A data enhancement method and application for optical fiber sensing foreign object intrusion identification

Abstract

The application discloses a data enhancement method and application for optical fiber sensing foreign matter invasion identification, belongs to the technical field of foreign matter invasion monitoring of optical fiber sensing, and finds a signal with maximum energy in foreign matter invasion data by adopting a short time window to slide on the foreign matter invasion data, so as to extract a key waveform in the foreign matter invasion data, and constructs a data enhancement sample containing essential features of the foreign matter invasion data by moving the position of the key waveform and controlling the breakage rate of the key waveform; the application considers that optical fiber sensing data under the same type of foreign matter invasion may have multiple forms of expression, expands the form of expression of the data by exploring the essential features of the foreign matter invasion data, realizes effective expansion of the foreign matter invasion data, enables the model to learn information directly related to the category of the invasion event, and greatly improves the accuracy of foreign matter invasion identification.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of optical fiber sensing foreign object intrusion monitoring, and more particularly to a data enhancement method for optical fiber sensing foreign object intrusion identification and application. BACKGROUND

[0002] Optical fiber sensing technology has become a mainstream technology for intrusion behavior safety monitoring due to its low cost, electromagnetic interference resistance, good harsh environment resistance, high detection sensitivity of demodulation system, and dynamic detection characteristics, and can complete large-scale, high-sensitivity real-time monitoring in complex environments and is widely used in pipeline monitoring, perimeter security, subway tunnels and other important scenes. In recent years, with the development of deep learning technology, optical fiber sensing foreign object intrusion identification methods based on deep learning have been applied to optical fiber sensing foreign object intrusion identification. However, the optical fiber sensing data under the same type of foreign object intrusion may have multiple forms of representation, and it is often difficult to collect complete foreign object intrusion data in the actual collection process, which makes the existing optical fiber sensing foreign object intrusion identification method have low accuracy and poor generalization ability, and is prone to a large number of false positives and false negatives. Therefore, there is an urgent need for a data enhancement method for optical fiber sensing foreign object intrusion identification to effectively expand the collected foreign object intrusion data.

[0003] An existing data enhancement method for optical fiber sensing foreign object intrusion identification generates new foreign object intrusion data using a GAN network to expand the foreign object intrusion data. However, this method usually generates new data based on statistical characteristics such as mean and variance, and the generated new foreign object intrusion data often has no direct connection with the representation form of the actually collected foreign object intrusion data, and does not specifically focus on the essential features of the foreign object intrusion data. Therefore, it is difficult to obtain other forms of foreign object intrusion data corresponding to the type of foreign object intrusion, and it is difficult to effectively expand the foreign object intrusion data, resulting in low accuracy of optical fiber sensing foreign object intrusion identification. At the same time, training the GAN network itself requires a large amount of data, which is contrary to the concept of expanding the data set through data enhancement. SUMMARY

[0004] In view of the defects or improvement needs of the prior art, the present application provides a data enhancement method for optical fiber sensing foreign object intrusion identification and application to solve the technical problem that the prior art cannot effectively expand the foreign object intrusion data.

[0005] To achieve the above purpose, in a first aspect, the present application provides a data enhancement method for optical fiber sensing foreign object intrusion identification, comprising: a shift data enhancement operation, specifically comprising:

[0006] The short-time window is slid on the foreign matter intrusion data sample, and a short-time window energy is calculated each time the short-time window is slid;

[0007] a shift parameter t is randomly extracted in the range of [-aW-pos0, L-aW-pos0] s ; wherein, a is a waveform damage parameter, 0

[0008] The foreign matter intrusion data sample is shifted based on the shift parameter t s , and data conforming to a Gaussian distribution is used to fill the null values in the foreign matter intrusion data sample caused by the shift operation;

[0009] When the shift parameter t s is greater than or equal to 0, the foreign matter intrusion data sample is moved left by |t s | time points; when the shift parameter t s is less than 0, the foreign matter intrusion data sample is moved right by |t s | time points.

[0010] Further preferably, the short-time window energy under the jth sliding is:

[0011]

[0012] ; wherein, S is a sliding step; Y is the foreign matter intrusion data sample, and is a one-dimensional vector data; Y(t) is a data value at t time in the foreign matter intrusion data sample.

[0013] Further preferably, 0.5

[0014] Further preferably, the above data enhancement method further comprises: a signal-to-noise ratio control data enhancement operation performed before or after the shift data enhancement operation;

[0015] The signal-to-noise ratio control data enhancement operation comprises: calculating a power value of the foreign matter intrusion data sample; calculating a sample noise of the foreign matter intrusion data sample based on the power value and a preset signal-to-noise ratio, and superimposing the sample noise on the foreign matter intrusion data sample;

[0016] ; wherein, the power value of the foreign matter intrusion data sample Y is:

[0017] In a second aspect, the present application provides a fiber sensing foreign matter intrusion identification method, comprising:

[0018] inputting the fiber sensing data to be identified into a foreign matter intrusion identification model to obtain an identification result;

[0019] The method for obtaining the foreign matter intrusion recognition model comprises the following steps:

[0020] S1, each foreign matter intrusion data sample in the real-time collected optical fiber sensing data set is subjected to data enhancement by using the data enhancement method provided in the first aspect of the present application, to obtain a data enhancement sample of each foreign matter intrusion data sample, and the data enhancement sample is paired with the corresponding label to form a data enhancement sample set;

[0021] S2, the data enhancement sample set and the optical fiber sensing data set are combined to form a training sample set, and the training sample set is input into a deep learning model for training, to obtain a foreign matter intrusion recognition model.

[0022] Further preferably, the method for obtaining the foreign matter intrusion recognition model further comprises a step S3 performed between the step S1 and the step S2, and the step S3 specifically comprises the following steps:

[0023] all data enhancement samples in the data enhancement sample set are combined two by two, and the two data enhancement samples under each combination and the corresponding labels are subjected to linear weighting summation respectively, to obtain a new sample pair; and all obtained new sample pairs are added to the data enhancement sample set.

[0024] Further preferably, the new sample pair obtained based on the mth group of data enhancement samples (X i ,X l ) is (Z m ,y m );

[0025] wherein Z m =r im ·X i +r lm ·X l ; y m =r im ·y i +r lm ·y l ; r im is a linear weighting factor randomly selected in the range of [0, 1]; r lm =1-r im ; y i is the label corresponding to the data enhancement sample X i ; and y l is the label corresponding to the data enhancement sample X l .

[0026] Further preferably, the deep learning model is a one-dimensional convolutional gated recurrent unit network with an attention mechanism, and comprises three parallel feature extraction modules and a full connection layer connected to the output ends of the feature extraction modules;

[0027] The feature extraction module comprises a first convolutional layer, a second convolutional layer, a first GRU unit and a second GRU unit in cascade; the output of the first convolutional layer is input into the first GRU unit after being spliced with the output of the second convolutional layer through an attention mechanism module;

[0028] In the training process, each input of the deep learning model is a training sample in three consecutive time periods in the training sample set, wherein one training sample is input into a corresponding feature extraction module.

[0029] Further preferably, the convolution kernel size of the first convolutional layer is 1*7; the convolution kernel size of the second convolutional layer is 1*5; and the attention mechanism module is realized by a fully connected layer.

[0030] In a third aspect, the present application provides a fiber sensing foreign object intrusion recognition method, comprising: a memory and a processor, the memory stores a computer program, and the processor executes the computer program to execute the fiber sensing foreign object intrusion recognition method provided in the second aspect of the present application.

[0031] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0032] 1. The present application provides a data enhancement method for fiber sensing foreign object intrusion recognition, which finds a signal with the maximum energy in the foreign object intrusion data by sliding a short time window on the foreign object intrusion data, thereby extracting the key waveform in the foreign object intrusion data, and constructs enhanced data containing the essential features of the foreign object intrusion data by moving the position of the key waveform and controlling the breakage rate of the key waveform; The present application considers that the fiber sensing data under the same type of foreign object intrusion may have multiple forms, and expands the form of the data by exploring the essential features of the foreign object intrusion data, thereby effectively expanding the foreign object intrusion data, enabling the model to learn information directly related to the intrusion event category, and greatly improving the accuracy of foreign object intrusion recognition.

[0033] 2. The present application provides a fiber sensing foreign object intrusion recognition method, which uses data enhancement to expand the data set, thereby improving the model recognition ability. The method can be easily integrated into existing deep learning models, and will not change due to changes in models and scenes. Only a few super parameters need to be modified according to the scene. At the same time, since the model structure is not changed, the real-time performance of the foreign object intrusion recognition is not affected.

[0034] 3、The optical fiber sensing foreign matter invasion recognition method provided by the application, the data set after the data shift enhancement operation, the signal-to-noise ratio control data enhancement operation and the linear weighting data enhancement operation on the foreign matter invasion data sample in the original optical fiber sensing data set is combined with the original optical fiber sensing data set as a training sample set to train the foreign matter invasion recognition model, so that the model can ignore the influence caused by the low signal-to-noise ratio, the shift and the linear weighting, thereby learning the information directly related to the invasion event category, and the accuracy of foreign matter invasion recognition is greatly improved.

[0035] 4、The optical fiber sensing foreign matter invasion recognition method provided by the application adopts a one-dimensional convolution gate recurrent unit network with an attention mechanism to recognize foreign matter invasion, the network uses the attention mechanism to improve the model fitting capability and pay attention to the waveform information more related to the event category, uses a bidirectional gate recurrent neural network for time sequence signal processing to extract the time sequence contact before and after, and excavates the time evolution information of each foreign matter invasion event, and the recognition accuracy is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 A data enhancement method flow chart for optical fiber sensing foreign matter invasion recognition is provided for the embodiment 1 of the application;

[0037] Figure 2 A one-dimensional convolution gate recurrent unit network schematic diagram is provided for the embodiment 2 of the application. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical scheme and advantages of the application more clear, 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 do not limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.

[0039] Embodiment 1,

[0040] A data enhancement method for optical fiber sensing foreign matter invasion recognition, as shown in Figure 1 The method comprises a shift data enhancement operation, specifically comprising:

[0041] 1) a short-time window is used to slide on the foreign matter invasion data sample, and a short-time window energy is calculated every time the window is slid;

[0042] Specifically, the short-time window energy under the jth sliding is:

[0043]

[0044] Wherein, S is the step length of the short-time frame sliding on the time axis; Y is the foreign matter intrusion data sample, specifically a one-dimensional vector data; Y(t) is the data value at time t in the foreign matter intrusion data sample. W = βL, 0 < β ≤ 1; L is the length of the foreign matter intrusion data sample. Preferably, 0.2 ≤ β ≤ 1; in the embodiment, β = 0.2.

[0045] Specifically, when the maximum short-time window energy corresponds to the j0th sliding, pos0 = (j0-1)*S+1.

[0046] 2) Randomly extract a shift parameter t in the range of [-αW-pos0, L-αW-pos0] s ; wherein, α is the waveform damage parameter, 0 < α ≤ 1; W is the length of the short-time window, W = βL, 0 < β ≤ 1; L is the length of the foreign matter intrusion data sample.

[0047] Preferably, 0.5 ≤ α ≤ 1; in the embodiment, α = 0.5; the foreign matter intrusion data sample under the short-time window corresponding to the maximum short-time window energy includes the key waveform of the foreign matter intrusion data sample, and setting α to 0.5 can make at least half of the key waveform damaged.

[0048] 3) Based on the shift parameter t s , the foreign matter intrusion data sample is subjected to a shift operation, and data conforming to a Gaussian distribution is used to fill the null values in the foreign matter intrusion data sample due to the shift operation;

[0049] Wherein, when the shift parameter t s is greater than or equal to 0, the foreign matter intrusion data sample is moved left by |t s | time points on the time axis; when the shift parameter t s is less than 0, the foreign matter intrusion data sample is moved right by |t s | time points on the time axis. In the embodiment, the null values generated by the shift operation are filled by Gaussian distribution N i (t) ~ N(0, 0.02) to make the length of the foreign matter intrusion data sample after the shift data enhancement operation consistent with that of the original foreign matter intrusion data sample.

[0050] It should be noted that, considering that the one-time action time of the one-dimensional sensing signal is generally short, the application adopts a maximum short-time energy window to capture the key signal, and on this basis, considering that the starting point is random when cutting the frame signal from the data stream, the signal waveform in the maximum short-time energy window is randomly translated to simulate the frame signal actually collected. At the same time, in the random cutting process of the frame signal, there may also be a part of the key waveform not in the frame, so the damage rate of the key waveform is controlled to simulate the frame signal of the key waveform damaged in the actual collection process. The application covers the above cases, realizes effective expansion of the foreign object intrusion data, and improves the recognition accuracy of the model in response to the actual signal.

[0051] Further, in an optional implementation, the data enhancement method further includes a signal-to-noise ratio control data enhancement operation performed before or after the shift data enhancement operation.

[0052] The signal-to-noise ratio control data enhancement operation includes: calculating a power value of the foreign object intrusion data sample; calculating a sample noise of the foreign object intrusion data sample based on the power value and a preset signal-to-noise ratio and superimposing the sample noise on the foreign object intrusion data sample (Y i (t)=Y i (t)+N i (t)); wherein the power value of the foreign object intrusion data sample Y is: The preset signal-to-noise ratio is flexibly set according to the environment, and in this embodiment, the value is 0 dB.

[0053] It should be noted that the introduction of random noise will cause the signal-to-noise ratio of the training data to be inconsistent, and the signal-to-noise ratio is an important standard for measuring signal quality, and it is beneficial to maintain the consistency of the data distribution and the model optimization to maintain the consistent signal-to-noise ratio, therefore, in the data enhancement, different amplitude noises are added to the foreign object intrusion data sample at different times, so that the signal-to-noise ratio of the enhanced data is basically consistent.

[0054] Embodiment 2,

[0055] A method for identifying foreign object intrusion based on optical fiber sensing, comprising:

[0056] inputting the optical fiber sensing data to be identified into a foreign object intrusion identification model to obtain an identification result;

[0057] The method for obtaining the foreign object intrusion identification model includes:

[0058] S1, each foreign object intrusion data sample in the real-time collected optical fiber sensing data set is subjected to the data enhancement method provided in Embodiment 1 of the application to obtain a data enhancement sample of each foreign object intrusion data sample, and is paired with the corresponding label to form a data enhancement sample set;

[0059] In this embodiment, the distributed acoustic sensing system is used to collect the optical fiber sensing data stream in real time. The optical fiber sensing data stream is composed of a series of one-dimensional vector data, including normal optical fiber sensing data samples, foreign object intrusion data samples, and corresponding labels.

[0060] S2, the data enhancement sample set and the optical fiber sensing data set are combined to form a training sample set, which is input into the deep learning model. The deep learning model is trained by minimizing the difference between the predicted label and the true label output by the deep learning model, and an foreign object intrusion identification model is obtained.

[0061] Further, in an optional embodiment, the method for obtaining the foreign object intrusion identification model further comprises a step S3 performed between step S1 and step S2, which specifically comprises:

[0062] All data enhancement samples in the data enhancement sample set are combined two by two, and the two data enhancement samples and the corresponding labels under each combination are respectively linearly weighted and summed to obtain a new sample pair; all the obtained new sample pairs are added to the data enhancement sample set.

[0063] Specifically, the new sample pair obtained based on the mth data enhancement sample (X i ,X l ) is (Z m ,y m );

[0064] Wherein, at the corresponding time point t, Z m (t) = r im ·X i (t) + r lm ·X l (t); y m = r im ·y i +r lm ·y l ; r im is a linear weighting factor randomly selected in the range [0, 1]; r lm = 1-r im ; y i is the label corresponding to the data enhancement sample X i ; y l is the label corresponding to the data enhancement sample X l .

[0065] In an optional embodiment, the representation form of the label is set based on the number of label types, such as 4 label types, including normal state, pile driver abnormal intrusion state, excavator abnormal intrusion state, and manual excavation abnormal intrusion state; wherein the label representing the normal state is denoted as The label indicating the abnormal intrusion state of the pile driver is marked as The label indicating the abnormal intrusion state of the excavator is marked as The label indicating the abnormal intrusion state of the manual excavation is marked as To facilitate linear weighted summation.

[0066] Further, in an optional embodiment, the deep learning model is a one-dimensional convolutional gated recurrent unit network with an attention mechanism, as shown in Figure 2 As shown, three parallel feature extraction modules, and a full connection layer connected to the output end of each feature extraction module;

[0067] The feature extraction module includes a cascaded first convolutional layer, a second convolutional layer, a first GRU unit (gated recurrent unit), and a second GRU unit (gated recurrent unit); the output of the first convolutional layer is input into the first GRU unit after being spliced with the output of the second convolutional layer through an attention mechanism module; wherein the convolution kernel size in the first convolutional layer is 1*7; the convolution kernel size in the second convolutional layer is 1*5; the attention mechanism module is realized by a full connection layer;

[0068] In the training process, each input of the deep learning model is a training sample in three consecutive time periods in the training sample set, wherein a training sample is input into a corresponding feature extraction module.

[0069] In the application process, the optical fiber sensing data to be identified can be used as the optical fiber sensing data at the intermediate time, and the optical fiber sensing data at the previous and subsequent time can be filled with 0 and then input into the one-dimensional convolutional gated recurrent unit network with an attention mechanism for identification.

[0070] It should be noted that the one-dimensional convolutional gated recurrent unit network with an attention mechanism uses the attention mechanism to improve the model fitting capability and pay attention to the waveform information more relevant to the event category; uses the bidirectional gated recurrent neural network for time series signal processing to extract the previous and subsequent time series relationship, and mine the time evolution information of each foreign object intrusion event, thereby further improving the recognition accuracy.

[0071] In summary, the foreign object intrusion recognition model obtained based on the above method can ignore the influence of low signal-to-noise ratio, shift, and linear weighting, and has high recognition accuracy and strong generalization ability.

[0072] Embodiment 3,

[0073] An optical fiber sensing foreign object intrusion recognition method, comprising a memory and a processor, the memory stores a computer program, and the processor executes the computer program to execute the optical fiber sensing foreign object intrusion recognition method provided in embodiment 2 of the application.

[0074] The relevant technical solutions are the same as those of Embodiment 3, and will not be described herein again.

[0075] Those skilled in the art can understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A data augmentation method for foreign object intrusion detection using fiber optic sensing, characterized in that, include: Shift data augmentation operations specifically include: A short time window is used to slide on the foreign object intrusion data sample. The energy of a short time window is calculated for each slide. The starting position pos0 of the short time window when the energy of the short time window is maximum is obtained. A shift parameter t is randomly selected from the range [-αW-pos0, L-αW-pos0]. s ; α is the waveform damage parameter, 0 < α ≤ 1; W is the length of the short time window, W = βL, 0 < β ≤ 1; L is the length of the foreign object intrusion data sample; Based on the shift parameter t s The foreign object intrusion data samples are shifted, and the null values ​​generated by the shifting operation are filled with data that conforms to a Gaussian distribution. Wherein, when the shift parameter t s When the value is greater than or equal to 0, shift the foreign object intrusion data sample to the left by |t. s | A time point; when the shift parameter t s When the value is less than 0, shift the foreign object intrusion data sample to the right by |t. s | point in time.

2. The data augmentation method according to claim 1, characterized in that, The energy of the short-time window under the j-th sliding step is: Where S is the sliding step size; Y is the foreign object intrusion data sample, specifically a one-dimensional vector data; Y(t) is the data value at time t in the foreign object intrusion data sample.

3. The data augmentation method according to claim 1, characterized in that, 0.5≤α≤1；0.2≤β≤1。 4. The data augmentation method according to any one of claims 1-3, characterized in that, Also includes: Signal-to-noise ratio controlled data enhancement operations performed before or after the shift data enhancement operation; The signal-to-noise ratio controlled data enhancement operation includes: calculating the power value of the foreign object intrusion data sample; calculating the sample noise of the foreign object intrusion data sample based on the power value and a preset signal-to-noise ratio, and superimposing the sample noise onto the foreign object intrusion data sample; The power value of the foreign object intrusion data sample Y is:

5. A fiber optic sensing method for foreign object intrusion identification, characterized in that, include: The fiber optic sensing data to be identified is input into the foreign object intrusion identification model to obtain the identification result; The method for obtaining the foreign object intrusion detection model includes: S1. Data augmentation is performed on each foreign object intrusion data sample in the real-time acquired fiber optic sensing dataset using the data augmentation method described in any one of claims 1-4 to obtain data augmentation samples for each foreign object intrusion data sample, and these samples are paired with corresponding tags to form a data augmentation sample set. S2. Combine the data augmentation sample set with the fiber optic sensing dataset to form a training sample set, and input it into the deep learning model for training to obtain the foreign object intrusion identification model.

6. The fiber optic sensing foreign object intrusion identification method according to claim 5, characterized in that, Also includes: Step S3, performed between step S1 and step S2, specifically includes: All data augmented samples in the data augmentation sample set are combined in pairs, and the two data augmented samples and their corresponding labels under each combination are linearly weighted and summed to obtain new sample pairs; all the resulting new sample pairs are added to the data augmentation sample set.

7. The fiber optic sensing foreign object intrusion identification method according to claim 6, characterized in that, Augmented sample based on the m-th data group (X) i ,X l The resulting new sample pair is (Z) m ,y m ); Among them, Z m =r im ·X i +r lm ·X l ;y m =r im ·y i +r lm ·y l ;r im r is a linear weighting factor randomly selected in the range [0,1]; lm =1-r im ;y i Augmenting sample X i The corresponding tag; y l Augmenting sample X l The corresponding tag.

8. The fiber optic sensing foreign object intrusion identification method according to claim 5, characterized in that, The deep learning model is a one-dimensional convolutional gated recurrent unit network with an attention mechanism, including: three parallel feature extraction modules, and a fully connected layer connected to the output of each feature extraction module; The feature extraction module includes a cascaded first convolutional layer, a second convolutional layer, a first GRU unit, and a second GRU unit; the output of the first convolutional layer is concatenated with the output of the second convolutional layer through an attention mechanism module and then input into the first GRU unit; During the training process, each input to the deep learning model is a training sample from three consecutive time periods in the training sample set, wherein one training sample is input into a corresponding feature extraction module.

9. The fiber optic sensing foreign object intrusion identification method according to claim 8, characterized in that, The kernel size of the first convolutional layer is 1*7; the kernel size of the second convolutional layer is 1*5; the attention mechanism module is implemented by a fully connected layer.

10. A fiber optic sensing method for foreign object intrusion identification, characterized in that, include: The system includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the fiber optic sensing foreign object intrusion identification method according to any one of claims 5-9.

Images