic generative adversarial network provided by an embodiment of the application is provided.

[0021] Figure 4 A schematic diagram of generating data samples of human jumping from data samples of human walking provided by an embodiment of the application is provided.

[0022] Figure 5 A schematic diagram of generating data samples of shovel hitting the ground from data samples of shovel digging the ground provided by an embodiment of the application is provided.

[0023] Figure 6 A flowchart of a fiber vibration event classification method based on double data augmentation provided by an embodiment of the application is provided. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical scheme and advantages of the application clearer, the application is further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application.

[0025] The present application will be further described below in conjunction with the accompanying drawings and specific embodiments. The described embodiments should not be considered limiting to the present application, all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present application.

[0026] In the following description, "some embodiments" are related to a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0028] Before the embodiments of the present application are further described in detail, the terms and phrases involved in the embodiments of the present application are explained, which are applicable to the following explanations.

[0029] (1) CycleGan, an improved network of adversarial network. In the field of computer vision, it is difficult to find a matching high-quality image as a target for deep learning model learning. For example, in the field of super-resolution, for a low-resolution object image, it is not necessarily possible to find a high-resolution image of the same scene. This results in limited adaptability of deep learning model. CycleGan can solve the above problems. CycleGan aims to form a universal mapping from data domain A to data domain B, and the learning goal is the transformation between the styles of data domain A and B rather than the one-to-one mapping relationship between specific data a and b, which reduces the dependence on data matching and improves the adaptability of deep learning model.

[0030] (2) Rayleigh scattering, an optical phenomenon, a case of scattering. When the particle size is much smaller than the wavelength of incident light, such as the particle size being less than one-tenth of the wavelength, the scattering intensity in each direction is not the same, and the intensity is inversely proportional to the fourth power of the wavelength of incident light. This phenomenon is called Rayleigh scattering.

[0031] Distributed optical fiber sensing system is widely used in long distance oil and gas pipeline monitoring, perimeter security, building structure health monitoring and other fields due to its high sensitivity, high positioning accuracy, electromagnetic immunity and other characteristics. Phase sensitive optical time domain reflectometer (Φ-OTDR) technology uses a long coherence light source to detect the coherent results of the returned light of the optical pulse. Its interference method can effectively realize dynamic response and can realize high positioning accuracy and high sensitivity detection at the same time. Especially for the detection of weak disturbance signals, it is very suitable for pipeline early warning.

[0032] The general Φ-OTDR sensing distance is about 20 kilometers, which does not meet the tens of kilometers of sensing length requirement of long distance pipeline. At present, distributed amplification technology is introduced to improve the sensing distance of Φ-OTDR. However, in order to ensure that the signal-to-noise ratio of Rayleigh scattering light is within a normal range, an external detection structure is needed to improve the beat frequency of Rayleigh scattering light before A / D sampling. The cost of the external detection structure is high, which makes the whole A / D sampling process complex and expensive.

[0033] In addition, due to the qualitative measurement characteristics of Φ-OTDR, it is difficult for phase sensitive optical time domain reflectometer to realize the identification of fiber vibration events, resulting in a high false alarm rate of fiber vibration events. The traditional vibration event classification method is usually manual feature analysis, which is difficult to select event general features suitable for various environments, increasing the difficulty of fiber event classification. At present, deep learning technology such as convolutional neural network (CNN) is introduced to obtain event features through deep learning technology. Although deep learning technology solves the problem of feature selection to some extent, the data-driven deep learning method often needs a large number of training samples to train the model to reach the best state. In addition, in some practical application scenarios, such as detecting events caused by concrete cracks and steel wire breakage, it is very difficult to obtain the acoustic emission signals of concrete cracks and steel wire breakage and collect the samples of vibration events corresponding to this type.

[0034] In view of the problems existing in the prior art, the present application provides a fiber vibration event classification method and a storage medium. According to the difficulty of sample collection, the data samples are divided into a small sample category and a large sample category. The small number of data samples are expanded by a double data expansion method, and the training of the fiber vibration event classification model is completed by using the expanded samples. This method is based on double data expansion and is suitable for small sample learning. It can complete the training of the classification model with a small number of samples and is suitable for high-difficulty sample collection of fiber event identification scenarios.

[0035] The application will be further explained and described with reference to the accompanying drawings and specific embodiments.

[0036] Before introducing the optical fiber vibration event classification method provided by the present application, the optical fiber data acquisition system provided by the present application is described and explained. The function of the optical fiber data acquisition system is to acquire training data for training the optical fiber vibration event classification model. Referring to Figure 1 Figure 1 The structure of the optical fiber data acquisition system provided by the embodiment of the present application is shown in the figure. In one embodiment of the present application, the optical fiber data acquisition system is a distributed optical fiber sensing system, which uses Φ-OTDR technology. Specifically, the optical fiber data acquisition system is composed of a narrow linewidth light source 100, a first coupler 200, a second coupler 300, a circulator 400, a first Raman pump laser 500, a first optical wavelength division multiplexer 600, a sensing optical fiber 700, a second Raman pump laser 800, a second optical wavelength division multiplexer 900, a first frequency shift module 1000, a second frequency shift module 1100, and a data processing module 1200. The first frequency shift module 1000 includes an optical amplifier 1001 and a pulse modulation acousto-optic modulator 1002. The second frequency shift module 1100 includes an adjustable attenuator 1101 and a frequency shift acousto-optic modulator 1102. The data processing module 1200 includes an acquisition card 1201, a balanced detector 1202, and an upper computer 1203.

[0037] The following connection relationships exist in the optical fiber data acquisition system:

[0038] The narrow linewidth light source 100 is connected to the first coupler 200, and the first coupler 200 is connected to the first frequency shift module 1000 and the second frequency shift module 1100. The first frequency shift module 1000 is connected to the circulator 400, and the second frequency shift module 1100 is connected to the second coupler 300. The circulator 400 is connected to the second coupler 300 and the first optical wavelength division multiplexer 600. The first optical wavelength division multiplexer 600 is connected to the first Raman pump laser 500 and the first end of the sensing optical fiber 700, i.e., the head end. The second end of the sensing optical fiber 700, i.e., the tail end, is connected to the second optical wavelength division multiplexer 900. The second optical wavelength division multiplexer 900 is connected to the second Raman pump laser 800. In addition, the second coupler 300 is connected to the balanced detector 1202, the balanced detector 1202 is connected to the acquisition card 1201, and the acquisition card 1201 is connected to the upper computer 1203.

[0039] ​In the first frequency shift module 1000, specifically, the first coupler 200 is connected with a pulse modulation acousto-optic modulator 1002 and a frequency shift acousto-optic modulator 1102. The pulse modulation acousto-optic modulator 1002 is connected with an optical amplifier 1001, and the optical amplifier 1001 is connected with the circulator 400. In the second frequency shift module 1100, specifically, the frequency shift acousto-optic modulator 1102 is connected with an adjustable attenuator 1101, and the adjustable attenuator 1101 is connected with the second coupler 300.

[0040] The functions of the modules in the optical fiber data acquisition system are as follows:

[0041] The narrow linewidth light source 100 is configured to emit continuous light. Optionally, the wavelength of the continuous light is 1550 nm.

[0042] The first coupler 200 is a 90:10 coupler, which is configured to split the continuous light into probe light and local oscillator light. The energy proportion of the probe light is 90%, and the energy proportion of the local oscillator light is 10%.

[0043] In the first frequency shift module 1000, the pulse modulation acousto-optic modulator 1002 is configured to modulate the probe light to generate a certain frequency shift of the probe light, thereby forming pulsed light. The optical amplifier 1001 is configured to amplify the pulsed light. The amplified pulsed light is sent into the first end of the sensing optical fiber 700 through the circulator 400.

[0044] In the second frequency shift module 1100, the frequency shift acousto-optic modulator 1102 is configured to modulate the local oscillator light to generate a certain frequency shift of the local oscillator light. The adjustable attenuator 1101 is configured to set an attenuation rate and perform attenuation processing on the local oscillator light with the frequency shift.

[0045] The first Raman pump laser 500 is configured to emit forward pumping light with a wavelength of 1455 nm.

[0046] The second Raman pump laser 800 is configured to emit backward pumping light with a wavelength of 1455 nm.

[0047] The first optical wavelength division multiplexer 600 is configured to send the forward pumping light into the first end of the sensing optical fiber 700. The forward pumping light is used to perform distributed amplification on the pulsed light in the front half of the sensing optical fiber 700.

[0048] The second optical wavelength division multiplexer 900 is configured to send the backward pumping light into the second end of the sensing optical fiber 700. The backward pumping light is configured to perform distributed amplification on the pulsed light in the back half of the sensing optical fiber 700. Under the action of double distributed amplification, the pulsed light generates backward Rayleigh scattering light in the sensing optical fiber 700.

[0049] The circulator 400 has two functions. First, it directs the pulsed light into the first optical wavelength division multiplexer 600, where it is multiplexed before entering the first end of the sensing fiber 700. Second, it directs the backscattered Rayleigh light to the second coupler 300.

[0050] The second coupler 300 is a 50:50 coupler, its function is to acquire the backscattered Rayleigh light and the frequency-shifted local oscillator light, causing them to interfere. In this embodiment, the energy of the frequency-shifted local oscillator light and the backscattered Rayleigh light each account for 50%.

[0051] The function of the balanced detector 1202 is to convert the interference light generated at the second coupler 300 into an electrical signal.

[0052] The function of the acquisition card 1201 is to acquire electrical signals at a preset sampling rate. Optionally, the sampling rate is 100MHz.

[0053] The function of the host computer 1203 is to perform digital orthogonal decomposition on the electrical signal. Optionally, a sine wave and a cosine wave with a frequency of 20MHz are used to perform orthogonal demodulation.

[0054] Reference Figure 2 , Figure 2 The diagram shows a flowchart of the fiber optic vibration event classification method provided in an embodiment of this application. An embodiment of this application will be described and explained below. The vibration event classification method may include, but is not limited to, the following steps.

[0055] S100, construct an optical fiber data acquisition system, and use the optical fiber data acquisition system to detect several Rayleigh scattering light intensity signals when historical sensing optical fiber vibration events occur.

[0056] It should be noted that, additionally, this application maps fiber vibration events using Rayleigh scattering light intensity signals. The Rayleigh scattering light intensity signals are acquired through a fiber optic data system, which is the system described and illustrated in the above embodiments. The fiber optic data acquisition system of this application is a distributed fiber optic sensing system, capable not only of distributed detection but also of simultaneously monitoring vibration at every spatial location over a long distance.

[0057] S200 generates vibration event data samples from several Rayleigh scattering light intensity signals and divides the vibration event data samples into Minor_Class (few sample classes) and Major_Class (multiple sample classes).

[0058] It should be noted that the vibration event data sample contains an RGB image and TS time domain data. The multi-sample category data sample Major_Class contains vibration event data samples belonging to the background noise category.

[0059] In this step, since the classification method provided in the application is mainly aimed at data samples with low collection difficulty and data samples with high collection difficulty, the vibration event data samples are divided into multi-sample category data samples Major_Class and few-sample category data samples Minor_Class according to the collection difficulty. The collection difficulty of the multi-sample category data sample Major_Class is lower than that of the few-sample category data sample Minor_Class.

[0060] In an embodiment, the few-sample category data sample Minor_Class includes at least one category of vibration event data samples, and the multi-sample category data sample Major_Class includes vibration event data samples of the background noise category and at least one category of vibration event data samples. The category classification of the data sample can actually be understood as the vibration event data sample of the reason causing the optical fiber vibration event. For example, the few-sample category data sample Minor_Class includes the human jumping category and the shovel hitting the ground category, and the multi-sample category data sample Major_Class includes the human walking category and the shovel digging the ground category and the background noise category.

[0061] In another embodiment, the data samples are classified into few-sample categories, background noise categories, and multi-sample categories. That is, the background noise category vibration event data sample is separately branched as a category.

[0062] In fact, the application does not specifically limit the specific categories contained in the few-sample category and multi-sample category data samples and the number of categories contained, which can be determined according to actual conditions. However, it should be noted that the few-sample category and multi-sample category data samples need to be divided according to the collection difficulty, and the vibration event data sample must include data samples belonging to the background noise category.

[0063] In an optional embodiment, the reasons causing the optical fiber vibration event include the five reasons of background noise, human jumping, human walking, shovel hitting the ground, and shovel digging the ground, that is, the optical fiber vibration event includes the optical fiber vibration event caused by background noise, human jumping, human walking, shovel hitting the ground, and shovel digging the ground. Therefore, according to the above five reasons, the optical fiber vibration event is divided into five events:

[0064] The vibration event caused by background noise refers to an event that the sensing optical fiber generates vibration due to noise caused by system noise of the optical fiber data acquisition system or changes in the surrounding environment. For example, light source frequency drift noise, detector noise, environmental vibration background noise, and slow creep noise all belong to background noise.

[0065] The vibration event caused by human walking refers to an event that the sensing optical fiber generates vibration due to vibration waves generated by people walking on the ground and transmitted to the sensing optical fiber through geology.

[0066] The vibration event caused by human jumping refers to an event that the sensing optical fiber generates vibration due to vibration waves generated by people jumping on the ground and transmitted to the sensing optical fiber through geology.

[0067] The vibration event caused by shovel hitting refers to an event that the sensing optical fiber generates vibration due to vibration waves generated by people using a shovel or the like to hit the ground on the ground, and the vibration waves are transmitted to the sensing optical fiber through geology.

[0068] The vibration event caused by shovel digging refers to an event that the sensing optical fiber generates vibration due to vibration waves generated by people using a shovel or the like to dig the ground on the ground, and the vibration waves are transmitted to the sensing optical fiber through geology.

[0069] In the optional embodiment, among the above five events, the vibration events caused by human jumping and shovel hitting are low-probability events. The data samples of these two types of vibration events are difficult to obtain, and the number of data samples is small. In this regard, the application classifies the vibration event data samples of the human jumping type and the shovel hitting type into the minor sample category data sample Minor_Class. In addition, the vibration events caused by human walking, shovel digging, and background noise are high-probability events. The data samples of these three types of vibration events are easy to obtain, and the number of data samples is large. In this regard, the application classifies the vibration event data samples of the background noise type, the shovel digging type, and the human walking type into the major sample category data sample Major_Class.

[0070] S300, according to the vibration event data sample of the background noise type, combining the data augmentation method based on time domain transfer, performing data augmentation on the TS time domain data in the minor sample category data sample Minor_Class, to obtain the augmented TS time domain data.

[0071] The purpose of this step is to use sufficient vibration event data samples of the background noise type to transfer diversified background noise to the minor sample category data sample Minor_Class, and enrich the background noise of the minor sample category data sample Minor_Class.

[0072] S400, according to the specific data sample, the augmented TS time domain data is augmented by a cycle of a generative adversarial network to obtain an augmented data sample, the augmented data sample is converted into an RGB format and added to the vibration event data sample to obtain a total training set Train Set.

[0073] It should be noted that the specific data sample refers to the vibration event data sample in the multi-sample category data sample that does not belong to the background noise class. For example, the multi-sample category data sample Major_Class includes vibration event data samples belonging to the background noise class, the human walking class and the shovel digging class. Then, the specific data sample is the vibration event data sample of the human walking class and the shovel digging class.

[0074] In this step, the multi-sample category data sample is augmented by the multi-sample category data sample. This can transfer the vibration diversity of the multi-sample category data sample Major_Class to at least the sample category data sample Minor_Class, enrich the vibration diversity of the few-sample category data sample Minor_Class, and increase the number of data samples of the few-sample category data sample Minor_Class.

[0075] S500, pre-training a deep convolutional network, inputting the total training set Train Set into the pre-trained deep convolutional network and training it to obtain an optical fiber classification model.

[0076] In this step, the step of pre-training the deep convolutional network is to pre-train on a large image classification dataset. In addition, before the current Rayleigh scattering light intensity signal is input into the optical fiber classification model, the format of the current light intensity signal needs to be converted into an RGB format before it can be input into the classification model. Optionally, the large image classification dataset is an ImageNet dataset, and the deep convolutional network is an AlexNet.

[0077] S600, detecting the current Rayleigh scattering light intensity signal of the sensing optical fiber that currently occurs a vibration event and inputting it into the optical fiber classification model to obtain a classification result of the vibration event.

[0078] Please refer to Figure 1 In an embodiment of the present application, the process of detecting the Rayleigh scattering light intensity signal of the historical sensing optical fiber when a vibration event occurs will be further described and explained as follows. The specific process of detecting the Rayleigh scattering light intensity signal of the historical optical fiber vibration event is as follows:

[0079] Firstly, the narrow linewidth light source 100 generates continuous light. Optionally, the wavelength of the continuous light is 1550 nm. The continuous light enters the first coupler 200 and is divided into probe light with an energy ratio of 90% and local oscillator light with an energy ratio of 10% through the first coupler 200. It should be noted that the energy ratio of 90% means that the energy of the probe light accounts for 90% of the energy of the continuous light, and the energy ratio of 10% means that the energy of the local oscillator light accounts for 10% of the energy of the continuous light. Then, the probe light enters the first frequency shift module 1000, and the local oscillator light enters the second frequency shift module 1100.

[0080] In the first frequency shift module 1000, the probe light is modulated into pulsed light with a width of 200 ns through the pulse modulation acousto-optic modulator 1002 and generates a frequency shift f1 of MHz level. That is, the pulse modulation acousto-optic modulator 1002 modulates the probe light into pulsed light with a frequency shift f1. Then, the pulsed light enters the optical amplifier 1001, is amplified by the optical amplifier 1001, enters the circulator 400, and is injected into the first end of the sensing optical fiber 700 through the first optical wavelength division multiplexer 600, and enters the sensing optical fiber 700.

[0081] It should be noted that the frequency shift amount of the frequency shift f1 is determined by the pulse modulation acousto-optic modulator 1002. In the embodiment, f1 = 200 MHz.

[0082] In the second frequency shift module 1100, the local oscillator light is modulated through the frequency shift acousto-optic modulator 1102 and generates another frequency shift f2 of MHz level. Then, the local oscillator light with the frequency shift is attenuated in energy through the adjustable attenuator 1101 and enters the second coupler 300. It should be noted that the frequency shift amount of the frequency shift f2 is determined by the frequency shift acousto-optic modulator 1102. In the embodiment, f2 = 180 MHz.

[0083] The first Raman pump laser 500 generates forward pumping light. In the embodiment, the wavelength of the forward pumping light is 1455 nm. The forward pumping light enters the first end of the sensing optical fiber 700 through the first optical wavelength division multiplexer 600 and enters the sensing optical fiber 700. The pulsed light in the sensing optical fiber 700 and the forward pumping light undergo energy transfer.

[0084] At the same time, the second Raman pump laser 800 generates backward pumping light. In the embodiment, the wavelength of the backward pumping light is 1455 nm. The backward pumping light enters the tail end of the sensing optical fiber 700 through the second optical wavelength division multiplexer 900 and enters the sensing optical fiber 700. The pulsed light in the sensing optical fiber 700 and the backward pumping light undergo energy transfer.

[0085] In the embodiment, the forward pumping light mainly performs distributed optical amplification on the pulse light in the front half of the sensing fiber 700 due to the energy attenuation of the forward pumping light with the propagation distance. The backward pumping light mainly performs distributed optical amplification on the pulse light in the back half of the sensing fiber 700 due to the energy attenuation of the backward pumping light with the propagation distance. In this way, the effective sensing distance can be improved.

[0086] When the pulse light is shifted in energy by the backward pumping light and the forward pumping light respectively, the pulse light becomes back Rayleigh scattering light. The back Rayleigh scattering light is multiplexed by the first optical wavelength division multiplexer 600 and then returns to the circulator 400. The circulator 400 emits the back Rayleigh scattering light to the second coupler 300.

[0087] At the second coupler 300, the back Rayleigh scattering light and the frequency-shifted local light interfere with each other to generate an interference optical signal. In the embodiment, the carrier beat frequency of the interference optical signal is |f1-f2|. Then, the interference optical signal enters the data processing module 1200. In the data processing module 1200, the interference optical signal is converted into an electrical signal by the balanced detector 1201, and the electrical signal is collected by the collection card 1202. Then, the electrical signal is sent to the host computer 1203 for digital quadrature demodulation, and finally the Rayleigh scattering light intensity signal is output.

[0088] In one embodiment of the present application, the process of converting the Rayleigh scattering light intensity signal into a vibration event data sample in step 200 will be further described and explained. The conversion of the Rayleigh scattering light intensity signal into a vibration event data sample can include but is not limited to the following steps.

[0089] S210, rearrange the obtained Rayleigh scattering light intensity signal into a matrix form to obtain a space-time matrix signal.

[0090] It should be noted that the i-th Rayleigh scattering light intensity signal is denoted as TR i . Then, the space-time matrix signal can be denoted as TR_Matrix=[TR1, TR2, TR3,..., TR i ,..., TR N ] T . Wherein: the detection light pulse repetition frequency is represented by N. In the embodiment, N=2kHz.

[0091] Alternatively, the plurality of Rayleigh scattering light intensity signals obtained within 1s are rearranged to form a space-time matrix signal. In the embodiment, the detection light pulse repetition frequency of the optical fiber data acquisition system is set to 2KHz. Therefore, 2000 Rayleigh scattering light intensity signals can be obtained within 1s, and the space-time matrix signal formed by the 2000 Rayleigh scattering light intensity signals is denoted as TR_Matrix=[TR1, TR2, TR3,..., TR i ,..., TR 2000 ]T At 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 scattering light intensity signal.

[0092] In S220, the mobile differential average method is used to locate the vibration event, and the position information of the located vibration event is used as the center to extract the data of the corresponding position in the space-time matrix signal TR_Matrix. The data of the corresponding position is recorded as TS time sequence.

[0093] In S230, the TS time sequence is subjected to band-pass filtering processing to filter out the direct current component and high-frequency system noise in the TS time sequence, and TS time domain data is obtained.

[0094] It should be noted that the first storage form of the vibration event data sample is TS time domain data.

[0095] This step realizes band-pass filtering through a band-pass filter. The band-pass filter is a filter that only allows specific frequencies to pass through while effectively suppressing the signals of the remaining frequencies.

[0096] In S240, the TS time domain data is subjected to time-frequency transformation processing and converted into an RGB image.

[0097] It should be noted that the second storage form of the vibration event data sample is an RGB image.

[0098] Optionally, the time-frequency transformation processing is performed by any one of short-time Fourier transform, Mel cepstrum transform, or Hilbert-Huang transform. The short-time Fourier transform (STFT) is a mathematical transform related to the Fourier transform, which is used to determine the frequency and phase of the sinusoidal wave in the local region of the time-varying signal. The idea of short-time Fourier transform is to select a time-frequency localized window function, assume that the signal is pseudo-stationary in a short time interval, move the window function so that the signal is a stationary signal in different finite time width, and thus calculate the power spectrum at each time.

[0099] The Hilbert-Huang transform (HHT) mainly includes two parts. The first part is empirical mode decomposition, and the second part is Hilbert spectrum analysis. First, the given signal is decomposed into a number of intrinsic mode functions by empirical mode decomposition method, and these intrinsic mode functions are components that meet certain conditions. Then, the Hilbert transform is performed on each intrinsic mode function to obtain the corresponding Hilbert spectrum, i.e., each intrinsic mode function is represented in the joint time-frequency domain. Finally, the Hilbert spectrum of all intrinsic mode functions is obtained by summing up the Hilbert spectrum of all intrinsic mode functions.

[0100] The mel-frequency cepstral transform is a time-frequency transform of a time series by a mel-frequency filter.

[0101] In an embodiment, the step of time-frequency transform processing the TS time domain data by the mel-frequency cepstral transform comprises:

[0102] S241, dividing the TS time domain data in one second into a plurality of short frames, and calculating the energy spectrum of each short frame by fast Fourier transform;

[0103] S242, constructing a plurality of mel filter banks, and for each mel filter bank, calculating the total energy of each group and taking the logarithm of the energy:

[0104]

[0105] Wherein s(m) is the logarithmic energy of the mth mel filter output of a frame of vibration signal. M is the number of filter banks. K is the length of each frame, F a is the fast Fourier transform of each frame of signal, and H is the frequency response of the mel filter.

[0106] Assuming that the center frequency of the mth mel filter is f(m), then its response H is:

[0107]

[0108] S243, the plurality of short frames form a time-frequency matrix S after passing through the mel filter bank, and the time-frequency matrix S is drawn into an RGB image.

[0109] It should be noted that the time-frequency matrix S satisfies the following formula:

[0110] S = [s i (1) T ,s2(2) T ,s3(3) T ,...,s p (M) T ];

[0111] Wherein p is the number of frames in one second. In this embodiment, p = 155, and the size of the generated time-frequency matrix S is 155*32. The size of the RGB image mapped by the time-frequency matrix S is 224*224.

[0112] In an embodiment of the present application, the process of data augmentation of step 300 will be described and explained below. Step 300 can include but is not limited to the following steps.

[0113] S310, extracting i TS time domain data from the few-sample class data samples to generate a time domain data set.

[0114] It should be noted that these i TS time-domain data constitute the time-domain dataset TS. Minor_Class Among them, the time-domain dataset TS Minor_Class The i-th TS time-domain data is passed through TS Minor_Class_i express.

[0115] S320 extracts TS time-domain data of j background noise classes from multi-sample class data samples to generate an augmented dataset.

[0116] It should be noted that the TS time-domain data of these j background noise classes constitute the augmented dataset TS. background The TS time-domain data of the j-th background noise class is obtained through TS. background_j To express.

[0117] S330 augments the time-domain dataset by expanding the dataset to obtain the augmented TS time-domain data.

[0118] It should be noted that the amplified TS time-domain data TS Minor_Class_i_j satisfy:

[0119] TS Minor_Class_i_j =TS Minor_Class_i +TS background_j ;

[0120] In this embodiment, i = 2, j = 1000, that is, 2 samples are amplified into 2000 samples.

[0121] This application expands the number of Minor_Class data samples by employing a time-domain transfer-based data augmentation method. Simultaneously, by utilizing the TS time-domain data of sufficient background noise vibration event data samples, the diversity of background noise is transferred to the Minor_Class data samples, enriching their background noise characteristics.

[0122] It should be noted that the background noise characteristics of the Minor_Class data samples amplified are achieved using TS time-domain data, not RGB images. Since RGB images of the background noise class lack obvious vibration patterns, it is difficult to learn their background noise characteristics using neural networks. If a neural network were used to learn these characteristics, the original vibration patterns would be blurred due to the lack of a central vibration image. Therefore, this application uses TS time-domain data to augment the background noise characteristics. Furthermore, the purpose of augmenting the background noise characteristics in this application is to increase the realism of the data samples in the Minor_Class data samples.

[0123] Reference Figure 3 As shown, Figure 3Fig. 1 is a structural schematic diagram of a data augmentation method based on a cycle generative adversarial network provided by an embodiment of the present application. After the expansion of the background noise characteristics is completed, the data augmentation method based on CycleGAN is used to complete the feature learning other than the background noise features. In one embodiment of the present application, the process of data augmentation in step 400 will be described and explained below. Step 400 can include but is not limited to the following steps.

[0124] It should be noted that in this embodiment, the few-shot class data samples are all human jumping classes. The specific data samples are defined as vibration event data samples belonging to the multi-shot class other than the background noise class, and the specific data samples are all human walking classes.

[0125] S410, the augmented TS time domain data is subjected to time-frequency transformation processing to obtain an extended RGB image corresponding to the augmented TS time domain data, and a few-shot dataset is generated.

[0126] Optionally, the time-frequency transformation processing is performed by any one of Hilbert-Huang transformation, mel-frequency cepstrum transformation, and short-time Fourier transform.

[0127] S420, the RGB images in the specific data samples are screened out to generate a multi-shot dataset.

[0128] In this specific embodiment, the few-shot dataset includes a plurality of extended RGB images of human jumping classes. The multi-shot dataset includes a plurality of extended RGB images of human walking classes. Optionally, the vibration event data samples of human jumping classes and the vibration event data samples of human jumping classes are similar, i.e., the similarity between the two is higher than a similarity threshold. The similarity threshold is not specifically limited in the present application, and the similarity threshold is only a data for measuring whether two data samples are similar.

[0129] S430, a cycle generative adversarial network is constructed, and the cycle generative adversarial network is trained using the few-shot dataset and the multi-shot dataset to obtain a data augmentation network.

[0130] In the present application, the cycle generative adversarial network includes a discriminator D y , a discriminator D x , a generator G, and a generator F. The structures of the discriminator D y , the discriminator D x , the generator G, and the generator F are consistent with the structure of a standard cycle generative adversarial network.

[0131] In this step, the images of the few-shot dataset are input into the generator G, and the generated images of the multi-shot dataset are output. If the generated images of the few-shot dataset are used as input, the restored images of the few-shot dataset are output. The images of the multi-shot dataset are input into the generator F as input, and the generated images of the few-shot dataset are generated. If the generated images of the multi-shot dataset are used as input, the restored images of the few-shot dataset are output. In addition, the quality of the generated images of the multi-shot dataset is judged by the discriminator D y The quality of the generated images of the multi-shot dataset is judged by the discriminator D x The quality of the generated images of the few-shot dataset is judged.

[0132] In this embodiment, the images of the human jumping category are input into the generator G, and the images of the human walking category can be generated by the generator G. The images of the human walking category are input into the generator F, and the images of the human jumping category can be generated by the generator F. The quality of the generated images of the human walking category is judged by the discriminator D y The quality of the generated images of the human walking category is judged by the discriminator Dx, and the quality of the generated images of the human jumping category is judged by the discriminator Dx.

[0133] S440, input the RGB images in the multi-shot category data sample into the data augmentation network for data augmentation to obtain the augmented data sample.

[0134] In this step, all the RGB images in the multi-shot category data sample are input into the data augmentation network to complete the augmentation of the few-shot category data sample Minor_Class.

[0135] In this embodiment, the data sample of the human jumping category is augmented by the data sample of the human walking category and the data augmentation network. In this embodiment, 3000 vibration event data samples of the human jumping category are generated by the data augmentation network.

[0136] In an optional embodiment, the few-shot category data sample can be the shovel hitting ground category, and the specific data sample can be the shovel digging ground category. Then, the process of the data augmentation method based on CycleGAN is as follows: a few-shot dataset containing augmented RGB images of the shovel hitting ground category is constructed, and a multi-shot dataset containing RGB images of the shovel digging ground category is constructed. The RGB images of the shovel digging ground category are combined with the augmented RGB images of the shovel hitting ground category to train the CycleGAN, and then a trained data augmentation network is obtained. Then, all the RGB images in the multi-shot category data sample containing the shovel digging ground category and the background noise category are input into the data augmentation network to obtain the augmented augmented data sample. In this optional embodiment, 3000 vibration event data samples of the shovel hitting ground category can be generated by the data augmentation network.

[0137] Reference Figures 3 to 5As shown, one embodiment of the present application, the following further illustrates and describes step 430. The few-shot dataset X1 is taken as one of the datasets in the CycleGAN model, and the multi-shot dataset Y1 is taken as the other dataset in the CycleGAN model. 1i is the i-th extended RGB image in the few-shot dataset X1. 1i is the i-th RGB image in the multi-shot dataset Y1. Step 430 specifically includes:

[0138] S431, an extended RGB image x 1i is randomly extracted from the few-shot dataset X1. 1i The generator G is trained using the extended RGB image x 1i , and outputs a generated image y' 1i .

[0139] It is to be noted that the generated image y' 1i has a content similar to the multi-shot category and has a style of the few-shot category. The generated image y' 1i satisfies: y' 1i =G(x 1i ).

[0140] In one embodiment, all the extended RGB images in the few-shot dataset are of the human jumping category, and all the RGB images in the multi-shot dataset are of the human walking category. Then, the above steps are to extract an extended RGB image x 1i belonging to the human jumping category from the few-shot dataset X1, and train the generator G using x 1i , and output a generated image y' 1i of the human walking category. The generated image y' 1i is an image having a content similar to the human walking category and having a style of the human jumping category.

[0141] In another embodiment, all the extended RGB images in the few-shot dataset are of the shovel hitting ground category, and all the RGB images in the multi-shot dataset are of the shovel digging ground category. Then, the above steps are to extract an extended RGB image x 1i belonging to the shovel hitting ground category from the few-shot dataset X1, and train the generator G using x 1i , and output a generated image y' 1i of the shovel digging ground category. The generated image y' 1i is an image having a content similar to the shovel digging ground category and having a style of the shovel hitting ground category.

[0142] S432, an RGB image y 1i is randomly extracted from the multi-shot dataset Y1. 1i The generator F is trained using the RGB image y 1i , and outputs a generated image x' 1i .

[0143] It should be noted that the generated image x′ 1i For images with similar few-sample content and a multi-sample category style, generate image x′. 1i Satisfy: x′ 1i =F(y 1i ).

[0144] In one embodiment, the extended RGB images in the few-sample dataset are all of the human jumping class, while the RGB images in the many-sample dataset are all of the human walking class. Therefore, the above step involves extracting the RGB images y of the human walking class. 1i And through y 1i Train the generator F to output the generated image x′ of the human jumping class. 1i Generate image x′ 1i It features content similar to a human jumping game and a style reminiscent of a human walking game.

[0145] In another embodiment, the extended RGB images in the few-sample dataset all represent the "shovel-photographed land" category, while the RGB images in the many-sample dataset all represent the "shovel-digged land" category. Therefore, the above step involves extracting the RGB image y of the "shovel-digged land" category. 1i And through y 1i Train the generator F to output the generated image x′ of the shovel-shaped landform. 1i Generate image x′ 1i It features content similar to the "shovel and shovel" genre and has a style reminiscent of "shovel and shovel" genres.

[0146] S433 will generate image y′ 1i As input to generator F, the extended RGB image x is obtained. 1i The restored image

[0147] It should be noted that the restored image satisfy:

[0148] S434 will generate image x′ 1i As input to generator G, the RGB image y is obtained. 1i The restored image

[0149] It should be noted that the restored image satisfy:

[0150] S435, Training Discriminator D y and discriminator D x Through the trained discriminator D x Determine the generated image x′ 1ithe quality of the generated image y' y the quality of the generated image y' 1i .

[0151] In this step, judging the quality of the generated image is equivalent to judging whether the generated image is good or bad.

[0152] Optionally, the discriminator D y , the discriminator D x , the generator G and the generator F are trained synchronously.

[0153] S436, calculate the loss function of CycleGAN.

[0154] In this step, in the training process of the discriminator D y , the discriminator D x , the generator G and the generator F, the loss function Loss cycleGAN is introduced to ensure that the generated image and the original image are as similar as possible.

[0155] In an embodiment, the extended RGB images in the small sample data set are all human jumping classes, and the RGB images in the large sample data set are all human walking classes. Through the loss function Loss cycleGAN , it is ensured that the generated image y' 1i of the human walking class and the RGB image y 1i of the human walking class are as similar as possible, and the generated image x' 1i of the human jumping class and the extended RGB image x 1i of the human jumping class are as similar as possible.

[0156] In another embodiment, the extended RGB images in the small sample data set are all shovel hitting classes, and the RGB images in the large sample data set are all shovel digging classes. Through the loss function Loss cycleGAN , it is ensured that the generated image of the shovel digging class and the original image of the shovel digging class are as similar as possible, and the generated image of the shovel hitting class and the extended RGB image of the shovel hitting class are as similar as possible.

[0157] Specifically, the loss function satisfies the following formula:

[0158] Loss cycleGAN = Loss GAN + Loss cycle .

[0159] In the above formula, the total loss function is represented by Loss cycleGAN , the adversarial loss between the generator and the discriminator is represented by Loss GAN , and the cycle consistency loss in CycleGAN is represented by Loss cycleTo express.

[0160] Specifically, combating loss GAN Its purpose is to ensure that the generator and discriminator evolve with each other, thereby ensuring that the generator can produce more realistic images. The adversarial loss is:

[0161]

[0162] In the above formula: generator G and discriminator D y The counter-loss through L GAN (G, D) y The generator F and discriminator D are represented by X1, Y1). x The counter-loss through L GAN (F, D) x The subscripts X1 to Pdata(x, Y1) are used to represent the expected value. E represents the expected value. 1i The image x represents an extended RGB image randomly sampled from a small sample dataset X1. 1i Subscript Y1 ~ Pdata(y 1i The image y represents an RGB image randomly sampled from a multi-sample dataset Y1. 1i Discriminator D y The output is through D y (·) represents the discriminator D. x The output is through D x (·) is used to represent this. Where: D y (G(x 1i )) is the discriminator D y The generated image y′ is generated by the discriminant generator G. 1i D x (F(y 1i )) is the discriminator D x The generated image x′ is generated by the discriminant generator F 1i .

[0163] In one embodiment, the extended RGB images in the few-sample dataset are all of the human jumping class, while the RGB images in the many-sample dataset are all of the human walking class. X1 ～ Pdata(x 1i This represents an expanded RGB image x of a person jumping class randomly selected from a few-sample dataset X1. 1i Y1~Pdata(y 1i Then, it represents the RGB image y of a walking person randomly selected from the multi-sample dataset Y1. 1i .

[0164] In another embodiment, the extended RGB images in the few-sample dataset all represent the "shovel-photographed" land type, while the RGB images in the many-sample dataset all represent the "shovel-digged" land type. X1~Pdata(x 1i Then, it represents the extended RGB image x of a randomly selected land parcel from the few-sample dataset X1. 1i X1~Pdata(x 1i This represents the RGB image y of a randomly selected excavated land type from the multi-sample dataset Y1. 1i .

[0165] Specifically, the cycle consistency loss cycle The purpose is to ensure that the generated image and the original image differ only in style, but have the same content. The cycle consistency loss is:

[0166]

[0167] In the above formula, ||·|| represents regularization. (This is achieved through...) It can be concluded that F(G(x) 1i )) represents an extended RGB image x 1i The restored image And, through It can be known that G(F(y) 1i )) represents the RGB image y 1i The restored image

[0168] In one embodiment, the extended RGB images in the few-sample dataset are all of the human jumping class, while the RGB images in the many-sample dataset are all of the human walking class. F(G(x) 1i )) An extended RGB image x representing the human jumping class 1i The restored image G(F(y 1i )) represents the RGB image of a walking human. 1i The restored image

[0169] In another embodiment, the extended RGB images in the few-sample dataset all represent the "shovel-photographed land" category, while the RGB images in the many-sample dataset all represent the "shovel-digged land" category. Therefore, G(F(y) 1i )) represents the RGB image of the excavated land type. 1i The restored image F(G(x 1i )) represents the extended RGB image x of the land cover. 1i The restored image

[0170] S437, updating the parameters of the generator of the CycleGAN according to the loss function of the CycleGAN, repeating the above steps until the training threshold is reached, and outputting the data augmentation network.

[0171] It should be noted that the parameters of the generator G and the generator F are updated using the Adam optimization algorithm.

[0172] Optionally, the training threshold is set to 200, which refers to the number of training cycles. The learning rate remains unchanged in the first 100 training cycles, and decreases with the increase of the number of iterations in the last 100 training cycles, and finally decreases to zero.

[0173] In an embodiment, the few-shot class data samples are all human jumping classes, and the multi-shot class data samples are all human walking classes. Then, after the training of the data augmentation network is completed, the vibration event data samples of the human walking class are input into the data augmentation network to achieve data augmentation, and a plurality of generated images similar to the human jumping class and having the style of the human walking class are obtained, thereby completing the augmentation of the vibration event data samples of the human jumping class.

[0174] In another embodiment, the few-shot class data samples are all shovel hitting ground classes, and the multi-shot class data samples are all shovel digging ground classes. Then, after the training of the data augmentation network is completed, the vibration event data samples of the shovel digging ground class are input into the data augmentation network, and the data augmentation network generates images having the content of the shovel hitting ground class and having the style of the shovel digging ground class, thereby achieving the augmentation processing of the vibration event data samples of the shovel hitting ground class.

[0175] The present application uses a data augmentation method based on CycleGAN to transfer the diversity of vibration to the few-shot class data samples Minor_Class using CycleGAN, thereby enriching the vibration diversity of the few-shot class data samples Minor_Class.

[0176] As an optional embodiment, since the classification category of the few-shot class data samples and the data category of the multi-shot class data samples can be divided into multiple categories. For example, the few-shot class data samples can be divided into human jumping classes and shovel hitting ground classes, and the multi-shot class data samples can be divided into human walking classes and shovel digging ground classes. Then, in the data augmentation process based on CycleGAN, a CycleGAN can be trained according to the data of the human jumping class and the data of the human walking class to obtain a data augmentation network. At the same time, another CycleGAN is trained according to the data of the shovel digging ground class and the shovel hitting ground class to obtain another data augmentation network. Then, the data of the human walking class and the shovel digging ground class are input into the two trained data augmentation networks, respectively, thereby completing the augmentation of the data of the human jumping class and the shovel hitting ground class.

[0177] Understandably, when both few-shot and many-shot data samples are divided into multiple categories, it's crucial to ensure that the categories in the many-shot and few-shot data samples are paired. For example, "walking person" and "jumping person" are paired, as are "digging with a shovel" and "patting with a shovel." When both few-shot and many-shot data samples are divided into multiple categories, CycleGANs need to be trained separately for each paired category to obtain different data augmentation networks.

[0178] In one embodiment of this application, step 500, which involves training a deep convolutional network to obtain a fiber optic classification model, specifically includes:

[0179] S510, the total training set is input into the deep convolutional network for forward propagation to obtain the classification output and its classification probability.

[0180] S520 calculates the loss function for deep convolutional networks.

[0181] It should be noted that the loss function of deep convolutional networks satisfies:

[0182]

[0183] Here, the total number of samples is represented by w, and the total number of event categories is represented by n. The classification probability refers to the probability that a vibration event data sample belongs to category c, expressed as p. j·c To represent. The indicator variable is represented by z. j·c To represent. When the time category and category c of the vibration event data samples are the same, z j·c It equals one; conversely, z equals 1. j·C It equals zero.

[0184] S530, based on the loss function of the deep convolutional network, uses the backpropagation gradient descent algorithm to update the weights of the deep convolutional network, thus obtaining the fiber optic early warning model.

[0185] Based on the above embodiments, referring to Figure 6 As shown, Figure 6 The diagram shows a flowchart of the fiber optic vibration event classification method provided in this application. The following example illustrates the dual data amplification process and vibration event classification process of this application. In this example, the first category is human jumping, the second category is shovel-pounding, the third category is human walking, and the fourth category is shovel-digging. Multi-sample category data is classified into shovel-digging, human walking, and background noise categories. Few-sample category data is classified into human jumping and shovel-pounding.

[0186] Firstly, a fiber data acquisition system is built to obtain the data samples of the sensing fiber that has occurred a vibration event at a historical time, and the vibration event data samples are represented by Rayleigh scattering light intensity signals. According to five reasons causing the vibration event, i.e., event categories, the vibration event data samples are classified into five categories, i.e., human walking category, shovel slapping ground category, human jumping category, shovel digging ground category and background noise category. Since the data samples of the background noise category, human walking category and shovel digging ground category are relatively large in quantity and easy to obtain, and the data samples of the human jumping category and shovel slapping ground category are relatively small in quantity and not easy to obtain, the vibration event data samples of the human jumping category and shovel slapping ground category are classified as minor class data samples Minor_Class, and the vibration event data samples of the background noise category, human walking category and shovel digging ground category are classified as major class data samples Major_Class. Two CycleGANs are trained respectively through the minor class data samples Minor_Class and the major class data samples Major_Class, and are respectively used for expanding the data samples of the human jumping category and the data samples of the shovel slapping ground category.

[0187] Then data augmentation is performed. Firstly, the background noise characteristics in the minor class data samples Minor_Class need to be increased, i.e., the background noise characteristics of the vibration event data samples of the human jumping category and the shovel slapping ground category need to be increased. The data augmentation method based on time domain transfer is adopted to expand the number of data samples of the minor class data samples Minor_Class.

[0188] After the expansion of the background noise characteristics, other characteristics in the minor class data samples Minor_Class need to be increased. CycleGAN is suitable for two-dimensional images, so the RGB images of the major class data samples Major_Class are input into the two trained CycleGANs to respectively expand the data samples of the human jumping category and the data samples of the shovel slapping ground category, and the expanded data samples are obtained. Specifically, one CycleGAN is trained according to the data samples of the human jumping category and the human walking category, and another CycleGAN is trained according to the data samples of the shovel slapping ground category and the shovel digging ground category, and two data augmentation networks are obtained. The data samples of the shovel digging ground category are input into the corresponding network for expansion, and the data samples of the human walking category are input into the corresponding network for expansion, and then the expanded data samples are obtained.

[0189] After that, the amplified data samples and the original few-sample category data samples Minor_Class and the original many-sample category data samples Major_Class constitute a total training set Train_Set. The total training set Train_Set is input into the deep convolutional neural network pre-trained by ImageSet, and the total training set Train_Set is used for training the deep convolutional neural network, so as to obtain the optical fiber classification model. Since the present application is applicable to few-sample training, the deep convolutional neural network and the CycleGAN do not need to be tested by a test set. After that, the newly detected Rayleigh scattering light intensity signal is input into the optical fiber classification model, and the classification result can be obtained.

[0190] The present application reduces the difficulty of obtaining the vibration event data samples of the optical fiber through the optical fiber data acquisition system, and increases the number of samples used for training the classification model through the double data amplification method for amplifying and filling the small number of data samples. Finally, the classification model is trained using the amplified data samples and the original data samples, and then the trained classification model is used to classify the vibration events of the optical fiber. The present application reduces the difficulty of obtaining the vibration event data samples of the optical fiber, increases the number of data samples through the double data amplification method, and the degree of authenticity of the data obtained after amplification is high. Only a small number of data samples are needed to complete the training of the vibration event classification model of the optical fiber, which ensures the accuracy of the model while reducing the training difficulty. The present application is applicable to small sample learning and some event recognition scenarios where training samples are difficult to collect.

[0191] In addition, the present application also provides a storage medium having processor-executable instructions stored therein, wherein the processor-executable instructions, when executed by a processor, are used to execute the optical fiber vibration event classification method based on double data amplification.

[0192] The terms "first", "second", "third", "fourth" and the like used in the description of the present application and the above drawings, if any, are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0193] It should be understood that, in the application, "at least one" refers to one or more, and "multiple" refers to two or more. "And / or" is used to describe the association relationship of the associated objects, which means that there can be three relationships, for example, "A and / or B" can represent three cases of only A, only B, and A and B existing at the same time, wherein A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after it. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can represent a, b, c, "a and b", "a and c", "b and c", or "a and b and c", wherein a, b, and c can be single or multiple.

[0194] In several embodiments provided in the application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division, and actual implementation can have another division manner. For example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0195] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0196] In addition, each functional unit in each embodiment of the application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0197] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0198] For the step numbers in the above method embodiments, they are only set for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

Claims

1. A method of classifying fiber optic vibration events based on dual data expansion, characterized in that, The method comprises the following steps: Constructing an optical fiber data acquisition system, and detecting a plurality of Rayleigh scattering light intensity signals of the historical sensing optical fiber when a vibration event occurs through the optical fiber data acquisition system; A plurality of Rayleigh scattering light intensity signals are made into vibration event data samples, the vibration event data samples comprise RGB images and TS time domain data, and the vibration event data samples are divided into few-sample category data samples and many-sample category data samples; The collection difficulty of the few-sample category data samples is higher than that of the many-sample category data samples, and the many-sample category data samples comprise vibration event data samples belonging to a background noise category; According to the vibration event data samples of the background noise category, and in combination with a time domain transfer-based data augmentation method, TS time domain data in the few-sample category data samples are subjected to data augmentation to obtain augmented TS time domain data; According to specific data samples, the augmented TS time domain data are subjected to augmentation through a cycle generative adversarial network to obtain augmented data samples, the augmented data samples are converted into an RGB format and added to the vibration event data samples to obtain a total training set; The specific data samples are vibration event data samples not belonging to the background noise category in the many-sample category data samples; A deep convolutional network is pre-trained, the total training set is input into the pre-trained deep convolutional network and is trained to obtain an optical fiber classification model; A current Rayleigh scattering light intensity signal of a sensing optical fiber currently undergoing a vibration event is detected and is input into the optical fiber classification model to obtain a classification result of the vibration event; The vibration event data samples of the background noise category are subjected to data augmentation through the cycle generative adversarial network to obtain the augmented TS time domain data, and the method comprises the following steps: extract i TS time domain data from the few-sample class data sample to generate a time domain data set; wherein the i-th TS time domain data in the time domain data set is ; TS time domain data of jth background noise class in the augmented data set is ; The time domain data set is subjected to data augmentation through the augmented data set to obtain the augmented TS time domain data; The augmented TS time domain data satisfy the following conditions: TS time domain data after expansion; The specific data samples are vibration event data samples not belonging to the background noise category in the many-sample category data samples; The RGB images in the specific data samples are screened out to generate a many-sample data set; The RGB images in the many-sample category data samples are input into the data augmentation network to obtain augmented data samples. A cycle adversarial generative network is constructed, and the cycle adversarial generative network comprises a generator G, a generator F, a discriminator and a discriminator The cycle adversarial generative network is trained using the small sample data set and the large sample data set to obtain a data augmentation network. The vibration event data samples are made from a plurality of Rayleigh scattering light intensity signals, the vibration event data samples comprise RGB images and TS time domain data, and the method comprises the following steps:

2. The dual data expansion based fiber vibration event classification method according to claim 1, wherein, A mobile difference average method is used to locate the vibration event, and the position information of the located vibration event is used as a center to extract data of a corresponding position in the space-time matrix signal, and the data of the corresponding position is denoted as a TS time sequence; The obtained Rayleigh scattering light intensity signal is rearranged into a matrix form to obtain a space-time matrix signal, denoted as , represents the i-th Rayleigh scattering light intensity signal, and N is the detection light pulse repetition frequency. ​ Band-pass filtering the TS time sequence to filter out DC components and high-frequency system noise in the TS time sequence, to obtain TS time domain data; wherein the TS time domain data is a first storage form of the vibration event data sample; Performing time-frequency transform processing on the TS time domain data and converting it into an RGB image, taking the RGB image as a second storage form of the vibration event data sample; The time-frequency transform processing includes any one of STFT, HHT or mel-frequency cepstrum transform.

3. The dual data expansion based fiber vibration event classification method of claim 1, wherein, The training of the cyclic generative adversarial network using the few-sample data set and the multi-sample data set includes: randomly sampling an extended RGB image from the small sample dataset training a generator G using the extended RGB image outputting a generated image ; wherein ; randomly sampling RGB images from the plurality of sample data sets training a generator F using the RGB images outputting generated images ; wherein ; said generated image as input to the generator F, obtaining the extended RGB image of the restored image ; wherein ; generating the image as input to the generator G, the RGB image of the restored image ; wherein ; training the discriminator and the discriminator , judging the quality of the generated image by the trained discriminator judging the quality of the generated image by the trained discriminator judging the quality of the generated image by the trained discriminator judging the quality of the generated image by the trained discriminator ​ Calculating the loss function of the cyclic generative adversarial network: wherein, is a total loss function, is a generator and discriminator adversarial loss, is a cycle consistency loss in the cycle adversarial generative network; The adversarial loss satisfies: wherein, is the adversarial loss of the generator G and the discriminator is the adversarial loss of the generator F and the discriminator ; E is the mathematical expectation; subscript is the extended RGB image randomly extracted from the few-shot dataset , subscript is the RGB image randomly extracted from the multi-shot dataset ; is the output of the discriminator , wherein: represents that the discriminator discriminates the generated image generated by the generator G ; is the output of the discriminator , wherein: is the output of the discriminator discriminates the generated image generated by the generator F ;​​​ The cyclic consistency loss satisfies: wherein, denotes a regular calculation; is the restored image of the extended RGB image ; ; is the restored image of the RGB image ; ; According to the loss function of the cyclic generative adversarial network, the parameters of the generator of the cyclic generative adversarial network are updated, and the above steps are repeated until the training threshold is reached, and the data augmentation network is output.

4. The dual data expansion based fiber optic vibration event classification method according to claim 3, wherein, The parameters of the generator G and the generator F are updated by the Adam optimization algorithm.

5. The dual data amplification based fiber vibration event classification method according to claim 1, wherein, The optical fiber data acquisition system comprises a narrow linewidth light source, a first coupler, a circulator, a sensing optical fiber, a first Raman pump laser, a second Raman pump laser, a second coupler, a pulse modulation acousto-optic modulator, an optical amplifier, a frequency shift acousto-optic modulator, an adjustable attenuator, a first optical wavelength division multiplexer, a second optical wavelength division multiplexer, a balanced detector, a data acquisition card and an upper computer. The first Raman pump laser is connected with the first end of the sensing optical fiber through the first optical wavelength division multiplexer, and the second Raman pump laser is connected with the tail end of the sensing optical fiber through the second optical wavelength division multiplexer. The narrow linewidth light source is connected with the pulse modulation acousto-optic modulator and the frequency shift acousto-optic modulator through the first coupler, the pulse modulation acousto-optic modulator is connected with the circulator through the optical amplifier, and the circulator is connected with the first optical wavelength division multiplexer and the second coupler respectively; the frequency shift acousto-optic modulator is connected with the second coupler through the adjustable attenuator. The second coupler is connected with the upper computer through the balanced detector and the data acquisition card.

6. The dual data expansion based fiber optic vibration event classification method according to claim 5, wherein, The Rayleigh scattering light intensity signal when the historical sensing optical fiber occurs a vibration event is detected by the optical fiber data acquisition system, comprising: The narrow linewidth light source generates continuous light, the continuous light is divided into probe light with an energy ratio of 90% and local oscillator light with an energy ratio of 10% by the first coupler; the pulse modulation acousto-optic modulator modulates the probe light into pulsed light with a frequency shift of ; the pulsed light enters the ring after being amplified by the optical amplifier, enters the sensing optical fiber through the first optical wavelength division multiplexer and the ring. The first Raman pump laser generates forward pumping light, the forward pumping light enters the sensing optical fiber from the first end of the sensing optical fiber through the first optical wavelength division multiplexer, and energy migration occurs with the pulse light; at the same time, the second Raman pump laser generates back pumping light, the back pumping light enters the sensing optical fiber from the tail end of the sensing optical fiber through the second optical wavelength division multiplexer, and energy migration occurs with the pulse light; the pulse light is converted into back Rayleigh scattering light after energy migration, and the back Rayleigh scattering light reaches the second coupler through the first optical wavelength division multiplexer and the circulator; The local light is frequency-shifted after passing through the frequency-shifted acousto-optic modulator , and then enters the second coupler through the adjustable attenuator; the frequency-shifted local light and the back Rayleigh scattering light interfere at the second coupler to form an interference light signal with a carrier beat frequency of , the interference light signal is converted into an electrical signal by the balanced detector, the electrical signal is collected by the acquisition card and sent to the host computer, and digital quadrature demodulation of the electrical signal is completed at the host computer to generate a Rayleigh scattering light intensity signal.

7. The dual data expansion based fiber optic vibration event classification method according to claim 1, wherein, The total training set is input into the pre-trained deep convolutional network and trained to obtain an optical fiber classification model, comprising: inputting the total training set into the deep convolutional network, performing forward propagation to obtain a classification output and a classification probability thereof; calculating a loss function of the deep convolutional network; wherein the loss function of the deep convolutional network is: where w is the total number of samples, n is the total number of event classes, is the probability that the vibration event data sample belongs to class c; is an indicator variable that is 1 when the time class of the vibration event data sample is the same as class c, is 1, otherwise it is 0; updating the weight of the deep convolutional network according to the loss function of the deep convolutional network by using a back propagation gradient descent algorithm to obtain an optical fiber early warning model.

8. A storage medium having stored therein instructions executable by a processor, the instructions causing the processor to perform the method of claim 1. The instructions executable by the processor, when executed by the processor, are used to perform the optical fiber vibration event classification method based on double data amplification according to any one of claims 1-7.

Images