Distributed optical fiber sensing pipeline leakage identification method

Abstract

The invention belongs to the technical field of distributed optical fiber sensing, and discloses a distributed optical fiber sensing pipeline leakage identification method. Comprising the following steps that 1, a distributed optical fiber sensor is adopted to collect vibration information of water supply pipeline leakage to serve as an original data set, and original data is a one-dimensional time sequence signal; 2, preprocessing the original data, and converting the preprocessed signal into a spectrogram; step 3, an improved MobileNetV3-large network model is constructed, and the step 3 is that the improved MobileNetV3-large network model is And step 4, performing leakage aperture identification by using the trained network model. According to the method, the accuracy of distributed optical fiber sensing water supply pipeline leakage identification can be improved while the calculated amount is small, and the identification efficiency is improved.

Description

Technical Field

[0001] The invention belongs to the technical field of distributed optical fiber sensing and relates to a distributed optical fiber sensing pipeline leakage identification method. Background Art

[0002] Urban water supply pipelines are an important part of urban infrastructure. Water supply pipeline transportation has become the most widely used mode of water transportation due to its advantages such as large transportation volume, strong continuity, convenient maintenance and low cost. However, water supply pipelines may be damaged due to factors such as aging corrosion, improper construction or external pressure, causing huge economic losses. Therefore, it is of great significance to study and improve the real-time, efficiency and accuracy of water supply pipeline leakage monitoring. Distributed Acoustic Sensing (DAS) monitoring water supply pipeline is a water supply pipeline monitoring technology based on optical fiber technology. It lays optical fiber wrapped around the water supply pipeline and uses the sensitivity of optical fiber to vibration signals to monitor the leakage signal of the water supply pipeline in real time. When a water supply pipeline leaks, the frequency of the vibration generated is different from the frequency when there is no leakage, so it can distinguish whether the water supply pipeline is leaking. Compared with traditional water supply pipeline leakage monitoring methods, distributed optical fiber acoustic sensors have the advantages of real-time monitoring, high precision and full coverage.

[0003] With the rapid development of deep learning technology, its application in distributed fiber optic sensors is becoming more and more extensive. How to quickly and accurately identify different signals is of great research significance. Traditional deep learning methods process data and recognize signals for a long time, and often require a lot of computing resources. Lightweight neural networks have fast processing speeds and can be applied to resource-constrained mobile or edge devices, but they are not as accurate as traditional neural networks. The lightweight neural network MobileNetV3 focuses on the relationship between channels, but may ignore local information in the spatial dimension, thus affecting the performance of certain tasks. Although the nonlinear characteristics of H-Swish contribute to the expressiveness of the model, it may not provide sufficient nonlinearity in some cases, thus affecting the model's ability to learn complex patterns.

[0004] In view of the above problems, the present invention proposes a distributed optical fiber sensing pipeline leakage identification method. Summary of the invention

[0005] The present invention solves the problems of low accuracy and large calculation amount in the prior art of distributed optical fiber sensing water supply pipeline leakage identification by providing a distributed optical fiber sensing pipeline leakage identification method.

[0006] The objective of the present invention is achieved through the following technical solution: A distributed optical fiber sensing pipeline leakage identification method comprises the following steps:

[0007] Step 1: Use distributed optical fiber sensors to collect vibration information of water supply pipeline leakage as the original data set, and the original data is a one-dimensional time series signal;

[0008] Step 2, preprocessing the raw data and converting the preprocessed signal into a spectrum diagram;

[0009] Step 3, build an improved MobileNetV3-large network model;

[0010] The improved MobileNetV3-large network model includes: a convolutional layer, a linear bottleneck layer, a pooling layer, and a fully connected layer; the first layer of the basic architecture of the network model is a convolutional layer, which is used to extract low-level features of the image; the second to sixteenth layers of the basic architecture of the network model are linear bottleneck layers, which are used to reduce the amount of calculation and continuously extract features. When the step size of the linear bottleneck layer is 2, the CBAM attention mechanism is adopted. The seventeenth layer of the basic architecture of the network model is a convolution layer, which is used for channel integration to reduce computational complexity; the eighteenth layer of the basic architecture of the network model is a pooling layer, which is used for global pooling to generate a feature map of a fixed size; the nineteenth and twentieth layers of the basic architecture of the network model are fully connected layers, which replace the fully connected layers with two 1×1 convolution layers to output classification results; the activation function of the first six linear bottleneck layers is the ReLu activation function, the pooling layer and the last convolution layer do not use activation functions, and the rest use the Mish activation function; after completing the training of the ResNet50 teacher model with the data set, the improved MobileNetV3-large network model is trained using the knowledge distillation strategy to obtain the trained improved MobileNetV3-large network model;

[0011] Step 4: Use the trained network model to identify the leakage aperture.

[0012] Preferably, step 1 is specifically as follows: the fiber optic distributed acoustic sensor is placed close to the surface of the water supply pipe, the fiber optic distributed acoustic sensor sends the monitored vibration signal to the DAS demodulator in real time, so that the DAS demodulator obtains the vibration information of the water supply pipe in real time, and the DAS demodulator transmits the vibration information back to the local server in real time. When sampling, the sampling duration of the sample data and the sampling frequency of the DAS demodulator are set, and the number of data points of each sample data = sampling duration * sampling frequency.

[0013] Preferably, the preprocessing in step 2 includes: spectral subtraction denoising and sliding window division; the spectral subtraction denoising includes using a smoothing mechanism to subtract the data of each sample in the original data set to denoise; the sliding window division includes using a sliding window with a preset step size to perform sliding division on a number of sample points obtained after standardization processing, and performing normalization processing after obtaining the time series samples, and scaling the time series to the range of [-1,1].

[0014] The spectral subtraction is calculated according to the following formula:

[0015] D(w)=P s (w)-αP n (w)

[0016]

[0017] Among them, the P S (w) is the preset estimated environmental noise spectrum, P n (w) is the spectrum of the input signal, that is, the spectrum of the mixed sound signal, D(w) is the spectrum of the noise-removed signal, P′ S (W) is the spectrum of the output signal, that is, the spectrum of the mixed sound signal after removing the environmental noise, α≥0, α is the subtraction factor, 0<β<<1, β is the spectrum lower limit parameter.

[0018] Preferably, the CBAM attention mechanism of step 3 is specifically:

[0019] When the stride is 1, the convolution layer consists of a depthwise convolution and two pointwise convolutions. When the stride is 2, a CBAM attention mechanism is added after the depthwise convolution, and the convolution layer module uses a residual connection.

[0020] The CBAM attention mechanism consists of a spatial attention module and a channel attention module;

[0021] The channel attention module includes a parallel maximum pooling layer. First, the input features are processed by the channel attention mechanism, including global maximum pooling and global average pooling, which are performed in parallel in the same feature layer. Then, the pooling results are processed using a shared fully connected layer to obtain a feature map. This feature map is then input into a two-layer neural network to generate two features. After adding these two features, the channel attention weight is obtained through the Sigmoid activation function, and the product is obtained with the input feature map to obtain the input features required by the spatial attention module;

[0022] Among them, the calculation formula of the channel attention module is:

[0023]

[0024] In the above formula, F represents the input feature map, W 0 and W 1 Represents the weights of channels 0 and 1 of the input feature layer, represents the maximum pooling eigenvalue, represents the average pooling feature value, σ is the Sigmoid activation function;

[0025] In order to reduce the loss of image position information, the attention module is used to process the input feature map. First, the feature map F obtained by channel attention is ′ It is used as the input feature map of the current module, and then it undergoes global maximum pooling and global average pooling respectively to obtain two H×W×1 (H and W represent the height and width of the feature map, respectively) feature maps, and then performs channel splicing operation. After that, it is convolved to reduce the dimension to a H×W×1, and then the Sigmoid function is used to generate the weight value, and the dot multiplication is used to obtain the final feature map;

[0026] The calculation formula of the spatial attention module is as follows:

[0027]

[0028] F ′ represents the input feature map, f 7×7 Indicates that the convolution kernel size is 7×7, M S (F ′ ) represents the spatial attention weight, σ is the Sigmoid activation function, represents the maximum pooling eigenvalue, Represents the average pooled feature value.

[0029] Preferably, the Mish activation function formula in step 3 is:

[0030] f(x)=xtanh(ln(1+e x ))

[0031] x represents each element of the input feature map, f(x) represents each element of the output feature map, and tanh is the hyperbolic tangent function.

[0032] Preferably, the knowledge distillation strategy in step 3 extracts knowledge from ResNet50, specifically:

[0033] Use ResNet50 as the teacher model and the improved MobileNetV3-large as the student model;

[0034] First, the spectrum graph is input into the teacher network model, and then the knowledge distillation strategy is used to distill the knowledge of the student network model to extract the knowledge from ResNet50 into the improved lightweight MobileNetV3-large network.

[0035] The steps of knowledge distillation are as follows:

[0036] The trained ResNet50 model is used as the teacher model. The teacher model processes the spectrogram image dataset and generates the predicted probability of each category of the leakage aperture as a soft label. The predicted probability is obtained by adjusting the temperature parameter T of the softmax function in the deep learning model:

[0037]

[0038] Among them, softmax(T(z i )) is the predicted probability of the teacher model for the leakage aperture category, i.e., the soft label; T(z i ) is the logits output of the teacher model for category i, that is, the unnormalized model output; Te is the temperature parameter, which is used to control the smoothness of the output probability distribution. When Te=1, the temperature-adjusted softmax function degenerates into a regular softmax function. When Te>1, the output probability distribution becomes smooth;

[0039] Soften the output of the student model through the softmax function of the same temperature Te:

[0040]

[0041] Among them, S(z i ) is the logits output of the student model for category i.

[0042] Preferably, the water-tightness and water-tightness of different apertures in step 4 are specifically:

[0043] The apertures for collecting vibration signals of water supply pipeline leakage are watertight and three different apertures for water leakage. 90% of the data set is used as the training set, and 10% of the data set is used as the validation set.

[0044] The beneficial effects of the present invention are:

[0045] The present invention discloses a DAS water supply pipeline leakage signal recognition method of an improved MobileNetV3-large. The present invention replaces the SENet attention mechanism in the original MobileNetV3-large network with a CBAM attention mechanism, considers the relationship and importance between different channels, and also considers the relationship between different spatial positions in the feature map, so as to better capture the correlation and context information in the feature map, and better enhance the spatial feature perception; the Mish activation function in the MobileNetV3-large network of the present invention replaces the H-Swish, fully utilizes the extracted deep image information, and improves the recognition accuracy of the model; the ResNet50 teacher model is trained by a data set, and after the teacher model training is completed, the improved MobileNetV3-large student model is trained by using a knowledge distillation strategy to obtain a trained student model, so as to significantly improve the recognition accuracy while keeping the model lightweight when deployed on a mobile or edge device. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0047] in:

[0048] Figure 1 An overall flow chart of a distributed optical fiber sensing water supply pipeline leakage signal recognition method based on an improved MobileNetV3-large provided by the present invention;

[0049] Figure 2 A schematic diagram of an improved MobileNetV3-large network structure provided by the present invention;

[0050] Figure 3 The overall structure diagram of an improved MobileNetV3-large network provided by the present invention; wherein, Figure 3 Where mish is the Mish activation function. Figure 3 Where relu is the ReLu activation function;

[0051] Figure 4 A schematic diagram of the structure of an improved MobileNetV3-large network with a convolutional layer step size of 1,2 provided by the present invention;

[0052] Figure 5 A schematic diagram of the structure of a CBAM attention mechanism provided by the present invention;

[0053] Figure 6 A schematic diagram of a knowledge distillation structure provided by the present invention. Specific implementation plan

[0054] The following will be combined with the embodiments of the present invention and the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] Example 1

[0056] See attached Figure 1 As shown, a flow chart of a distributed optical fiber sensing pipeline leakage identification method provided by an embodiment of the present invention, the method comprises the following steps:

[0057] Step 1: Use distributed optical fiber sensors to collect vibration information of water supply pipeline leakage as the original data set, and the original data is a one-dimensional time series signal;

[0058] It should be noted that the Distributed Acoustic-vibration Senor (DAS) is a sensing instrument based on the phase-sensitive Rayleigh scattering principle, which has the advantages of anti-electromagnetic interference, high sensitivity, small size, stable performance and long service life.

[0059] It should be noted that the collected data sets are data sets of different leakage apertures.

[0060] Step 2: preprocess the raw data and convert the preprocessed signal into a spectrum diagram.

[0061] It should be noted that the preprocessing includes: spectral subtraction denoising and sliding window division; the spectral subtraction denoising includes using a smoothing mechanism to subtract the data of each sample in the original data set to denoise; the sliding window division includes using a sliding window with a preset step size to perform sliding division on a number of sample points obtained after standardization, and performing normalization after obtaining the time series samples to scale the time series to the range of [-1,1].

[0062] Specifically, spectral subtraction is a method that utilizes the short-time stationary characteristics of speech signals to subtract the short-time spectrum of noise from the short-time spectrum of noisy speech signals, thereby obtaining the spectrum of pure speech. Spectral subtraction is calculated according to the following formula:

[0063] D(w)=P s (w)-αP n (w)

[0064]

[0065] Among them, the P s (w) is the preset estimated environmental noise spectrum, P n (w) is the spectrum of the input signal, that is, the spectrum of the mixed sound signal, D(w) is the spectrum of the noise-removed signal, P S ′ (W) is the spectrum of the output signal, that is, the spectrum of the mixed sound signal after removing the environmental noise, α≥0, α is the subtraction factor, 0<β<<1, β is the spectrum lower limit parameter.

[0066] Step 3, such as Figure 2-Figure 3 As shown, an improved MobileNetV3-large network model is constructed.

[0067] It should be noted that the network layers in the improved MobilenetV3-large network model are specifically the first layer is a convolutional layer, with an input size of (224×224×3) and an output size of (112×112×16), which is used to extract low-level features of the image, such as edges and textures. The second layer is a bottleneck layer (bneck, 3x3), with an input size of (112×112×16) and an output size of (112×112×16), which is used to extract different features of the image. The third layer is a bottleneck layer (bneck, 3x3), with an input size of (112×112×16), and the output size remains unchanged at (112×112×16), and further extracts features. The fourth layer is a bottleneck layer (bneck, 3x3), with an input size of (112×112×16) and an output size of (56×56×24), which increases the number of channels to capture more features, and enhances important features through the CBAM module. The fifth layer is the bottleneck layer (bneck, 3x3), with an input size of (56×56×24) and an output size of (56×56×24), which further extracts features. The sixth layer is the bottleneck layer (bneck, 5x5), with an input size of (56×56×24) and an output size of (28×28×40), using a larger convolution kernel to capture more complex features. The seventh layer is the bottleneck layer (bneck, 5x5), with an input size of (28×28×40) and an output size of (28×28×40), which continues to extract high-level features. The eighth layer is the bottleneck layer (bneck, 5x5), with an input size of (28×28×40) and an output size of (28×28×40), which repeatedly extracts high-level features to increase the depth of the model. The ninth layer is the bottleneck layer (bneck, 3x3), with an input size of (28×28×40) and an output size of (14×14×80), further increasing the number of channels and using the attention mechanism to enhance feature representation. The tenth layer is the bottleneck layer (bneck, 3x3), with an input size of (14×14×80) and an output size that remains unchanged at (14×14×80), which deepens the network structure. The eleventh layer is the bottleneck layer (bneck, 3x3), with an input size of (14×14×80) and an output size that remains unchanged at (14×14×80), further extracting features. The twelfth layer is the bottleneck layer (bneck, 3x3), with an input size of (14×14×80) and an output size that remains unchanged at (14×14×80), deepening the network to capture more complex features. The thirteenth layer is the bottleneck layer (bneck, 3x3), with an input size of (14×14×80) and an output size of (14×14×112). The number of channels is increased to enhance the model's expressiveness. The fourteenth layer is the bottleneck layer (bneck, 3x3), with an input size of (14×14×112) and an output size that remains unchanged at (14×14×112). Its function is to deepen feature extraction.The fifteenth layer is the bottleneck layer (bneck, 5x5), with an input size of (14×14×112) and an output size of (14×14×160). It uses a larger convolution kernel to capture more complex high-level features. The sixteenth layer is the bottleneck layer (bneck, 5x5), with an input size of (14×14×160) and an output size that remains unchanged at (14×14×160). Its function is to increase the depth of the network. The seventeenth layer is the bottleneck layer (bneck, 5x5), with an input size of (14×14×160) and an output size of (7×7×960), which greatly increases the number of channels and reduces the size of the feature map. The eighteenth layer is the convolution layer (conv2d, 1x1), with an input size of (7×7×960) and an output size that remains unchanged at (7×7×960). Its function is to integrate channels and reduce computational complexity. The 19th layer is a pooling layer (pool, 7x7), with an input size of (7×7×960) and an output size of (1×1×960). It performs global pooling and generates a fixed-size feature map. The 20th layer is a convolutional layer (conv2d, 1x1, NBN), with an input size of (1×1×960) and an output size of (1×1×1280), which further increases the dimension of the output features. The 21st layer is a convolutional layer (conv2d, 1x1, NBN), with an input size of (1×1×1280) and an output size of (1×1×k), which is used to generate classification or regression results.

[0068] Furthermore, after completing the training of the ResNet50 teacher model with the data set, the improved MobileNetV3-large network model is trained using the knowledge distillation strategy to obtain the trained improved MobileNetV3-large network model;

[0069] It should be noted that the CBAM attention mechanism in step 3 is specifically:

[0070] When the stride is 1, the convolution layer consists of a depthwise convolution and two pointwise convolutions. When the stride is 2, a CBAM attention mechanism is added after the depthwise convolution, and the convolution layer module uses a residual connection.

[0071] The CBAM attention mechanism consists of a spatial attention module and a channel attention module, such as Figure 5 As shown;

[0072] The channel attention module includes a parallel maximum pooling layer. First, the input features are processed by the channel attention mechanism, including global maximum pooling and global average pooling, which are performed in parallel in the same feature layer. Then, the pooling results are processed using a shared fully connected layer to obtain a feature map. This feature map is then input into a two-layer neural network to generate two features. After adding these two features, the channel attention weight is obtained through the Sigmoid activation function, and the product is obtained with the input feature map to obtain the input features required by the spatial attention module;

[0073] Among them, the calculation formula of the channel attention module is:

[0074]

[0075] In the above formula, F represents the input feature map, W 0 and W 1 Represents the weights of channels 0 and 1 of the input feature layer, represents the maximum pooling eigenvalue, represents the average pooling eigenvalue, and σ is the Sigmoid activation function.

[0076] In order to reduce the loss of image position information, the attention module is used to process the input feature map. First, the feature map F obtained by channel attention is ′ It is used as the input feature map of the current module, and then undergoes global maximum pooling and global average pooling respectively to obtain two H×W×1 (H and W represent the height and width of the feature map respectively) feature maps, and then performs channel splicing operation, and then convolution is used to reduce the dimension to a H×W×1, and then the Sigmoid function is used to generate the weight value, and the dot product is used to get the final feature map.

[0077] The calculation formula of the spatial attention module is as follows:

[0078]

[0079] F ′ represents the input feature map, f 7×7 Indicates that the convolution kernel size is 7×7, M S (F ′ ) represents the spatial attention weight, σ is the Sigmoid activation function, represents the maximum pooling eigenvalue, Represents the average pooled feature value.

[0080] It should be noted that the input of the spatial attention part of CBAM is the output feature of the channel attention module. Then the channel maximum pooling and average pooling are performed to obtain the 1×f×t spatial feature, and then the two descriptions are spliced ​​together by channel, and then a 7×7 convolution layer and activation function are passed to the spatial weight coefficient, and finally the spatial weight coefficient is multiplied by the feature, so that the new local details are richer and the frame-level attention weight is obtained.

[0081] It should be noted that the Mish activation function in step 3 replaces H-Swish and includes: Mish activation function is a self-regularized, non-monotonic activation function. Mish uses the Self-Gating feature, in which the non-modulated input is multiplied by the output obtained by the input through a nonlinear function. Since a small amount of negative information is retained, Mish eliminates the Dying ReLU phenomenon. This feature contributes to better expressiveness and information flow. Since there is no upper bound, Mish avoids saturation.

[0082] Furthermore, the Mish activation function formula is:

[0083] f(x)=xtanh(ln(1+e x ))

[0084] x represents each element of the input feature map, f(x) represents each element of the output feature map, and tanh is the hyperbolic tangent function.

[0085] It should be noted that in step 3, the knowledge distillation strategy is used to train the improved MobileNetV3-large student model, and the trained student model is specifically ResNet50 as the teacher model, and the improved MobileNetV3-large is used as the student model.

[0086] Specifically, Figure 6 As shown in the figure, the spectrum graph is first input into the teacher network model, and then the knowledge distillation strategy is used to perform knowledge distillation on the student network model to extract knowledge from ResNet50 into the improved lightweight MobileNetV3-large network.

[0087] The trained ResNet50 model is used as the teacher model. The teacher model processes the spectrogram image dataset and generates the predicted probability of each category of the leakage aperture as a soft label. The predicted probability is obtained by adjusting the temperature parameter T of the softmax function in the deep learning model:

[0088]

[0089] Among them, softmax(T(z i)) is the predicted probability of the teacher model for the leakage aperture category, i.e., the soft label; T(z i ) is the logits output of the teacher model for category i, that is, the unnormalized model output; Te is the temperature parameter, which is used to control the smoothness of the output probability distribution. When Te=1, the temperature-adjusted softmax function degenerates into a regular softmax function. When Te>1, the output probability distribution becomes smooth.

[0090] Soften the output of the student model through the softmax function of the same temperature Te:

[0091]

[0092] Among them, S(z i ) is the logits output of the student model for category i; when training the student model, the cross entropy loss adjusted by temperature T is used to determine the degree of student model training:

[0093] L KD = -∑softmax(T(z i ))·log(softmax(S(z i )))

[0094] The student model is an improved lightweight MobileNetV3-large network. The output of the student model is combined with the cross entropy loss of the corresponding labels of the real samples to enhance the learning effect:

[0095] L CE = -∑y·log(softmax(S(z i )))

[0096] Among them, y is the one-hot encoding of the label corresponding to the real sample.

[0097] By minimizing the weighted sum of the distillation loss and the label loss corresponding to the real sample, the training objective of the student model is:

[0098] L=αL KD +(1-α)L CE

[0099] Among them, α is the weight factor that balances the distillation loss and the label loss corresponding to the real sample.

[0100] Furthermore, knowledge distillation training is performed on ResNet50 to obtain a student network with better performance and higher lightness, thereby compressing the size of the target detection model to improve the target detection performance.

[0101] The improved MobileNetV3-large network convolutional layer provided in the embodiment of the present invention has a structure with a step size of 1,2, such as Figure 4 As shown, the convolution layer specifically includes the following structure:

[0102] When the stride is 1, the convolution layer consists of a depthwise convolution and two pointwise convolutions. When the stride is 2, a CBAM attention mechanism is added after the depthwise convolution, and the convolution layer module uses a residual connection.

[0103] Specifically, pointwise convolution uses a 1x1 convolution kernel for convolution operations. It can be seen as a linear combination of channels at each pixel.

[0104] Point-by-point convolution can change the number of channels of the convolutional layer, thereby achieving mixing between channels and helping the network learn more complex feature representations by weighting the features of different channels.

[0105] Since the convolution kernel size is 1x1 and the amount of computation is relatively small, it is often used in deep learning models to reduce the number of parameters.

[0106] Furthermore, depthwise convolution applies a convolution kernel to each input channel independently, usually using a 3x3 or 5x5 convolution kernel. Each input channel has a corresponding convolution kernel.

[0107] Depthwise convolution performs convolution on each channel separately, which helps to extract local features.

[0108] Compared to traditional convolution, depthwise convolution significantly reduces the number of parameters because only one convolution kernel is used per channel.

[0109] Deep convolution has lower computational complexity than standard convolution and is suitable for resource-constrained environments such as mobile devices.

[0110] Step 4: Use the trained network model to identify the leakage aperture.

[0111] It should be noted that the aperture for leakage aperture identification can be 0 mm, 1 mm, 2 mm and 3 mm.

[0112] Model evaluation:

[0113] Since the probability of a water supply pipeline leakage event is lower than that of a non-leakage event, and the inventors prefer to identify water supply pipeline leakage events, the recall rate is selected as the evaluation indicator. The recall rate formula is: The accuracy is The data set of the present invention has 20,000 samples, of which 7,450 are watertight, 4,170 are 1mm in diameter, 4,220 are 2mm in diameter, and 4,160 are 3mm in diameter. 90% of the data set is used as a training set, and 10% of the data set is used as a validation set. The recall rate of watertightness can reach 98.5%, and the accuracy rate is 99.2%. The recall rate of 1mm in diameter can reach 98.2%, and the accuracy rate is 98.7%. The recall rate of 2mm in diameter can reach 98.4%, and the accuracy rate is 98.9%. The recall rate of 3mm in diameter can reach 98.3%, and the accuracy rate is 98.8%.

[0114] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0115] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A distributed optical fiber sensing pipeline leakage identification method, It is characterized in that The following steps are involved: Step 1: Use distributed optical fiber sensors to collect vibration information of water supply pipeline leakage as the original data set, and the original data is a one-dimensional time series signal; Step 2, preprocessing the raw data and converting the preprocessed signal into a spectrum diagram; Step 3, build an improved MobileNetV3-large network model; The improved MobileNetV3-large network model includes: a convolutional layer, a linear bottleneck layer, a pooling layer and a fully connected layer; the first layer of the basic architecture of the network model is a convolutional layer, which is used to extract low-level features of the image; the second to sixteenth layers of the basic architecture of the network model are linear bottleneck layers, which are used to reduce the amount of calculation and continuously extract features. When the step size of the linear bottleneck layer is 2, the CBAM attention mechanism is adopted; the seventeenth layer of the basic architecture of the network model is a convolutional layer, which is used for channel integration to reduce computational complexity; the eighteenth layer of the basic architecture of the network model is a pooling layer, which is used Global pooling is performed to generate a feature map of a fixed size; the nineteenth and twentieth layers of the basic architecture of the network model are fully connected layers, and the fully connected layers are replaced by two 1×1 convolutional layers to output classification results; the activation function of the first six linear bottleneck layers is the ReLu activation function, the pooling layer and the last convolutional layer do not use activation functions, and the rest use the Mish activation function; after completing the training of the ResNet50 teacher model with the data set, the improved MobileNetV3-large network model is trained using the knowledge distillation strategy to obtain the trained improved MobileNetV3-large network model; Step 4: Use the trained network model to identify the leakage aperture.

2. The distributed optical fiber sensing pipeline leakage identification method according to claim 1, It is characterized in that The step 1 is specifically as follows: the fiber optic distributed acoustic sensor is placed close to the surface of the water supply pipe, and the fiber optic distributed acoustic sensor sends the monitored vibration signal to the DAS demodulator in real time, so that the DAS demodulator obtains the vibration information of the water supply pipe in real time, and the DAS demodulator sends the vibration information back to the local server in real time. When sampling, the sampling time of the sample data and the sampling frequency of the DAS demodulator are set, and the number of data points of each sample data = sampling time * sampling frequency.

3. The distributed optical fiber sensing pipeline leakage identification method according to claim 1, It is characterized in that The preprocessing in step 2 includes: spectral subtraction denoising and sliding window division; the spectral subtraction denoising includes using a smoothing mechanism to subtract the data of each sample in the original data set to denoise; the sliding window division includes using a sliding window with a preset step size to perform sliding division on a number of sample points obtained after standardization processing, and performing normalization processing after obtaining the time series samples, and scaling the time series to the range of [-1,1].

4. The distributed optical fiber sensing pipeline leakage identification method according to claim 3, It is characterized in that The spectral subtraction is calculated according to the following formula: D(w)=P s (w)-αP n (w) Among them, P s (w) is the preset estimated environmental noise spectrum, P n (w) is the spectrum of the input signal, that is, the spectrum of the mixed sound signal, D(w) is the spectrum of the noise-removed signal, P′ S (w) is the spectrum of the output signal, that is, the spectrum of the mixed sound signal after removing the environmental noise, α≥0, α is the subtraction factor, 0<β<<1, β is the spectrum lower limit parameter.

5. The distributed optical fiber sensing pipeline leakage identification method according to claim 1, It is characterized in that The CBAM attention mechanism of step 3 is specifically: When the stride is 1, the convolution layer consists of a depthwise convolution and two pointwise convolutions; when the stride is 2, a CBAM attention mechanism is added after the depthwise convolution, and the convolution layer module uses a residual connection.

6. The distributed optical fiber sensing pipeline leakage identification method according to claim 5, It is characterized in that The CBAM attention mechanism consists of a spatial attention module and a channel attention module; The channel attention module includes a parallel maximum pooling layer; first, the input features are processed by the channel attention mechanism, including global maximum pooling and global average pooling, which are performed in parallel in the same feature layer; then, the pooling results are processed using a shared fully connected layer to obtain a feature map; then, this feature map is input into a two-layer neural network to generate two features; after adding these two features, the channel attention weight is obtained by the Sigmoid activation function, and the channel attention weight is multiplied with the input feature map to obtain the input features required by the spatial attention module; Among them, the calculation formula of the channel attention module is: M in the above formula c (F) represents the channel attention weight, F represents the input feature map, W 0 and W 1 Represents the weights of channels 0 and 1 of the input feature layer, represents the maximum pooling eigenvalue, represents the average pooling feature value, σ is the Sigmoid activation function; In order to reduce the loss of image position information, the attention module is used to process the input feature map. First, the feature map F′ obtained by channel attention is used as the input feature map of the current module. Then, it is subjected to global maximum pooling and global average pooling respectively to obtain two H×w×1 feature maps, and then the channel splicing operation is performed. H and w represent the height and width of the feature map respectively; then, it is convolved to reduce the dimension to H×w×1, and then the Sigmoid function is used to generate the weight value, and the dot multiplication is used to obtain the final feature map; The calculation formula of the spatial attention module is as follows: F′ represents the input feature map, f 7×7 Indicates that the convolution kernel size is 7×7, M S (F′) represents the spatial attention weight, σ is the Sigmoid activation function, represents the maximum pooling eigenvalue, Represents the average pooled feature value.

7. The distributed optical fiber sensing pipeline leakage identification method according to claim 1, It is characterized in that The Mish activation function formula in step 3 is: f(x)=xtanh(ln(1+e x )) x represents each element of the input feature map, f(x) represents each element of the output feature map, and tanh is the hyperbolic tangent function.

8. The distributed optical fiber sensing pipeline leakage identification method according to claim 1, It is characterized in that The knowledge distillation strategy in step 3 extracts knowledge from ResNet50, specifically: Use ResNet50 as the teacher model and the improved MobileNetV3-large as the student model; First, the spectrum graph is input into the teacher network model, and then the knowledge distillation strategy is used to distill the knowledge of the student network model to extract the knowledge from ResNet50 into the improved lightweight MobileNetV3-large network.

9. The distributed optical fiber sensing pipeline leakage identification method according to claim 8, It is characterized in that The steps of knowledge distillation are as follows: The trained ResNet50 model is used as the teacher model. The teacher model processes the spectrogram image dataset and generates the predicted probability of each category of the leakage aperture as a soft label. The predicted probability is obtained by adjusting the temperature parameter T of the softmax function in the deep learning model: Among them, softmax(T(z i )) is the predicted probability of the teacher model for the leakage aperture category, i.e., the soft label; T(z i ) is the logits output of the teacher model for category i, that is, the unnormalized model output; Te is the temperature parameter, which is used to control the smoothness of the output probability distribution. When Te=1, the temperature-adjusted softmax function degenerates into a regular softmax function. When Te>1, the output probability distribution becomes smooth; Soften the output of the student model through the softmax function of the same temperature Te: Among them, S(z i ) is the logits output of the student model for category i.

10. The distributed optical fiber sensing pipeline leakage identification method according to claim 1, It is characterized in that The specific examples of no water leakage and water leakage with different apertures in step 4 are: The apertures for collecting vibration signals of water supply pipeline leakage are no leakage and three different apertures with leakage. 90% of the data set is used as the training set and 10% of the data set is used as the validation set.

Images