Oil and gas pipeline leakage wave identification and monitoring system

Abstract

The present invention relates to the field of pipeline leakage monitoring. Disclosed is an oil and gas pipeline leakage wave identification and monitoring system. In the present invention, an mCNN is combined with LFLBs for performing feature extraction on an acoustic wave signal collected by a DFB, and the collected data improves information completeness; a three-way parallel one-dimensional CNN used in the present invention exhibits good temporal resolution and sensitivity to high-frequency feature transformations in signals; and the present invention integrates advantages of different scales, enabling the algorithm to learn more features, and incorporating the LFLBs to further extract high-level local features. An mCNN-LFLBs network model of the present invention exhibits significant innovation and advancement on the technical level, and also demonstrates extremely high value in actual application. The network model not only provides a novel and efficient technical means for critical fields such as natural gas pipeline inspection, but also introduces new ideas and methods to research fields related to deep learning and signal processing.

Description

A Leakage Wave Identification and Monitoring System for Oil and Gas Pipelines Technical Field

[0001] This invention relates to the field of pipeline leakage monitoring, and in particular to a system for identifying and monitoring leakage waves in oil and gas pipelines. Background Technology

[0002] Pipeline transportation, as the primary mode of oil and gas transmission and distribution, offers advantages such as low cost, high safety, and rapid transport. With the increasing demand for natural gas from industry and residents, pipeline networks are becoming increasingly longer. However, as pipelines age and corrode, they are susceptible to damage from third parties or natural disasters. These situations can lead to pipeline leaks, resulting in economic losses and severe human and environmental impacts. In recent years, various methods for pipeline leak detection have been extensively studied; however, the negative pressure wave method is difficult to detect slow leaks and changes caused by small leaks. Fiber optic detection methods require the addition of appropriate sound insulation materials, are expensive, and have certain limitations.

[0003] Sound waves are not easily attenuated, thus allowing for extremely long propagation distances; they are not easily absorbed by media such as air; they possess extremely strong penetrability; their long wavelengths result in a wider propagation range compared to other sound wave signals; and they are easy to capture, making acoustic methods more suitable for detecting long-distance oil and gas pipelines. Distributed feedback fiber lasers (DFBs), as a type of linear cavity fiber laser, feature single-frequency single-mode operation, low noise, narrow linewidth, and small size, making them suitable for sensing physical quantities such as strain, acceleration, and vibration. Therefore, DFBs are very suitable for capturing acoustic vibration signals. However, due to the extremely high sensitivity of DFBs and the complexity and variability of the objects being probed, there is an urgent need for more effective identification algorithms to effectively discriminate pipeline leakage signals, thereby more accurately identifying whether a pipeline leak exists.

[0004] In recent years, with the rise and development of artificial intelligence technology, machine learning and deep learning have been widely applied in the field of oil and gas pipeline leak detection. Neural networks, with their characteristics of large-scale parallel processing, distributed storage, flexible topology, high redundancy, and nonlinear operation, are more suitable for the complex environments in which pipelines operate. However, continuously increasing the number of network layers to achieve higher recognition accuracy has always been a key issue restricting the application of convolutional neural networks. Most neural network classification algorithms lack high sensitivity to key feature information and are easily interfered with by non-key feature information, leading to inaccurate classification results. Therefore, there is an urgent need to construct more effective neural network models to achieve highly sensitive identification and monitoring of oil and gas pipeline leaks. Summary of the Invention

[0005] To address the issues of high algorithm complexity and low recognition rate in existing pipeline leakage monitoring systems, this invention proposes an oil and gas pipeline leakage wave identification and monitoring system.

[0006] This invention is achieved through the following technical solution: A system for identifying and monitoring leakage waves in oil and gas pipelines, comprising a laser source for generating pump light, the laser source being connected to a wavelength division multiplexer (WDM), the pump light being transmitted to a sensor after being split by the WDM, the sensor being a distributed feedback fiber laser vibration sensor, the distributed feedback fiber laser vibration sensor generating signal light upon receiving the pump light and carrying an acoustic signal caused by pipeline leakage back to the COM terminal of the WDM, the WDM being sequentially connected to an optical isolator and a Michelson interferometer, the acoustic signal being output through the signal terminal of the WDM, passing through the optical isolator and then entering the Michelson interferometer to interfere; the interference signal is then detected by a photodetector and converted from an optical signal into an electrical signal, the photodetector being sequentially connected to a phase generation carrier demodulation system and a computer, the electrical signal being phase demodulated by the phase generation carrier demodulation system, and the original signal of the input model being obtained from the computer; Among them, the network model mCNN-LFLBs, which combines multichannel convolutional neural network (mCNN) with local feature extraction blocks (LFLBs) and hierarchical fusion blocks, includes mCNN, LFLBs composed of several LFLB blocks, and fully connected layers.

[0007] Furthermore, mCNN is composed of three parallel one-dimensional convolutions with different kernel sizes. Except for the kernel size, each convolutional layer contains a batch normalization (BN) layer, an activation function layer, and an average pooling layer. The LFLBs consist of five LFLB blocks. The first four LFLBs are composed of convolutional layers, batch normalization layers, activation function layers, and average pooling layers, while the last LFLB uses global average pooling. The last fully connected layer consists of a dropout layer and a softmax layer. The overall network structure is mainly divided into four parts: input layer, convolutional layer, pooling layer, and output layer.

[0008] Furthermore, the convolution of one channel of mCNN is calculated as follows: Wherein, the data before the convolution operation has a dimension of 1×L, L=5000; K 1 This represents the size of the convolution kernel in one of the three channels, and the sizes of the other two channels are K. 2 K 3 ; The dimension is (LK) 1 +1)×K 1 The matrix; because the scale of the convolution kernel is K. 1 The convolution kernel moves with a stride of 1 each time, and moves a total of (LK) steps. 1 +1) times, so the dimension of the input data during the convolution operation is (LK) times. 1 +1)×K 1 ; Indicates a dimension of K 1 A weight matrix of ×1, The dimension is (LK) 1 +1)×1 paranoia matrix; This indicates that the input data is processed by a convolution kernel of size K. 1 The output matrix obtained by convolution of the channels has dimensions (LK). 1 +1)×1; tanh represents the nonlinear activation function ReLU; yes The output matrix after activation by the activation function has dimensions (LK). 1 +1)×1.

[0009] Because the scale of the convolution kernels differs, the lengths of the one-dimensional features created by the convolution operation also differ. To combine features, average pooling is used on parallel convolutional layers to achieve feature fusion. Assuming the scale of the average pooling layer is S, the average pooling operation is as follows: The final output of the other two channels repeats the above steps; where, It is the output matrix after activation by the activation function, and avg is the average pooling function. Yes The output matrix after average pooling has dimensions [(LK 1 +1) / S]×1; Yes The output matrix after average pooling has dimensions [(LK 2 +1) / S]×1; Yes The output matrix after average pooling has dimensions [(LK 3 +1) / S]×1; Therefore, the final output of the mCNN part is: Among them, F N N represents the matrix obtained by merging the feature matrices of the three channels, where N = {[(LK 1 +1) / S]+[(LK 2 + 1) / S]+[(LK 3 +1) / S]}×1 represents its dimension, This represents the matrix concatenation operator.

[0010] Furthermore, FN serves as the input to the LFLBs part, whose convolution is calculated as follows: Y (N-T+1)×1 =F (N-T+1)×T ·W T×1 +B (N-T+1)×1 ; where F (N-T+1)×TThis indicates that the input dimension is (N-T+1)×T during convolution operations, W T×1 B represents a weight matrix of dimension T×1. (N-T+1)×1 A paranoia matrix of dimension (N-T+1)×1, Y (N-T+1)×1 The output matrix of the convolution operation is represented as (N-T+1)×1. L2 regularization is added to each LFLB, which acts on the loss function to avoid overfitting. The definition of L2 regularization is as follows: Loss = Loss * +λ||W T×1 || 2 L2=λ||W T×1 || 2 λ is the regularization strength, i.e., the regularization parameter, which is a non-negative hyperparameter; ||W T×1 || 2 λ||W represents the square of the Frobenius norm of the weight matrix, calculated by squaring the absolute value of each element in the matrix, summing them, and then taking the square root; T×1 || 2 Represents the L2 regularization value; Loss * represents the original loss function, and Loss represents the loss function after weight decay.

[0011] Furthermore, the third part, the fully connected layer, classifies and outputs the features obtained from the LFLBs part. This part consists of three fully connected layers. In the last fully connected layer, the softmax function is used to output the probability distribution of the samples among three different event types.

[0012] Compared with the prior art, the advantages of this invention are as follows: (1) The core of this invention lies in the innovative design of the mCNN-LFLBs network model, which has demonstrated unprecedented performance in the fields of deep learning and signal processing. In addition, mCNN-LFLBs combines the advantages of multiple one-dimensional convolutional neural networks (1DCNN) to form a multi-dimensional, multi-functional multi-channel convolutional neural network (mCNN). This unique architecture design enables the algorithm to focus on both the short-term and long-term features of the signal when processing complex signals, thereby achieving more comprehensive and in-depth data analysis. In traditional deep learning models, the use of small-sized convolutional kernels is often limited to capturing high-frequency signal features, while ignoring the importance of low-frequency signals. mCNN effectively makes up for this deficiency by fusing convolutional kernels of different sizes, thereby achieving a better balance and integration between the high-frequency and low-frequency features of the signal.

[0013] (2) Another major innovation of the mCNN-LFLBs network model is the introduction of the concept of Local Feature Extraction Blocks (LFLBs). These LFLBs focus on learning the local correlations of signals and are able to extract and understand the correlations between layers. This structural design not only effectively reduces information redundancy in the data processing process, but also more accurately captures and presents the key features hidden in the signal. In practical applications, especially in high-requirement scenarios such as pipeline leak identification, the introduction of LFLBs has greatly improved the accuracy and efficiency of identification.

[0014] (3) From a technical perspective, the lightweight design of the algorithm in the mCNN-LFLBs network model is another major highlight. While maintaining high performance, the algorithm significantly reduces the number of network parameters required, which means shorter training time and faster inference speed. This efficient computational performance makes the mCNN-LFLBs network model algorithm suitable not only for high-performance computing platforms but also for resource-constrained environments, greatly expanding its application scope. In summary, the mCNN-LFLBs network model recognition algorithm of this invention not only demonstrates significant innovation and advancement at the technical level but also shows extremely high value in practical applications. It not only provides a new and efficient technical means for key areas such as natural gas pipeline inspection but also brings new ideas and methods to related research fields of deep learning and signal processing. Through this innovative algorithm, the safe transmission of energy can be more effectively guaranteed, laying a solid foundation for the future development of these fields. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 illustrates the hardware system principle of the oil and gas pipeline leakage wave identification and monitoring system of the present invention; Figure 2 shows the overall structure diagram of the mCNN-LFLBs self-learning algorithm of the present invention, where (a) is the overall network structure diagram and (b) is the LFLB block structure diagram; Figure 3 shows the algorithm identification flowchart of the present invention; Figure 4 shows the accuracy curve and loss curve of the present invention, where (a) is the accuracy curve of the training set and the validation set, and (b) is the loss curve of the training set and the validation set; Figure 5 shows the confusion matrix of the present invention.

[0017] The correspondence between the reference numerals and components in Figure 1 is as follows: 1-Pump laser; 2-Wavelength division multiplexing (WDM); 3-Distributed feedback fiber laser vibration sensors (DFB-FL); 4-Isolator; 5-Michelson interferometer; 6-Photoelectric detector (PD); 7-Phase generated carrier system (PGC); 8-Computer (PC). Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings.

[0019] Example 1: This invention extracts features from acoustic signals acquired by DFB by combining mCNN with LFLBs. The acquired acoustic signals are one-dimensional raw time-domain data. This input improves the completeness of information. The three parallel one-dimensional CNN used in this invention employs convolutional kernels of different sizes in each path. Smaller kernels provide better temporal resolution and are more sensitive to high-frequency feature transformations, but they cannot effectively learn low-frequency features present in the signal. However, larger kernels can learn information over a longer time span, i.e., low-frequency features present in the signal. Therefore, combining the advantages of different scales allows the algorithm to learn more features. Combined with LFLBs, this further extracts high-level local features, which is more conducive to classification. Finally, the features are classified using softmax.

[0020] The construction method of the mCNN-LFLBs network model is as follows: 1. Experimental site construction and data acquisition site construction: First, a natural gas pipeline experimental site is carefully designed and constructed in a natural environment. This experimental site should simulate the real-world natural gas pipeline environment, including pipeline layout, surrounding terrain, climate conditions, and other factors.

[0021] Data Acquisition: High-precision sensors and data acquisition equipment are used to collect experimental data on-site at the prepared experimental site. The types of data collected include acoustic wave information.

[0022] 2. Data Preprocessing and Database Construction: Data preprocessing involves segmenting and preprocessing the collected raw data to ensure data quality. Preprocessing steps include normalization and feature extraction, aiming to improve data usability and the efficiency of subsequent processing. This embodiment directly uses the raw signal for classification, simplifying the signal processing flow. The preprocessing process only involves dividing the time-domain acoustic signal; for example, if the original data length is 1*25000, it is divided into 1*5000.

[0023] Database Setup: Create a structured database to store and manage the preprocessed data. The database design should consider features such as fast data retrieval, efficient storage, and ease of maintenance to facilitate subsequent data processing and analysis.

[0024] 3. Network Architecture Design for Building the mCNN-LFLBs Network Model: A network model combining a multi-channel convolutional neural network (mCNN) with local and hierarchical fusion blocks (LFLBs) is designed and built. This model needs to comprehensively consider the temporal and frequency characteristics of the signal, as well as the special patterns in leakage events. The overall structure diagram of the mCNN-LFLBs self-learning algorithm in this invention is shown in Figure 2, where Figure 2(a) is the overall network structure diagram and Figure 2(b) is the LFLB block structure diagram.

[0025] Model optimization: Through experiments and adjustments, optimize network structure parameters, such as the number of layers, convolutional kernel size, activation function, etc., to achieve the best learning effect and recognition accuracy.

[0026] Parallel data processing and feature extraction: Data input: The preprocessed data is input into mCNN in parallel. Due to the parallel processing capabilities of mCNN, this can effectively speed up data processing.

[0027] Feature extraction: mCNN extracts different features from the data through its multi-layered structure, including but not limited to time-domain, frequency-domain, and modal features, as well as time-domain features for different frequency bands. This step is crucial for accurately understanding and identifying pipeline leak signals.

[0028] Feature merging and local feature extraction data merging: The multi-channel output data after mCNN processing is merged to form a comprehensive feature set.

[0029] Local Feature Extraction: The merged data is input into LFLBs for more in-depth local feature extraction. LFLBs are specifically optimized for local correlations, thereby identifying more refined patterns and features in the data.

[0030] 4. Data Classification and Model Training: Data Classification: The final data classification is completed in the softmax layer. This step classifies the data into different states such as normal operation and leakage based on the features extracted earlier.

[0031] Model Saving: Finally, save the model training parameters, including weights and biases. This is not only important for subsequent model evaluation and optimization, but also facilitates model reuse and deployment.

[0032] By following the steps outlined above, a highly efficient and accurate natural gas pipeline leak detection system can be developed. This system will then be able to be applied in practice.

[0033] Example 2: Applying the model from Example 1, the system is constructed as follows: As shown in Figure 1, an active evolutionary fiber optic pipeline leakage wave identification and monitoring system includes a laser source 1 for generating pump light. The laser source 1 is connected to a wavelength division multiplexer 2. After the laser source 1 is split by the wavelength division multiplexer 2, the pump light is transmitted to a sensor, which is a distributed feedback fiber laser vibration sensor 3. After receiving the pump light, the distributed feedback fiber laser vibration sensor 3 generates signal light and carries the acoustic wave signal caused by the pipeline leakage back to the COM terminal of the wavelength division multiplexer. The wavelength division multiplexer 2 is connected in sequence to an optical isolator 4 and a Michelson interferometer 5. The acoustic wave signal is output through the signal terminal of the wavelength division multiplexer 2, passes through the optical isolator 4, and enters the Michelson interferometer 5 to interfere. After interference, the signal is detected by a photodetector 6 and converted from an optical signal to an electrical signal. The photodetector 6 is connected in sequence to a phase generation carrier demodulation system 7 and a computer 8. The electrical signal is demodulated by the phase generation carrier demodulation system 7, and the original signal of the input model is obtained from the computer. In addition, the DFB-FL collects the pipeline signal, demodulates it using PGC, and then inputs the signal to the computer for classification using the mCNN-LFLBs network model on the computer. The mCNN-LFLBs network model combines a multi-channel convolutional neural network (mCNN) with local feature extraction blocks (LFLBs) and hierarchical fusion blocks. This network model includes mCNN, several LFLB blocks, and fully connected layers.

[0034] As detailed in Figure 2, mCNN is composed of three parallel one-dimensional convolutions with different kernel sizes. Except for the kernel size, each convolutional layer contains a batch normalization (BN) layer, an activation function layer, and an average pooling layer. The LFLBs consist of five LFLB blocks. The first four LFLBs are composed of convolutional layers, batch normalization layers, activation function layers, and average pooling layers, while the last LFLB uses global average pooling. The last fully connected layer consists of a dropout layer and a softmax layer. The overall network structure is mainly divided into four parts: input layer, convolutional layer, pooling layer, and output layer.

[0035] The convolution calculation for one channel of mCNN is as follows: Wherein, the data before the convolution operation has a dimension of 1×L, L=5000; K 1 This represents the size of the convolution kernel in one of the three channels, and the sizes of the other two channels are K. 2 K 3 ; The dimension is (LK) 1 +1)×K 1 The matrix; because the scale of the convolution kernel is K. 1 The convolution kernel moves with a stride of 1 each time, and moves a total of (LK) steps. 1 +1) times, so the dimension of the input data during the convolution operation is (LK) times. 1 +1)×K 1 ; Indicates a dimension of K 1 A weight matrix of ×1, The dimension is (LK) 1 +1)×1 paranoia matrix; This indicates that the input data is processed by a convolution kernel of size K. 1 The output matrix obtained by convolution of the channels has dimensions (LK). 1 +1)×1; tanh represents the nonlinear activation function ReLU; yes The output matrix after activation by the activation function has dimensions (LK). 1 +1)×1.

[0036] Because the scale of the convolution kernels differs, the lengths of the one-dimensional features created by the convolution operation also differ. To combine features, average pooling is used on parallel convolutional layers to achieve feature fusion. Assuming the scale of the average pooling layer is S, the average pooling operation is as follows:

[0037] in, It is the output matrix after activation by the activation function, and avg is the average pooling function. Yes The output matrix after average pooling has dimensions [(LK 1 +1) / S]×1; Yes The output matrix after average pooling has dimensions [(LK 2 +1) / S]×1; Yes The output matrix after average pooling has dimensions [(LK 3+1) / S]×1; Therefore, the final output of one channel of mCNN is: The final output of the other two channels repeats the above steps; where, F N N represents the matrix obtained by merging the feature matrices of the three channels, where N = {[(LK 1 +1) / S]+[(LK 2 + 1) / S]+[(LK 3 +1) / S]}×1 represents its dimension, This represents the matrix concatenation operator.

[0038] FN serves as the input to the LFLBs part, and the convolution in this part is calculated as follows: Y (N-T+1)×1 =F (N-T+1)×T ·W T×1 +B (N-T+1)×1

[0039] Among them, F (N-T+1)×T This indicates that the input dimension is (N-T+1)×T during convolution operations, W T×1 B represents a weight matrix of dimension T×1. (N-T+1)×1 A paranoia matrix of dimension (N-T+1)×1, Y (N-T+1)×1 It is represented as the output matrix of the convolution operation, with a dimension of (N-T+1)×1.

[0040] L2 regularization is added to each LFLB, which is applied to the loss function to avoid overfitting. The definition of L2 regularization is as follows: Loss = Loss * +λ||W T×1 || 2 L2=λ||W T×1 || 2

[0041] λ is the regularization strength, i.e., the regularization parameter, which is a non-negative hyperparameter; ||W T×1 || 2 λ||W represents the square of the Frobenius norm of the weight matrix, calculated by squaring the absolute value of each element in the matrix, summing them, and then taking the square root; T×1 || 2 Represents the L2 regularization value; Loss * represents the original loss function, and Loss represents the loss function after weight decay.

[0042] The third part, the fully connected layer, classifies and outputs the features obtained from the LFLBs part. This part consists of three fully connected layers. In the last fully connected layer, the softmax function is used to output the probability distribution of the samples among three different event types.

[0043] Example 3: Data acquisition was performed using a distributed feedback fiber laser, and the acquired data was segmented and preprocessed to construct a database. The pipe diameter was 0.52m, and three leakage types were identified: no leakage, leakage with a 3mm orifice diameter, and leakage with a 5mm orifice diameter. The experimental pipe had two different orifice diameters (3mm and 5mm) for venting to simulate leakage points. The occurrence of a leakage event was simulated by opening a drain valve. Three different types of vibration signals were acquired using a DFB-FL. Each vibration signal was a one-dimensional time-domain signal with a duration of 1s, and the number of sampling points was 2.5k.

[0044] As shown in Figure 1, the principle of the distributed feedback fiber laser vibration sensing system is as follows: Pump element 1 generates pump light, which is split by wavelength division multiplexing (WDM) 2 and then transmitted to the DFB-FL. Upon receiving the pump light, the DFB-FL generates signal light, carrying an acoustic signal caused by a pipe leak, and returns it to the COM terminal of the WDM. Subsequently, the signal is output through the 1550 signal terminal of the WDM, passes through optical isolator 4, and then enters the Michelson interferometer 5 for interference. The interference signal is detected by photodetector PD6 and converted from an optical signal to an electrical signal. Then, phase demodulation is performed by phase generation carrier demodulation system 7, and the original signal of the input model is obtained from PC terminal 8.

[0045] Each vibration signal is a one-dimensional time-domain signal with a duration of 1 second and 2.5k sampling points. Inputting excessively long data into the network may increase computational complexity. Therefore, in this embodiment, each signal is divided into 5 equal parts, resulting in a duration of 0.2 seconds for each signal and 5000 sampling points.

[0046] The mCNN-LFLBs network was built and trained. The mCNN-LFLBs network model consists of two parts, which will be described in detail below: The main structure of the network consists of three parts: mCNN, LFLBs composed of multiple LFLB blocks, and a fully connected layer. The mCNN is composed of three parallel one-dimensional convolutions with different kernel sizes. Except for the kernel size, each convolutional layer contains a batch normalization (BN) layer, an activation function layer, and an average pooling layer. The LFLBs consist of five LFLB blocks. The first four LFLBs are composed of convolutional layers, batch normalization layers, activation function layers, and average pooling layers, while the last LFLB uses global average pooling (GAP). The final fully connected layer mainly consists of a dropout layer and a softmax layer. The overall network structure is mainly divided into four parts: input layer, convolutional layer, pooling layer, and output layer. See Figure 1 for the specific network structure.

[0047] The convolution calculation for a specific channel in the mCNN part is as follows:

[0048] The final output of the other two channels repeats the above steps; where the data before the convolution operation has a dimension of 1×L, L=5000. K 1 This represents the size of the convolution kernel in one of the three channels, and the sizes of the other two channels are K. 2 K 3 . The dimension is (LK) 1 +1)×K 1 The matrix. Because the scale of the convolution kernel is K. 1 The convolution kernel moves with a stride of 1 each time, and moves a total of (LK) steps. 1 +1) times, so the dimension of the input data during the convolution operation is (LK) times. 1 +1)×K 1 . Indicates a dimension of K 1 A weight matrix of ×1, The dimension is (LK) 1 The paranoia matrix is ​​+1)×1. This indicates that the input data is processed by a convolution kernel of size K. 1 The output matrix obtained by convolution of the channels has dimensions (LK). 1 +1)×1. tanh represents the nonlinear activation function ReLU. yes The output matrix after activation by the activation function has dimensions (LK). 1 +1)×1.

[0049] Because the scale of the convolution kernels differs, the lengths of the one-dimensional features created by the convolution operation also differ. To combine features, average pooling is performed on parallel convolutional layers to achieve feature fusion. Assuming the scale of the average pooling layer is S, the average pooling operation is as follows:

[0050] in, It is the output matrix after activation by the activation function, and avg is the average pooling function. Yes The output matrix after average pooling has dimensions [(LK 1 +1) / S]×1. Therefore, the final output of the mCNN part is:

[0051] Among them, F N N represents the matrix obtained by merging the feature matrices of the three channels, where N = {[(LK1 +1) / S]+[(LK 2 +1) / S]+[(LK 3 +1) / S]}×1 represents its dimension. This represents the matrix concatenation operator.

[0052] The second part consists of five LFLBs with different configurations. Each LFLB is composed of a set of sequential convolutions. Initially, the LFLB consists of convolutional layers, batch normalization (BN) layers, exponential linear units (ELU), and max pooling layers, forming a two-dimensional convolutional structure. In this invention, the LFLB is modified into a one-dimensional convolutional structure and utilized. To avoid gradient vanishing, the original ELU activation function is replaced with ReLU, and average pooling is used instead of max pooling. The computation process of the second part is the same as the first part: convolution is performed first, followed by activation and then pooling, and finally the extracted features are output to the next part. FN is used as the input to this part, and the convolution calculation for this part is as follows: Y (N-T+1)×1 =F (N-T+1)×T ·W T×1 +B (N-T+1)×1

[0053] Among them, F (N-T+1)×T This indicates that the input dimension is (N-T+1)×T during convolution operations. W T×1 B represents a weight matrix of dimension T×1. (N-T+1)×1 A paranoia matrix of dimension (N-T+1)×1, Y (N-T+1)×1 It is represented as the output matrix of the convolution operation, with a dimension of (N-T+1)×1.

[0054] L2 regularization is added to each LFLB, which acts on the loss function to avoid overfitting. The definition of L2 regularization is as follows: Loss = Loss * +λ||W T×1 || 2 L2=λ||W T×1 || 2

[0055] λ is the regularization strength (regularization parameter), which is usually a non-negative hyperparameter. ||W T×1 || 2 λ||W represents the square of the Frobenius norm of the weight matrix, calculated by squared the absolute value of each element in the matrix, summing them, and then taking the square root. T×1 || 2 This represents the L2 regularization value. Loss * represents the original loss function, and Loss represents the loss function after weight decay.

[0056] The main task of the third part is to classify and output the features obtained earlier; this part consists of three fully connected layers. In the last fully connected layer, the softmax function is used to output the probability distribution of samples among three different event types; there are no convolution operations in this part.

[0057] The network structure of this invention was used to classify pipe data with non-leakage, 3mm leakage diameter, and 5mm leakage diameter. The classification accuracy reached 99%, as shown in Figure 4. The maximum value of the vertical axis in Figure 4 is 1, and the unit is dimensionless. Figure 4(a) shows the accuracy curves of the training set and the validation set, and Figure 4(b) shows the loss curves of the training set and the validation set.

[0058] The effectiveness of the trained model was verified using a confusion matrix, which is commonly used to summarize the prediction results for classification problems. Each row represents the true class of the data, each column represents the predicted class, and the sum of the data in each row represents the total number of data points for that class. Classification tests were performed using a pre-split test set, with each of the three classes having an equal proportion of data points. The classification results are shown in Figure 5.

Claims

1. A system for identifying and monitoring leakage waves in oil and gas pipelines, characterized in that: The system includes a laser source (1) for generating pump light, which is connected to a wavelength division multiplexer (2). After the laser source (1) is split by the wavelength division multiplexer (2), the pump light is transmitted to a sensor, which is a distributed feedback fiber laser vibration sensor (3). After receiving the pump light, the distributed feedback fiber laser vibration sensor generates a signal light and carries the acoustic signal caused by the pipeline leak back to the COM end of the wavelength division multiplexer. The wavelength division multiplexer (2) is connected in sequence to an optical isolator (4) and a Michelson interferometer (5). The acoustic signal is output through the signal end of the wavelength division multiplexer (2), and after passing through the optical isolator (4), it enters the Michelson interferometer (5) and interferes. After the interference signal is detected by a photodetector (6), it is converted from an optical signal into an electrical signal. The photodetector (6) is connected in sequence to a phase generation carrier demodulation system (7) and a computer (8). The electrical signal is phase demodulated by the phase generation carrier demodulation system (7) and the original signal of the input model is obtained from the computer. Among them, the network model mCNN-LFLBs, which combines multi-channel convolutional neural network (mCNN) with local feature extraction blocks (LFLBs) and hierarchical fusion blocks, includes mCNN, LFLBs composed of several LFLB blocks, and fully connected layers.

2. The system according to claim 1, characterized in that: mCNN is composed of three parallel one-dimensional convolutions with different kernel sizes. Except for the kernel size, each convolutional layer contains a batch normalization (BN) layer, an activation function layer, and an average pooling layer. The LFLBs consist of five LFLB blocks. The first four LFLBs are composed of convolutional layers, batch normalization layers, activation function layers, and average pooling layers, while the last LFLB uses global average pooling. The last fully connected layer consists of a dropout layer and a softmax layer. The overall network structure is mainly divided into four parts: input layer, convolutional layer, pooling layer, and output layer.

3. The system according to claim 2, characterized in that: The convolution calculation for one channel of mCNN is as follows: Wherein, the data before the convolution operation has a dimension of 1×L, L=5000; K 1 This represents the size of the convolution kernel in one of the three channels, and the sizes of the other two channels are K. 2 K 3 ; The dimension is (LK) 1 +1)×K 1 The matrix; because the scale of the convolution kernel is K. 1 The convolution kernel moves with a stride of 1 each time, and moves a total of (LK) steps. 1 +1) times, so the dimension of the input data during the convolution operation is (LK) times. 1 +1)×K 1 ; Indicates a dimension of K 1 A weight matrix of ×1, The dimension is (LK) 1 +1)×1 paranoia matrix; This indicates that the input data is processed by a convolution kernel of size K. 1 The output matrix obtained by convolution of the channels has dimensions (LK). 1 +1)×1; tanh represents the nonlinear activation function ReLU; yes The output matrix after activation by the activation function has dimensions (LK). 1 +1)×1; Because the scale of the convolution kernels differs, the lengths of the one-dimensional features created by the convolution operation also differ. To combine features, average pooling is used on parallel convolutional layers to achieve feature fusion. Assuming the scale of the average pooling layer is S, the average pooling operation is as follows: in, It is the output matrix after activation by the activation function, and avg is the average pooling function. Yes The output matrix after average pooling has dimensions [(LK 1 +1) / S]×1; Yes The output matrix after average pooling has dimensions [(LK 2 +1) / S]×1; Yes The output matrix after average pooling has dimensions [(LK 3 +1) / S]×1; Therefore, the final output of one channel of mCNN is: The final output of the other two channels repeats the above steps; Among them, F N N represents the matrix obtained by merging the feature matrices of the three channels, where N = {[(LK 1 +1) / S]+[(LK 2 +1) / S]+[(LK 3 +1) / S]}×1 represents its dimension, and ⊕ represents the matrix concatenation operator.

4. The system according to claim 2, characterized in that: FN serves as the input to the LFLBs part, and the convolution in this part is calculated as follows: Y (N-T+1)×1 =Y (N-T+1)×T ·W T×1 +B (N-T+1)×1 ; Among them, F (N-T+1)×T This indicates that the input dimension is (N-T+1)×T during convolution operations, W T×1 B represents a weight matrix of dimension T×1. (N-T+1)×1 A paranoia matrix of dimension (N-T+1)×1, Y (N-T+1)×1 It is represented as the output matrix of the convolution operation, with a dimension of (N-T+1)×1; L2 regularization is added to each LFLB. It is applied to the loss function to avoid overfitting. The definition of L2 regularization is as follows: Loss=Loss * +λ||W T×1 || 2 ,L2=λ||W T×1 || 2 ; λ is the regularization strength, i.e., the regularization parameter, which is a non-negative hyperparameter; ||W T×1 || 2 λ||W represents the square of the Frobenius norm of the weight matrix, calculated by squaring the absolute value of each element in the matrix, summing them, and then taking the square root; T×1 || 2 Represents the L2 regularization value; Loss * represents the original loss function, and Loss represents the loss function after weight decay.

5. The system according to claim 2, characterized in that: The third part, the fully connected layer, classifies and outputs the features obtained from the LFLBs part. This part consists of three fully connected layers. In the last fully connected layer, the softmax function is used to output the probability distribution of the samples among three different event types.

Images