A φ-OTDR pattern recognition system and method based on Wavenet

Abstract

The present invention discloses a φ-OTDR pattern recognition system and method based on Wavenet, comprising: a laser, an optical attenuator, an acousto-optic modulator, an erbium-doped fiber amplifier and a filter connected in sequence, the output end of the filter is connected to a circulator, the output ends of the circulator and the Raman amplifier are respectively connected to a wavelength division multiplexer, and the output end of the wavelength division multiplexer is connected to a sensing optical fiber through a PZT. Beneficial effects of the present invention: In the φ-OTDR system pattern recognition, the present invention directly extracts and recognizes features of the collected original signal, avoiding the process of manually extracting features required by the traditional method, and can more efficiently and fully obtain vibration information.

Description

Technical Field

[0001] The invention relates to the technical field of information processing and phase-sensitive optical time-domain reflection optical fiber sensing, and in particular to a φ-OTDR pattern recognition system and method based on Wavenet (acoustic wave network). Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Based on the Phase-Sensitive Optical Time-Domain Reflecrometric System, ), using optical fiber as a sensor, has the unique advantages of intrinsic safety, corrosion resistance, and low power consumption. It can monitor and locate vibration events along the optical fiber in real time online, and has a wide range of applications in perimeter security, pipeline safety monitoring, building structural health and other fields.

[0004] along with With the continuous deepening of the actual application of the system, it is difficult to meet the actual needs by only sensing and locating the vibration signal. Therefore, it is increasingly important to accurately and effectively identify various vibration events. At present, the method commonly used by researchers is to manually extract the feature information of the vibration signal in the time domain, frequency domain, time-frequency domain, spatial domain and other dimensions, and then input the feature information into the machine learning or deep learning structure for pattern recognition. However, whether the manually extracted features can fully represent the original vibration signal directly affects the accuracy of feature classification, and the manual feature extraction process itself will increase the computing resources of the entire pattern recognition method, resulting in poor real-time performance of the method.

[0005] The prior art discloses the use of a one-dimensional convolutional neural network to directly extract and identify features of a one-dimensional original signal, eliminating the need for manual feature extraction. However, the input is only amplitude information in the one-dimensional time domain, and the one-dimensional convolutional neural network lacks the ability to accurately express and analyze vibration timing signals.

[0006] The prior art discloses the use of a long short-term memory network model to train the time domain data of equipment operation vibration. However, due to the existence of cyclic connections in the long short-term memory network, the training speed of the model is limited to a certain extent. Summary of the invention

[0007] In order to solve the problems in the prior art of needing to manually extract features or poor real-time pattern recognition due to the structure of the neural network model itself, as well as the problem of low recognition accuracy due to the inability of the model to fully analyze the timing of its input data, the present invention proposes a φ-OTDR pattern recognition system and method based on Wavenet, which can accurately and efficiently identify various types of external vibration signals.

[0008] In some embodiments, the following technical solutions are adopted:

[0009] A φ-OTDR pattern recognition system based on Wavenet comprises: a laser, an optical attenuator, an acousto-optic modulator, an erbium-doped fiber amplifier and a filter which are connected in sequence, wherein the output end of the filter is connected to a circulator, the output ends of the circulator and the Raman amplifier are respectively connected to a wavelength division multiplexer, and the output end of the wavelength division multiplexer is connected to a sensing optical fiber via a PZT.

[0010] In other embodiments, the following technical solutions are adopted:

[0011] A φ-OTDR pattern recognition method based on Wavenet, comprising:

[0012] Acquire multiple groups of external vibration signals to be predicted through sensing optical fibers;

[0013] The collected vibration signal is processed into time-space data through threshold screening, the data is input into the trained Wavenet neural network model, and the vibration signal type is output;

[0014] The threshold is achieved by using a set sweep frequency signal to drive the PZT to stretch the sensing optical fiber to generate a corresponding strain, and the amplitude corresponding to the strain is used as the threshold for screening effective vibration signals.

[0015] In other embodiments, the following technical solutions are adopted:

[0016] A terminal device includes a server, wherein the server includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned Wavenet-based φ-OTDR pattern recognition method when executing the program.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] (1) The present invention uses an analog output card and PZT in the φ-OTDR system to simulate external vibrations, which facilitates the setting of thresholds so as to effectively filter out noise and invalid vibration signals.

[0019] (2) The use of optical attenuators, Raman amplifiers, and wavelength division multiplexers in the φ-OTDR system facilitates the setting of the optical power, which is beneficial to increasing the sensing distance of the φ-OTDR system and improving the signal-to-noise ratio of the sensing system.

[0020] (3) In the φ-OTDR system pattern recognition, the features of the collected original signal are directly extracted and recognized, which avoids the manual feature extraction process required by traditional methods and can more efficiently and fully obtain vibration information.

[0021] (4) The collected vibration signals form a time-space two-dimensional array. The causal hole convolution structure of Wavenet fully considers the long-span temporal sequence of the input two-dimensional array and can improve the recognition accuracy of the vibration signal type.

[0022] (5) The residual network and skip connection in Wavenet can make the model converge faster in a deeper neural network, which can effectively improve the training speed of the model and make the φ-OTDR pattern recognition highly real-time.

[0023] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the structure of a φ-OTDR pattern recognition system based on Wavenet in an embodiment of the present invention;

[0025] Figure 2 Schematic diagram of the φ-OTDR pattern recognition process in an embodiment of the present invention;

[0026] Figure 3 is a flowchart of processing collected raw data in an embodiment of the present invention;

[0027] Figure 4 is a stacked structure of Wavenet causal convolutional layers in an embodiment of the present invention;

[0028] Figure 5 It is the Wavenet multi-layer hole convolution structure in the embodiment of the present invention;

[0029] Figure 6 It is the Wavenet residual network structure in the embodiment of the present invention;

[0030] Figure 7 It is a flow chart of the Wavenet pattern recognition method in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0032] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0033] Embodiment 1

[0034] In one or more embodiments, a Wavenet-based φ-OTDR pattern recognition system is disclosed, referring to Figure 1 , including: the output end of the ultra-narrow linewidth laser is connected to the input end of the optical attenuator, the output end of the optical attenuator is connected to the optical input end of the acousto-optic modulator, the analog output card has two signal output ports, one output port is connected to the input end of the acousto-optic modulator driver; the other output port is connected to the PZT (piezoelectric ceramic) driving signal input end to drive the PZT. The output end of the acousto-optic modulator driver is connected to the acousto-optic modulator driving signal input end, the output end of the acousto-optic modulator is connected to the input end of the erbium-doped fiber amplifier, the output end of the erbium-doped fiber amplifier is connected to the input end of the filter, the output end of the filter is connected to an input end of the circulator, one output end of the circulator is connected to an input end of the wavelength division multiplexer, and the output end of the Raman amplifier is connected to another input end of the wavelength division multiplexer. The output end of the wavelength division multiplexer is connected to the optical input end of the PZT, and the output end of the PZT is connected to the sensing optical fiber.

[0035] The other output end of the circulator is connected to the input end of the photoelectric detector, the output end of the photoelectric detector is connected to the input end of the data acquisition card, and the output end of the data acquisition card is connected to the computer.

[0036] Specifically, when collecting data, the sensing optical fiber is laid in areas where various vibrations occur, and the vibration signals can be perceived in a distributed manner using the Rayleigh scattering principle.

[0037] An ultra-narrow linewidth laser with a central wavelength of 1550nm and a linewidth of less than 3KHz emits continuous light, the optical power is adjusted by an optical attenuator, and then modulated into pulse light by an acousto-optic modulator with a frequency shift of 200MHz; one output of the analog output card is for the acousto-optic modulator driver to always provide a pulse positive edge trigger signal with a frequency of 5kHz and a pulse width of 20ns, and the trigger mode is post-trigger.

[0038] The pulse light is amplified by an erbium-doped fiber amplifier, filtered out spontaneous radiation noise, and then passes through a circulator and enters the wavelength division multiplexer together with the Raman light output by the Raman amplifier to further amplify the scattered light signal. The optical signal then enters the sensor fiber to sense the external vibration signal.

[0039] At this stage, the other output of the analog output card does not generate a driving signal for the PZT, and the PZT has no effect on the sensing optical fiber.

[0040] Since Rayleigh scattering will return backscattered light, its light field intensity expression is:

[0041] I(t)=E s 2 (t) = [A s cos(2πf s t+φ s )] 2

[0042] =A s 2 / 2+(A s 2 / 2)cos(4πf s t+2φ s )

[0043] Among them, E s , A s 、f s ,φ s are the light field, amplitude, frequency and phase of the backward Rayleigh scattered light. When a vibration signal acts on the sensing fiber, the phase φ s When the light intensity I(t) changes, the data acquisition card with a sampling rate of 100MS / s collects multiple groups of raw data of various vibration signals and transmits them to the computer.

[0044] In addition, in order to obtain the vibration signal threshold, the analog output card generates an amplitude of A for PZT. 0 , frequency is f = f 1 ~f 2 The swept frequency signal:

[0045] x(t)=A 0 cos(2πf)

[0046] Among them, f 1 ~f 2 The frequency range of the φ-OTDR system is within the detectable frequency range. The frequency sweep signal drives the PZT to stretch the optical fiber, causing the optical fiber to deform by Δl, which in turn generates a strain of Δε. The signal amplitude corresponding to the strain is set as the threshold for screening effective vibration signals.

[0047] Embodiment 2

[0048] In one or more embodiments, a Wavenet-based φ-OTDR pattern recognition method is disclosed, comprising:

[0049] Acquire multiple groups of external vibration signals to be predicted through sensing optical fibers;

[0050] The collected vibration signal is processed into time-space data through threshold screening, the data is input into the trained Wavenet neural network model, and the vibration signal type is output;

[0051] The threshold is achieved by using a set sweep frequency signal to drive the PZT to stretch the sensing optical fiber to generate a corresponding strain, and the amplitude corresponding to the strain is used as the threshold for screening effective vibration signals.

[0052] Specifically, the training process of the Wavenet neural network model is as follows: Figure 2 As shown, it mainly includes two parts: training and testing. First, in the training part, after collecting multiple groups of original vibration data, the vibration information is obtained by subtracting the amplitudes of the two groups and taking the absolute value; and compared with the set amplitude threshold, the noise or useless signals other than the vibration signal are filtered out, so as to extract the effective vibration signal.

[0053] A 50×50 time-space array is further formed, with digital labels such as "0", "1", and "2" representing different vibration types, and corresponding digital labels are marked for each data set. The entire data set and the corresponding labels are divided into training sets, validation sets, and test sets in pairs, and shuffled to observe the degree of overfitting of the model. The time-space data set is input into the Wavenet network structure for training.

[0054] In the test part, the trained model structure and weights are loaded to predict the time-space array collected and processed in real time to obtain the label of the corresponding vibration data.

[0055] The specific process includes the following:

[0056] (1) The sensor fiber in the φ-OTDR system senses various external motion signals and collects multiple sets of raw data for each type of motion.

[0057] Specifically, the sensing optical fiber is laid in areas where various vibrations occur, and the vibration signals can be sensed in a distributed manner using the Rayleigh scattering principle.

[0058] In the φ-OTDR system, the ultra-narrow linewidth laser emits continuous light, which is adjusted in optical power by an optical attenuator, modulated by an acousto-optic modulator, amplified by an erbium-doped fiber amplifier, and then filtered out by a filter to remove the spontaneous radiation noise. The output pulse light passes through a circulator and enters the wavelength division multiplexer together with the Raman light output by the Raman amplifier, and then enters the sensing fiber.

[0059] The analog output card outputs two signals, one of which always provides a trigger signal for the AOM driver, and the other output signal can control whether the PZT works. At this time, the analog output card does not output a signal to the PZT, and the PZT has no effect on the sensing fiber.

[0060] When a vibration signal acts on the sensing optical fiber, the phase of the backward Rayleigh scattered light returned from the sensing optical fiber changes, and the light intensity changes accordingly. Multiple sets of original data of various vibration signals are obtained through photoelectric detection and data acquisition.

[0061] (2) The analog output card generates a cosine signal to drive the PZT, simulating the effect of an external vibration signal. The strain amount generated by the PZT on the sensing optical fiber is set as the threshold for screening effective vibration signals.

[0062] Specifically, the analog output card generates an amplitude of A for the PZT 0 , frequency is f 1 ~f 2 The swept frequency signal, where f 1 ~f 2 The frequency range of the φ-OTDR system is within the detectable frequency range. The frequency sweep signal drives the PZT to stretch the sensing fiber to generate corresponding strain, and the amplitude corresponding to the strain is set as the threshold for screening effective vibration signals.

[0063] (3) The collected multiple sets of data are threshold-screened and processed into time-space data sets, which are divided into training sets, validation sets, and test sets, and shuffled.

[0064] Specifically, Figure 3 A flowchart for processing the collected raw data is given. After collecting n groups of raw data, each group has m spatial sampling points according to the length of the sensing optical fiber. The difference between the two groups is taken and the absolute value is taken to obtain n-1 groups of difference data. Since the pulse trigger frequency is 5KHz, the length in the time dimension is:

[0065]

[0066] In each group of difference data, it is determined in turn whether the amplitude difference between two adjacent groups of sampling points is greater than the set threshold. If so, the effective vibration signal action time and the specific position on the sensing optical fiber are obtained, and the group and the 24 groups before and the 25 groups after the group are taken to construct the time information of the vibration signal;

[0067] Take the sampling point j with the largest amplitude in the group, and take j and the 24 points before j, and the 25 points after j to construct the spatial information of the vibration signal, forming a 50×50 time-space two-dimensional array;

[0068] Use numbers such as "0", "1", and "2" to indicate the corresponding action label for each array. Divide all arrays and corresponding labels into training sets, validation sets, and test sets, shuffle them, and input them into the Wavenet neural network for training.

[0069] (4) The time-space data set is input into the Wavenet neural network model, trained multiple times until the model is optimal, and stored.

[0070] like Figure 4 The stacked structure of Wavenet causal convolutional layers is shown. Based on the two-dimensional time-space dataset, in order to analyze the temporal nature of the vibration signal in the time dimension, the joint probability density of the entire input sequence can be expressed as:

[0071]

[0072] Among them, x 1 ,...,x t is the vibration data at a certain time t. The vibration data points are all related to the data at the previous time, and the causal convolution structure in Wavenet is suitable for analyzing time series data. When the input is a time series data point x = {x 1 ,x 2 ,...,x 50}, through causal convolution, its output is only related to the current moment and the previous moment, but not to the future moment. When used for the classification of vibration signals, it can effectively extract the temporal characteristics of the vibration signal. The Wavenet structure is composed of multiple layers of causal convolution layers. The number of causal convolution kernels is 16, the size is 1, and the step size is 1. There is no pooling layer in the convolution network, which can avoid the loss of effective information to a certain extent. However, the problem is that when the number of convolution layers is fixed, the model's perception range of the input sequence is limited.

[0073] like Figure 4 As shown in the figure, a model with 4 convolutional layers has only 5 receptive fields. To train sufficiently long time series samples, multiple convolutional layers need to be stacked, or larger filters are needed to increase the receptive field long enough, but this increases the computing resources.

[0074] like Figure 5The figure shows the Wavenet multi-layer hole convolution structure, which enables the model to obtain a larger perception range with the same number of convolution layers. That is, some zero-valued weights are added to the convolution kernel at fixed intervals (called hole coefficients) to sparse the convolution kernel, so that the convolution kernel can be applied to an area larger than its own length, thereby obtaining a larger perception field with the same number of convolution layers. Figure 5 As shown in the figure, when the convolution kernel dilation coefficients are 1, 2, 4, and 8, the input perception range increases to 16, and the perception field increases exponentially with the number of convolution layers. To further expand the perception range, the dilated convolutions with dilation coefficients of 1, 2, 4, and 8 are regarded as a block with a convolution kernel size of 7; repeating the block three times will generate 12 layers of causal dilated convolutions.

[0075] like Figure 6 The figure shows the Wavenet residual network structure. Each layer of the residual network contains a dilated convolution structure with different dilation coefficients, thus generating 12 layers of residual blocks. The input sequence first passes through a one-dimensional causal convolution and then enters the first layer of residual blocks. After passing through a dilated convolution layer with a dilation coefficient of 1, the input is further activated by a gated activation unit to increase the nonlinearity between convolution layers and enhance the expressiveness of the model:

[0076] z=tanh(W f,k *x)⊙σ(W g,k *x)

[0077] Among them, * represents convolution operation, ⊙ represents point multiplication operation, tanh and σ represent tanh activation function and sigmoid activation function respectively, k is the number of layers, f and g represent filter and gate respectively, W is the learnable convolution kernel, and x is the input.

[0078] After activation, the number of channels is reduced by a 1×1 convolution kernel; then its output is divided into two paths: one path is added to the input of the residual block of this layer to form a residual network; the other path is directly output as a skip connection and accumulated with each layer. The residual network directly transmits the input of the shallow network to the output of the deep layer, and the identity mapping is converted into the difference between the two as the input of the residual block of the next layer. The introduction of direct mapping of the shallow network can more efficiently execute the transformed identity mapping, so that the model converges faster and trains a deeper network structure.

[0079] like Figure 7 The figure shows the flow chart of the Wavenet pattern recognition method. The 50×50 time-space two-dimensional array is input to the causal convolution layer, and then enters the first residual block. In this residual block, it first passes through the dilated convolution layer with a dilated coefficient of 1 and is input to the gate activation unit tanh (W f,k *x)⊙σ(W g,k*x) is further activated to increase the nonlinearity between convolutional layers, and then input to the 1×1 convolutional layer to reduce the number of channels. Then its output is divided into two paths: one is added to the input of the residual block of this layer to form a residual network; the other is directly output as a skip connection and accumulated with each layer. The output of the residual network is passed to the next layer in turn with a hole coefficient of 2 i The residual block is repeated three times until the hole coefficient increases to 8. The output of the last residual block is accumulated with the skip connection output of each residual block, and the accumulated result is subjected to one-dimensional causal convolution, batch normalization, and then one-dimensional global maximum pooling. Finally, the Softmax activation function is used for classification, the probability of each vibration type is output, and the final prediction label is given.

[0080] The newly collected test data is formed into a time-space array and input into the trained model to obtain the pattern recognition result.

[0081] Embodiment 3

[0082] In one or more embodiments, a terminal device is disclosed, including a server, the server including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the Wavenet-based φ-OTDR pattern recognition method in Embodiment 2 when executing the program. For the sake of brevity, it will not be described in detail here.

[0083] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0084] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0085] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software.

[0086] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A Wavenet-based φ-OTDR pattern recognition method, implemented based on a Wavenet-based φ-OTDR pattern recognition system, characterized in that: The system comprises: a laser, an optical attenuator, an acousto-optic modulator, an erbium-doped fiber amplifier and a filter connected in sequence, the output end of the filter is connected to a circulator, the output ends of the circulator and the Raman amplifier are respectively connected to a wavelength division multiplexer, and the output end of the wavelength division multiplexer is connected to a sensing optical fiber through a PZT; The method comprises: Acquire multiple groups of external vibration signals to be predicted through sensing optical fibers; The collected vibration signal is processed into time-space data through threshold screening, and the data is input into the trained Wavenet neural network model to output the vibration signal type; wherein the threshold is to use the set sweep frequency signal to drive the PZT action, stretch the sensing optical fiber to generate the corresponding strain, and the amplitude corresponding to the strain is used as the threshold for screening effective vibration signals; The data is input into the trained Wavenet neural network model to output the vibration signal type. The specific process includes: The temporal-spatial data is input into the causal convolution layer and then into the first residual block. In this residual block, it first passes through the dilated convolution layer and is input into the gate activation unit. It is then input into the convolution layer to reduce the number of channels, and then its output is divided into two paths: one path is added to the input of the residual block of this layer to form a residual network; the other path is directly output as a skip connection and accumulated with each layer. The output of the residual network is passed to the next layer of residual blocks in sequence until the hole coefficient increases to the set value, and the process is repeated for the set number of times; The output of the last layer of residual blocks is added to the skip connection output of each layer of residual blocks. The accumulated result is subjected to one-dimensional causal convolution and batch normalization, and then one-dimensional global maximum pooling. Finally, the Softmax activation function is used for classification, the probability of each vibration type is output, and the final prediction label is given.

2. A Wavenet-based φ-OTDR pattern recognition method as claimed in claim 1, characterized in that: The AOM is driven by an AOM driver.

3. A Wavenet-based φ-OTDR pattern recognition method as claimed in claim 2, characterized in that: Also includes: The analog output card has output terminals connected to the acousto-optic modulator driver and the PZT respectively.

4. A Wavenet-based φ-OTDR pattern recognition method as claimed in claim 1, characterized in that: The output end of the circulator is also connected to a photodetector.

5. A Wavenet-based φ-OTDR pattern recognition method as claimed in claim 4, characterized in that: The photoelectric detector is connected to the computer through a data acquisition card.

6. A Wavenet-based φ-OTDR pattern recognition method as claimed in claim 1, characterized in that: The laser is an ultra-narrow line width laser.

7. A Wavenet-based φ-OTDR pattern recognition method as claimed in claim 1, characterized in that: The collected vibration signal is processed into time-space data through threshold screening. The specific process includes: Obtain n groups of original data, each group has m spatial sampling points; Difference the two groups and take the absolute value to get n-1 groups of difference data; In each group of difference data, it is determined in turn whether the amplitude difference between two adjacent groups of sampling points is greater than the set threshold. If so, the effective vibration signal action time and the specific position on the sensing optical fiber are obtained, and the group and the set number of groups before and after the group are taken to construct the time information of the vibration signal; Take the sampling point j with the largest amplitude in the group, and take j and a set number of points before and after j to construct the spatial information of the vibration signal.

8. A terminal device, comprising a server, wherein the server comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the Wavenet-based φ-OTDR pattern recognition method described in any one of claims 1 to 7 is implemented.

Images