Event positioning method, system and equipment based on OTDR (Optical Time Domain Reflectometer) data and medium

Abstract

The invention relates to the technical field of optical fiber sensing, and provides an event positioning method, system and device based on OTDR data and a medium, and the method comprises the steps: obtaining an OTDR data input condition, and generating adversarial network denoising; slicing the de-noised OTDR data and the variance and differential curve thereof, wherein the length is between one and a plurality of pulse widths; respectively inputting the sliced data into three feature extraction modules, wherein each feature extraction module is a series structure of a convolution layer, a batch normalization layer and an activation layer; combining the output features and inputting the combined output features into a high-efficiency channel attention layer to obtain an initial probability sequence; the initial probability sequence is binarized through a threshold value x, a binary sequence L is formed, and a point with the value being 1 in the binary sequence L represents that an event exists. According to the method, the problem that an OTDR event positioning algorithm is sensitive to a threshold value can be solved.

Description

Technical Field

[0001] The present invention belongs to the field of optical fiber sensing technology, and in particular relates to an event positioning method, system, device and medium based on OTDR data. Background Art

[0002] OTDR can be used to measure fiber attenuation, connector loss, fiber fault point location, and understand the loss distribution along the length of the fiber. It is an indispensable tool in optical cable construction, maintenance and monitoring. In different applications of OTDR (for example, intrusion detection, co-cable detection), event recognition is the technical foundation and needs to provide necessary guarantees for the accuracy of subsequent algorithms.

[0003] Traditional OTDR event location algorithms are generally based on threshold segmentation or wavelet decomposition, both of which require threshold selection; threshold selection is generally based on manual work, and the thresholds in different data collection scenarios are different, requiring manual adaptation of thresholds for different scenarios. There are problems of sensitivity to thresholds and reliance on manual selection. Summary of the invention

[0004] To solve the above problems, the present disclosure provides an event location method, system, device and medium based on OTDR data, which adopts an event location network for event location, and can solve the problem that the OTDR event location algorithm is sensitive to thresholds and relies on manual selection.

[0005] The following is the scheme of the present invention:

[0006] Obtain OTDR data, input the OTDR data into a conditional generative adversarial network, and output denoised OTDR data;

[0007] Slice the OTDR data and its variance and differential curves, with the slice length ranging from one pulse width to several pulse widths; input the sliced ​​OTDR data, the variance of the OTDR data, and the differential curve of the OTDR data into three feature extraction modules respectively; the feature extraction module is a series-connected convolution layer, batch normalization layer, and activation layer structure;

[0008] After merging the features output by the three feature extraction modules, the features are input into the efficient channel attention layer, which outputs an initial probability sequence of 0-1 with the same length as the input, representing the probability of an event at each point.

[0009] Binarize the initial probability sequence with the threshold x, take 1 for points greater than x and take 0 for points less than x, and get the probability sequence L;

[0010] In the probability sequence L, a point with a value of 1 represents an event.

[0011] Furthermore,

[0012] The step of inputting OTDR data into a conditional generative adversarial network and outputting denoised OTDR data comprises:

[0013] Normalize OTDR data;

[0014] Slice the normalized OTDR data, with the slice length being greater than a pulse width;

[0015]

[0016] The sliced ​​OTDR data is input into the generator of the conditional generative adversarial network, and the denoised OTDR data is output.

[0017] Furthermore,

[0018] The construction of the conditional generative adversarial network includes:

[0019] Get the conditional generative adversarial network of the pix2pix structure;

[0020] The two-dimensional convolutional layers and deconvolutional layers of the conditional generative adversarial network of the pix2pix structure are replaced by one-dimensional convolutional layers and deconvolutional layers.

[0021] Furthermore,

[0022] The normalization of the sliced ​​OTDR data comprises:

[0023] Convert the OTDR data to dB units and scale the converted OTDR data to between 0 and 1:

[0024]

[0025] Among them, X i , X i ′ are the values ​​of the i-th point of the original OTDR data and the normalized OTDR data, respectively.

[0026] Furthermore,

[0027] The training of the conditional generative adversarial network includes:

[0028] Get the original OTDR data and the corresponding real denoised data {X denoise};

[0029] The original OTDR data is input into the generator Gen of the conditional generative adversarial network, and the false denoised data {X pred};

[0030] Set the real denoising data {X denoise} has a label of 0, and the false denoising data {Xpred} is labeled 1, construct the loss function of the discriminator, and train the discriminator;

[0031] Based on the pseudo denoising data {X pred} Input the value between 0 and 1 obtained by the discriminator to calculate the generator loss and train the generator;

[0032] The loss function of the discriminator is:

[0033] Loss(Dis)=CrossEntropy(Dis(X pred ,X),1)

[0034] +CrossEntropy(Dis(X denoise ,X),0)

[0035] Among them, CrossEntropy(Dis(X pred ,X),1) is to remove the false denoising data {X pred}The cross entropy loss between the output of the discriminator and 1 after inputting the discriminator; CrossEntropy(Dis(X denoise ,X),0) is the real denoising data {X denoise The cross entropy loss between the output of the discriminator and 0 after input into the discriminator;

[0036] The generator loss is:

[0037] Loss(Gen)=CrossEntropy(Dis(X pred ,X),0)+L1Loss(X pred ,X denoise )

[0038] Among them, CrossEntropy(Dis(X pred ,X),0) is to remove the false denoising data {X pred}The cross entropy loss between the output of the discriminator and 0 after inputting the discriminator; L1Loss(X pred ,X denoise ) is the real denoising data {X denoise} and the fake denoised data {X pred}L1 loss.

[0039] Furthermore,

[0040] The acquisition of the original OTDR data and the corresponding real denoised data {X denoise};include:

[0041] The experiment measured the noise-free OTDR of the fiber section, flattened the curve after the fiber tail to the bottom, and obtained the real denoised data {Xdenoise};

[0042] The real denoised data {X ddnoise} is randomly cut into segments of length L to obtain a data set {X}; the sliced ​​data set without events is recorded as {X0}, and the sliced ​​data set with events is recorded as {X1}, {X0} and {X1} are equal in size;

[0043] The noise of OTDR data is described by Gaussian distribution, and a simulated noise curve is generated to obtain simulated noise; a first-order Butterworth filter is set to filter the simulated noise curve to remove noise components higher than the cutoff frequency;

[0044] The data set {X0} and the data set {X1} are respectively superimposed with an equal amount of noise with different signal-to-noise ratios to serve as the original OTDR data.

[0045] Furthermore,

[0046] After obtaining the binary sequence L, the binary sequence L is filtered using morphological operations to obtain a filtered binary sequence L1; which includes:

[0047] Use the closing operation to merge the disconnected recognition results in the probability sequence L; where the template length of the closing operation is L close <Pulse Width , Pulse Width is the pulse width of the OTDR data;

[0048] Use the open operation to remove the probability sequence L whose length is less than the template length L of the open operation open The filtered probability sequence L1 is obtained; L open =Pulse width .

[0049] An event location system for OTDR data, characterized by comprising:

[0050] A data acquisition module is used to acquire OTDR data, input the OTDR data into a conditional generative adversarial network, and output denoised OTDR data;

[0051] The event location module is used to slice the denoised OTDR data and its variance and differential curves, with the slice length ranging from one pulse width to several pulse widths; the sliced ​​OTDR data, the variance of the OTDR data, and the differential curve of the OTDR data are respectively input into three feature extraction modules; the feature extraction modules are a series of convolutional layers, batch normalization layers, and activation layers;

[0052] After merging the features output by the three feature extraction modules, the features are input into the efficient channel attention layer, which outputs an initial probability sequence of 0-1 with the same length as the input, representing the probability of an event at each point.

[0053] The initial probability sequence is binarized using the threshold x, with the values ​​greater than x being 1 and the values ​​less than x being 0, to obtain a binary sequence L. In the binary sequence L, the points with a value of 1 represent events.

[0054] Compared with the prior art, the present invention has the following advantages:

[0055] In terms of data input, the present invention adopts original OTDR, variance and differential data as multiple input sources. Since the OTDR data itself contains information such as the loss along the optical fiber, the variance can reflect the discreteness of the data. The variance usually changes significantly in the event occurrence area. The differential data is sensitive to the change trend of the data, and the differential curve in the event occurrence area will fluctuate significantly. Therefore, by using these three types of data at the same time, the event characteristics can be fully captured from multiple angles, avoiding inaccurate event judgment due to the limitations of a single data, thereby reducing dependence on precise thresholds.

[0056] In the feature extraction stage, a feature extraction module with three serial structures of "convolution layer + batch normalization + activation layer" is used to process the three types of data respectively; the convolution layer performs convolution operation on the data by sliding convolution kernel, which can extract local features at different positions; the batch normalization layer normalizes each batch of data in the feature dimension, which accelerates the convergence speed of network training and makes the network more adaptable to data of different magnitudes, thereby improving the stability of the model; the activation layer introduces nonlinear activation function to increase nonlinear expression ability of the model, so that the network can learn more complex data relationships and better fit event characteristics; this structural design enables the network to deeply mine the potential information in the data, laying the foundation for subsequent accurate judgment of events.

[0057] The efficient channel attention layer adaptively adjusts features based on the channel dimension. By calculating the importance weight of each channel, the model can automatically focus on channels containing key event information and suppress those channels that contribute less to event judgment. During the training process, the network automatically learns the importance of different channels to event judgment in different scenarios, thereby highlighting important features and further improving the effectiveness of feature extraction.

[0058] After that, the network output is a sequence of values ​​0-1, indicating the probability of an event at each point. Compared with the binary result obtained by the traditional algorithm directly based on the threshold segmentation, the event judgment result is finally obtained by binarization through the threshold x. Since the previous network structure has fully extracted and optimized the features, in different data collection scenarios, the model can stably output the probability sequence based on the characteristics of the data itself. It avoids the tedious operation of frequently adjusting the threshold for different scenarios, and improves the versatility and accuracy of the event location algorithm.

[0059] Other features and advantages of the present disclosure will be described in the following description, and partly become apparent from the description, or be understood by implementing the present disclosure. The purpose and other advantages of the present disclosure can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1 The schematic diagram of the algorithm flow framework of the present invention is shown;

[0062] Figure 2 A schematic diagram of a conditional generative adversarial network of the present invention is shown;

[0063] Figure 3 A schematic diagram of the method for generating a denoising network training set is shown;

[0064] Figure 4 A schematic diagram of the input and output results of the denoising network is shown;

[0065] Figure 5 A schematic diagram of the event location network structure is shown;

[0066] Figure 6 A schematic diagram showing the effect of the denoising network on the output of the event localization network;

[0067] Figure 7 A schematic diagram of the morphological operation results is shown. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0069] Figure 1 A schematic diagram of the method according to the present invention is shown, and the specific implementation details of the present invention include:

[0070] 1. Step 1: Obtain OTDR data, normalize and slice the OTDR data, input the slicing data into the denoising network, and output the denoised OTDR data.

[0071] 1) Specifically, the denoising network uses a conditional generative adversarial network

[0072] The conditional generative adversarial network is a variant of the generative adversarial network. Figure 2 The structure of the Conditional Generative Adversarial Network (CGAN) consists of two sub-networks: the generator and the discriminator.

[0073] The input of the generator is a random vector and a condition, and it will generate a fake sample corresponding to the random vector in the latent space based on the given condition.

[0074] The input of the discriminator is a data sample (which can be a true sample or a false sample) and a condition, and it will determine whether the sample is true or false under the given conditions. The output of a true sample is 0, and the output of a false sample is 1. The discriminator not only needs to determine the authenticity of the sample, but also needs to determine the correlation between the sample and the condition.

[0075] 2) Specific:

[0076] The OTDR denoising network of the present invention is a conditional generative adversarial network based on the pix2pix structure (since OTDR data is one-dimensional data, the two-dimensional convolutional layer and deconvolution layer in the pix2pix network need to be replaced by one-dimensional convolutional layer and deconvolution layer.) The input condition in the model is the original noisy OTDR data sequence, and the generator output is an equal-length denoised OTDR data sequence, and no random vector input is required in the pix2pix network.

[0077] Normalize the OTDR data:

[0078] Since the neural network model requires small value input, and in order to make the model generalize to OTDR data with different dynamic ranges, the input of the model needs to be normalized to 0-1. The normalization process requires first converting the OTDR data to dB units, and then using the formula to scale the OTDR data to between 0 and 1:

[0079]

[0080] Among them, X i , X i ′ are the values ​​of the i-th point of the original OTDR data and the normalized OTDR data, respectively.

[0081] OTDR data slice:

[0082] In terms of model performance, the denoising effect is better when the input sequence length of the model is longer. However, the length of OTDR data is generally thousands to hundreds of thousands of points. In order to limit the model size and computational complexity, the OTDR data needs to be sliced.

[0083] The selection of slice length is related to the pulse width of the OTDR. Generally speaking, the slice length needs to be greater than a pulse width to include information about the edges before and after the event, and the pix2pix network requires the input length to be an integer power of 2. The pulse width is the minimum number of points for an event on the OTDR curve:

[0084] 3) Specifically, the training of the denoising network includes the following steps:

[0085] 3.1) Input the normalized OTDR data slice {X} into the generator Gen to generate the model denoised data {X pred}.

[0086] X pred =Gen(X)

[0087] 3.2) Mixed true samples {X denoise}(real denoised data) and fake samples {X pred}(model denoised data), the true sample label is 0, the false sample label is 1, and the discriminator is trained. The discriminator loss function = the cross entropy loss between the discriminator (model denoised data) and 1 + the cross entropy loss between the discriminator (model target data) and 0, and the discriminator is trained.

[0088] Loss(Dis)=CrossEntropy(Dis(X pred ,X),1)

[0089] +CrossEntropy(Dis(X denoise ,X),0)

[0090] 3.3) The false sample is input into the discriminator to obtain a value between 0 and 1 (the target value is 0), the generator loss is calculated, and the generator is trained. The generator loss function = L1 loss between the model denoised data and the target denoised data + the cross entropy loss between the discriminator (model denoised data) and 0, where the former represents the correlation between the model denoised data and the original OTDR data, and the latter represents the authenticity of the model denoised data.

[0091] Loss(Gen)=CrossEntropy(Dis(X pred ,X),0)+L1Loss(X pred ,X denoise )

[0092] 4) Dataset acquisition:

[0093] Training the denoising network requires a training data set in the form of (original OTDR data, actual denoised OTDR data), but the actual denoised data needs to be obtained using traditional denoising algorithms such as wavelet transform, and some manual selection of thresholds and wavelet bases cannot be avoided.

[0094] Therefore, this patent adopts a training data set in the form of (original noise-free OTDR data + simulated noise, original noise-free OTDR data), which can be obtained by experimentally measuring the noise-free OTDR data and superimposing the simulated noise.

[0095] The steps to obtain the training data set are as follows: Figure 3 As shown, the specific steps are as follows:

[0096] 4.1) Experimentally measure the noise-free OTDR of the fiber segment, and flatten the curve after the fiber tail to the bottom. Randomly cut the original OTDR data into segments of length L, which are the original data set {X}. The slice data set without events is recorded as {X0}, and the slice data set with events is recorded as {X1}. In order to balance the training data set, {X0} and {X1} need to be equal in size.

[0097] 4.2) The noise in the OTDR signal is mainly caused by the thermal noise and dark current of the detector. These noises generally present a Gaussian distribution, so Gaussian distribution is usually used to model and describe the noise of OTDR data. Generate a simulated noise curve: n[t] = σ n ·η[t], where η[t] is a random variable that follows a standard normal distribution, σ n is the standard deviation of the noise power,

[0098]

[0099] P s , Pn , SNR are the maximum signal power, maximum noise power and OTDR signal-to-noise ratio respectively. Different SNRs can be covered when generating noise curves.

[0100] 4.3) In order to improve measurement accuracy and reduce the impact of noise on measurement results, low-pass filtering technology is often used in OTDR systems to eliminate high-frequency noise components and improve the signal-to-noise ratio. During the filtering process, a cutoff frequency is usually set, and noise components above this frequency will be filtered out or attenuated, thereby retaining valid signals with lower frequencies. The choice of cutoff frequency is usually determined based on specific application requirements and system performance to balance signal fidelity and noise suppression effects.

[0101] After the noise curve and the original OTDR data are superimposed, a first-order Butterworth filter is set to filter the signal to remove noise components above the cutoff frequency in order to simulate the signal processing of the OTDR single disk. The cutoff frequency is usually in the range of several hundred kHz to several GHz.

[0102] 4.4) The dataset {X0} and the dataset {X1} are superimposed with an equal number of noises of different SNRs (30dB, 60dB, 100dB, 1000dB). When SNR=60dB, it is close to the experimental data, SNR=30 / 100dB can simulate the situation of larger / smaller noise, and SNR=1000dB can simulate the situation of no noise.

[0103] 4.5) The (original data + simulated noise, original data) data pair is stretched, compressed, and other morphological changes are performed along the x-axis or y-axis to simulate different OTDR pulse widths and powers, thereby enhancing the model's generalization ability for OTDR data in different scenarios.

[0104] 2. Step 2: Input the denoised OTDR data together with its variance and difference curves into the event location network, and output a probability sequence indicating whether there is an event at each point.

[0105] Event Location Network (B):

[0106] The structure of the event localization network is as follows: Figure 5 The event location network is described as follows:

[0107] The network uses original OTDR, variance, and differential data as multiple inputs, because all three types of data contain event information, and the variance and differential values ​​are larger in areas with events. The original OTDR data, variance curve, and differential curve are used as inputs to the network at the same time, and the positioning effect is better than the single input network that only uses the original OTDR curve.

[0108]

[0109] Among them, Diff i 、Var i are the i-th point of the difference curve and variance curve respectively, L diff , L var are the step sizes for calculating the difference and variance respectively, and their values ​​can be equal to 1 / 2 of the pulse width, that is, considering the difference and variance within a minimum event range.

[0110] Input data slice:

[0111] Similar to the denoising network, the longer the input sequence is, the better the positioning network is. However, in order to limit the amount of computation of the model, the three input sequences need to be sliced, and the slice length is preferably between one pulse width and several pulse widths.

[0112] Model processing data logic:

[0113] Each input uses three "convolution layer + batch normalization + activation layer" modules for feature extraction. Since the input and output of the model are sequences of equal length, there is no need to use a pooling layer. Assuming that the lengths of the convolution kernels of the three convolution layers are L1, L2, and L3, respectively, after each convolution layer, each point contains information about the surrounding L1, L1+L2, and L1+L2+L3 points. In order to consider the information around a point, the size of L1+L2+L3 should be greater than the minimum number of points of an event on the OTDR curve.

[0114] After the convolutional layer, the importance of different channels is adaptively adjusted by using the Efficient Channel Attention (ECA) module.

[0115] The output is a sequence of 0-1 values ​​of the same length as the input, representing the probability of an event at each point.

[0116] The output sequence is binarized with a threshold of 0.5, and points with a value greater than 0.5 are set to 1 (event present), and points with a value less than 0.5 are set to 0 (no event).

[0117] An example of the output of the event localization network is Figure 6 As shown in the figure. (a) is the normalized original OTDR data, the red dot indicates the location of the real event, (b) is the output result obtained by directly inputting the original OTDR and its variance and differential data into the event location network, the red dot indicates the event location located by the algorithm, and it can be seen that there are many false alarms in the noise floor part, (c) is the denoised OTDR data output by the denoising network, (d) is the output result obtained by inputting the denoised OTDR and its variance and differential data into the event location network, and it can be seen that the false alarm events in the noise floor part are eliminated.

[0118] 3. Step three: Use morphological operations to filter the event location results to obtain accurate event location results.

[0119] Morphological operations (C):

[0120] The closing operation of binary morphological operation (dilation first, then erosion) can merge broken target objects without changing the size of the original target objects. The opening operation of binary morphological operation (erosion first, then dilation) can separate adhered objects and remove small particle noise without changing the size of the original target objects.

[0121] The output of the binarized event localization network is a one-dimensional 01 sequence, where 0 represents a point without an event and 1 represents a point with an event.

[0122] When there is noise interference in the middle of an event, the recognition result may be disconnected, and an event is split into two or more adjacent events. Therefore, the closed operation can be used to merge the disconnected recognition results. Since OTDR data is one-dimensional data, the expansion and erosion templates both use one-dimensional templates. Assume that the template length of the closed operation is L close , then the closing operation can merge the distances less than L clpse When the distance between two events exceeds one pulse width, they can be considered as two independent events and cannot be merged. Therefore, there is L clpse <Pulse Width In practice, close 1 / 2 pulse width can be used.

[0123] Noise interference in OTDR data may cause false alarms of events, so the opening operation can be used to remove events caused by noise. Assume that the template length of the opening operation is L open , then the open operation can remove the length less than L open Since the pulse width is the minimum number of points of an event on the OTDR curve, an event with a length less than 1 pulse width can be regarded as a false alarm event caused by noise. open =Pulse Width Considering that the end of an event may not be fully identified, in practice L open A smaller value may be used, such as 1 / 2 or 1 / 4 of the pulse width.

[0124] The result of morphological operation is as follows Figure 7 As shown, (a) is the output result of the event localization network, and (b) is the event recognition result after morphological operation. It can be seen that the last disconnected event is merged into a wider event.

[0125] In summary, the solution of the present invention solves:

[0126] 1. The OTDR signal will bring noise due to factors such as light source, detector, and environment. It is easy to be disturbed by noise when locating events, which may cause missed reports and false reports. Through a denoising network based on conditional generative adversarial network, after inputting the original OTDR data, the denoised OTDR data can be output to achieve the denoising function. The influence of noise on the accuracy of event location results is reduced, and the missed reports and false reports are reduced.

[0127] 2. Based on the sensitivity of the threshold value and the reliance on manual selection in the OTDR event recognition algorithm, an event location network based on a convolutional neural network is used to input the denoised OTDR data and its variance and difference curves. The model can automatically learn, extract features, and output the probability of an event at each point. When facing OTDR data of different scenarios, there is no need to manually set the threshold value. For data of different scenarios, the accuracy of event recognition is more stable;

[0128] 3. The event location results are further denoised through an event filtering algorithm based on morphological operations.

[0129] Based on the method of the present invention, the embodiment of the present disclosure also provides a system corresponding to the above method, which includes:

[0130] A data acquisition module is used to acquire OTDR data, input the OTDR data into a conditional generative adversarial network, and output denoised OTDR data;

[0131] The event location module is used to slice the OTDR data and its variance and differential curves, with the slice length ranging from one pulse width to several pulse widths; the sliced ​​OTDR data, the variance of the OTDR data, and the differential curve of the OTDR data are respectively input into three feature extraction modules; the feature extraction modules are a series of convolutional layers, batch normalization layers, and activation layers;

[0132] After merging the features output by the three feature extraction modules, the features are input into the efficient channel attention layer, which outputs an initial probability sequence of 0-1 with the same length as the input, representing the probability of an event at each point.

[0133] The initial probability sequence is binarized using the threshold x, with the values ​​greater than x being 1 and the values ​​less than x being 0, to obtain a binary sequence L. In the binary sequence L, the points with a value of 1 represent events.

[0134] Based on the same inventive concept as the above disclosed content, an embodiment of the present disclosure also provides a device corresponding to the above method, which includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above method.

[0135] It should be noted that the electrical connection between the above-mentioned units does not necessarily mean the connection between lines. An indirect connection method can be applied to the embodiments of the present disclosure as long as the purpose of the present disclosure is achieved.

[0136] Based on the same inventive concept, the present disclosure also provides a computer storage medium on which executable instructions are stored. When the instructions are executed by a processor, the processor performs the above method.

[0137] Although the present disclosure has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. An event location algorithm based on OTDR data, characterized in that: include: Obtain OTDR data, input the OTDR data into a conditional generative adversarial network, and output denoised OTDR data; Slice the denoised OTDR data and its variance and differential curves, with the slice length ranging from one pulse width to several pulse widths; input the sliced ​​OTDR data, the variance of the OTDR data, and the differential curve of the OTDR data into three feature extraction modules respectively; the feature extraction module is a series-connected convolution layer, batch normalization layer, and activation layer structure; After merging the features output by the three feature extraction modules, the features are input into the efficient channel attention layer, which outputs an initial probability sequence of 0-1 with the same length as the input, representing the probability of an event at each point. The initial probability sequence is binarized using the threshold x, with the values ​​greater than x being 1 and the values ​​less than x being 0, to obtain a binary sequence L. In the binary sequence L, a point with a value of 1 represents an event.

2. The event location algorithm based on OTDR data according to claim 1, characterized in that: The step of inputting OTDR data into a conditional generative adversarial network and outputting denoised OTDR data comprises: Normalize OTDR data; Slice the normalized OTDR data, with the slice length being greater than a pulse width; The sliced ​​OTDR data is input into the generator of the conditional generative adversarial network, and the denoised OTDR data is output.

3. An event location algorithm based on OTDR data according to claim 2, characterized in that: The construction of the conditional generative adversarial network includes: Get the conditional generative adversarial network of the pix2pix structure; The two-dimensional convolutional layers and deconvolutional layers of the conditional generative adversarial network of the pix2pix structure are replaced by one-dimensional convolutional layers and deconvolutional layers.

4. The event location algorithm based on OTDR data according to claim 2, characterized in that: The OTDR data is normalized; comprising: Convert the OTDR data to dB units and scale the converted OTDR data to between 0 and 1: Among them, X i , X i i are the values ​​of the i-th point of the original OTDR data and the normalized OTDR data, respectively.

5. The event location algorithm based on OTDR data according to claim 3, characterized in that: The training of the conditional generative adversarial network includes: Get the original OTDR data and the corresponding real denoised data {X denoise }; The original OTDR data is input into the generator Gen of the conditional generative adversarial network, and the false denoised data {X pred }; Set the real denoising data {X denoise } has a label of 0, and the false denoising data {X pred } is labeled 1, construct the loss function of the discriminator, and train the discriminator; Based on the pseudo denoising data {X pred } Input the value between 0 and 1 obtained by the discriminator to calculate the generator loss and train the generator; The loss function of the discriminator is: Loss(Dis)=CrossEntropy(Dis(X pred ,X),1)+CrossEntropy(Dis(X denoise ,X),0) Among them, CrossEntropy(Dis(X pred ,X),1) is to remove the false denoising data {X pred }The cross entropy loss between the output of the discriminator and 1 after inputting the discriminator; CrossEntropy(Dis(X denoise ,X),0) is the real denoising data {X denoise The cross entropy loss between the output of the discriminator and 0 after input into the discriminator; The generator loss is: Loss(Gen)=CrossEntropy(Dis(X pred ,X),0)+L1Loss(X pred ,X denoise ) Among them, CrossEntropy(Dis(X pred ,X),0) is to remove the false denoising data {X pred }The cross entropy loss between the output of the discriminator and 0 after inputting the discriminator; L1Loss(X pred ,X denoise ) is the real denoising data {X denoise } and the fake denoised data {X pred }L1 loss.

6. The event location algorithm based on OTDR data according to claim 5, characterized in that: The acquisition of the original OTDR data and the corresponding real denoised data {X denoise }; include: The experiment measured the noise-free OTDR of the fiber section, flattened the curve after the fiber tail to the bottom, and obtained the real denoised data {X denoise }; The real denoised data {X denoise } is randomly cut into segments of length L to obtain a data set {X}; the sliced ​​data set without events is recorded as {X0}, and the sliced ​​data set with events is recorded as {X1}, {X0} and {X1} are equal in size; The noise of OTDR data is described by Gaussian distribution, and a simulated noise curve is generated to obtain simulated noise; a first-order Butterworth filter is set to filter the simulated noise curve to remove noise components higher than the cutoff frequency; The data set {X0} and the data set {X1} are respectively superimposed with an equal amount of noise with different signal-to-noise ratios to serve as the original OTDR data.

7. The event location algorithm based on OTDR data according to claim 1, characterized in that: After obtaining the binary sequence L, the binary sequence L is filtered using morphological operations to obtain a filtered binary sequence L1; It includes: Use the closing operation to merge the disconnected recognition results in the probability sequence L; where the template length of the closing operation is L close <Pulse Width , Pulse Width is the pulse width of the OTDR data; Use the open operation to remove the probability sequence L whose length is less than the template length L of the open operation open The filtered probability sequence L1 is obtained; L open =Pulse Width .

8. An event location system for OTDR data, characterized in that: include: A data acquisition module is used to acquire OTDR data, input the OTDR data into a conditional generative adversarial network, and output denoised OTDR data; The event location module is used to slice the denoised OTDR data and its variance and differential curves, with the slice length ranging from one pulse width to several pulse widths; the sliced ​​OTDR data, the variance of the OTDR data, and the differential curve of the OTDR data are respectively input into three feature extraction modules; the feature extraction modules are a series of convolutional layers, batch normalization layers, and activation layers; After merging the features output by the three feature extraction modules, the features are input into the efficient channel attention layer, which outputs an initial probability sequence of 0-1 with the same length as the input, representing the probability of an event at each point. The initial probability sequence is binarized using the threshold x, with the values ​​greater than x being 1 and the values ​​less than x being 0, to obtain a binary sequence L. In the binary sequence L, a point with a value of 1 represents an event.

9. An event location device for OTDR data, comprising: at least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the event location method of OTDR data described in any one of claims 1 to 7.

10. A computer storage medium having executable instructions stored thereon, wherein when the instructions are executed by a processor, the processor is enabled to implement the method according to any one of claims 1 to 7.

Images