Natural gas pipeline valve inner leakage identification method based on deep learning model

Through the deep learning model combined with multi-sensor data for valve internal leakage identification, the instability of single sensor data detection and insufficient multi-source data processing capabilities in the existing technology are solved, and efficient and accurate valve internal leakage detection is achieved to ensure that The safe operation of the pipeline system.

CN119004378APending Publication Date: 2024-11-22CHANGZHOU UNIV
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202411104927.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing valve internal leakage detection methods rely on a single sensor data and are susceptible to noise and environmental changes, resulting in unstable detection results and lack of comprehensive processing capabilities of multi-source data, resulting in low accuracy and robustness.

Method used

Using a deep learning model combined with multi-sensor data (pressure, flow, acoustic data), through multi-modal feature fusion, dynamic spatiotemporal feature extraction and adaptive attention mechanism, a deep learning model is constructed for valve leakage recognition, realizing comprehensive data processing and real-time monitoring.

Benefits of technology

It significantly improves the accuracy and robustness of valve leakage detection, reduces environmental noise interference, and realizes efficient and accurate detection of valve leakage. It has the characteristics of intelligence and automation, providing guarantees for the safe operation of natural gas pipeline systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004378A_ABST
    Figure CN119004378A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of valve inner leakage detection, in particular to a natural gas pipeline valve inner leakage identification method based on a deep learning model, which comprises the following steps: acquiring pressure data, flow data and acoustic data of a pipeline key position, and constructing an original operation data set; preprocessing the original operation data set to obtain a feature data set; and inputting the marked feature data set into a deep learning model comprising an input layer, a multi-modal feature fusion layer, a dynamic spatial-temporal feature extraction layer, an adaptive attention mechanism layer, a full connection layer and an output layer for training. The problems that existing data detection based on a single sensor is low in accuracy, long in consumed time and high in missing detection rate and false detection rate are solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of valve internal leakage detection, and in particular to a natural gas pipeline valve internal leakage identification method based on a deep learning model. Background Art

[0002] Valve internal leakage is a common and serious problem in pipeline systems. Internal leakage not only leads to energy waste, but also may cause safety hazards and even cause serious environmental pollution and economic losses. Therefore, timely and accurate detection of valve internal leakage to ensure the safe operation of pipeline systems is a technical problem that needs to be solved urgently.

[0003] Existing valve internal leakage detection methods mainly rely on signal processing technology and manual monitoring. These methods usually include pressure detection, acoustic detection and flow detection. However, these traditional methods have many defects and limitations, which are specifically manifested in the following aspects:

[0004] First, traditional signal processing methods often rely on a single type of sensor data, such as data from pressure sensors or acoustic sensors. This single data source detection method is easily affected by noise and environmental changes, resulting in unstable detection results and the risk of missed detection and false detection. Especially in complex pipeline systems, external environmental factors (such as temperature changes, mechanical vibrations, etc.) have a greater impact on sensor signals, further reducing the accuracy and reliability of detection.

[0005] Secondly, existing methods lack the ability to comprehensively process complex and diverse data. In modern pipeline systems, the operating status of valves is usually affected by multiple factors, and it is necessary to combine the data of multiple sensors such as pressure, flow, and acoustics for comprehensive analysis. However, traditional detection methods are difficult to effectively integrate and process these multi-source heterogeneous data, and cannot fully explore the potential information in the data, resulting in low accuracy and robustness of internal leakage detection.

[0006] In addition, the existing detection method, such as the application number CN202310551236.3, includes: S1, data acquisition and preprocessing; S2, building a multi-channel multi-scale convolutional neural network model; S3, using the multi-channel multi-scale convolutional neural network model trained in step S2 to diagnose valve leakage faults. In the preprocessing step of this method, the acoustic emission time domain signal is converted into a grayscale image through the signal-image mapping method. In the process of converting the time domain signal to the image, some details or features in the original sound signal may be lost due to quantization, sampling rate adjustment or image encoding steps, resulting in the converted image being unable to fully reflect all the information of the original sound signal; in addition, this method uses a multi-channel multi-scale convolutional neural network, but the convolutional neural network is easily affected by noise and environmental changes, resulting in unstable detection results, missed detections and false detections; and ignoring the spatiotemporal feature information in the leakage signal, affecting the detection accuracy in complex environments.

[0007] In addition, the existing data detection method based on a single sensor detects the data of a single sensor, which has the problems of low accuracy, long time consumption, high missed detection and false detection rate.

[0008] Based on the defects of the above methods, there is an urgent need for an efficient and accurate valve internal leakage detection method that can make full use of a variety of sensor data to achieve real-time monitoring and intelligent diagnosis of valve internal leakage. Summary of the invention

[0009] In view of the shortcomings of the existing methods, the present invention solves the problems of low accuracy, long time consumption, high missed detection and false detection rate in the existing data detection based on a single sensor.

[0010] The technical solution adopted by the present invention is: a method for identifying internal leakage of a natural gas pipeline valve based on a deep learning model comprises the following steps:

[0011] Step 1: Collect pressure data, flow data and acoustic data at key locations of the pipeline to construct the original operation data set;

[0012] As a preferred embodiment of the present invention, a pressure sensor, a flow sensor and an acoustic sensor are used to collect pressure data, flow data and acoustic data respectively.

[0013] As a preferred embodiment of the present invention, the key positions include: vicinity of valves, pipeline corners, pipeline joints and areas with significant pressure changes.

[0014] Step 2: Preprocess the original running data set to obtain a feature data set;

[0015] As a preferred implementation of the present invention, preprocessing includes: data cleaning, removal of outliers, removal of noise data, and standardization.

[0016] Step 3: Input the labeled feature data set into a deep learning model including an input layer, a multimodal feature fusion layer, a dynamic spatiotemporal feature extraction layer, an adaptive attention mechanism layer, a fully connected layer, and an output layer for training;

[0017] As a preferred embodiment of the present invention, the deep learning model includes:

[0018] The multimodal feature fusion layer is used to generate a comprehensive feature representation Z(t) = Fusion(P train,norm (t),F train,norm (t),A train,norm (t));

[0019] Using dynamic spatiotemporal feature extraction layer to obtain feature representation Among them, W k (t) represents the kth dynamic convolution kernel, σ represents the activation function, * represents the convolution operation, b k represents the bias term, K represents the number of dynamic convolution kernels;

[0020] Using adaptive attention mechanism layer to obtain weighted feature representation α i (t) represents the attention weight of the i-th time step;

[0021] The weighted feature is represented by H att (t) Input the fully connected layer and output the internal leakage state through the output layer.

[0022] As a preferred embodiment of the present invention, the conditional generative adversarial network layer is used to represent the weighted feature H att (t) Carry out enhanced expansion.

[0023] As a preferred embodiment of the present invention, the deep learning model further includes:

[0024] Initialize deep learning model weights and bias parameters;

[0025] Define a semi-supervised learning loss function that includes supervised learning loss for labeled data and unsupervised learning loss for unlabeled data, L(θ) = αL sup (θ)+βL unsup (θ), L sup represents the supervised learning loss with labeled data, L unsup represents the unsupervised learning loss of unlabeled data, α and β are weight parameters;

[0026] A weakly supervised learning method is used to calculate the gradient of the semi-supervised learning loss function with respect to the model parameters through the back propagation algorithm, and the model parameters are updated using the Adam optimization algorithm to minimize the comprehensive training error.

[0027] When both the training error and the validation error converge to stable values ​​and meet the expected accuracy, stop training.

[0028] As a preferred embodiment of the present invention, a cross-validation method is used to validate the deep learning model to evaluate the accuracy, recall rate and F1 value.

[0029] As a preferred embodiment of the present invention, a natural gas pipeline valve internal leakage identification system based on a deep learning model includes: a memory for storing instructions executable by a processor; and a processor for executing instructions to implement a natural gas pipeline valve internal leakage identification method based on a deep learning model.

[0030] As a preferred embodiment of the present invention, a computer-readable medium stores a computer program code, and when the computer program code is executed by a processor, a method for identifying internal leakage of a natural gas pipeline valve based on a deep learning model is implemented.

[0031] Beneficial effects of the present invention:

[0032] 1. Install pressure sensors, flow sensors and acoustic sensors at key locations of natural gas pipelines to collect real-time operation data of pipelines and valves; use multi-sensor data fusion technology to comprehensively process data from different types of sensors to generate a unified feature data set, which significantly improves the accuracy and robustness of internal leakage detection and reduces the impact of environmental noise and interference that is easily affected by single data source detection methods;

[0033] 2. Use the Keras framework to build and train a deep learning model. By learning historical operation data and known valve internal leakage case data, the deep learning model grasps the characteristics and laws of internal leakage, effectively extracts key features from complex data, and enhances the detection ability of the model. During the training process, the model parameters are continuously optimized through the back propagation algorithm to minimize the training error and improve the prediction accuracy of the model.

[0034] 3. Through the Internet of Things technology, the real-time data collected by the sensor is transmitted to the data preprocessing module. After data cleaning, noise reduction, standardization and feature extraction, it is input into the trained deep learning model for real-time detection. The system can automatically identify the internal leakage of the valve and promptly notify the maintenance personnel through the alarm mechanism, realizing real-time monitoring and intelligent diagnosis of valve internal leakage, avoiding the lag and inefficiency existing in traditional manual monitoring methods.

[0035] 4. Through multi-sensor data fusion, deep learning model training and real-time monitoring, efficient and accurate detection of valve internal leakage is achieved. It has the characteristics of intelligence, automation and high reliability, providing a strong guarantee for the safe operation of the natural gas pipeline system. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a logic diagram of the natural gas pipeline valve internal leakage identification method based on the deep learning model of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0038] like Figure 1 As shown, the natural gas pipeline valve internal leakage identification method based on the deep learning model includes the following steps:

[0039] Step 1: Install various types of sensors at key locations of natural gas pipelines, including pressure sensors, flow sensors, and acoustic sensors, to collect real-time operation data of pipelines and valves and construct original operation data sets;

[0040] Step 11: Install pressure sensors, flow sensors and acoustic sensors at key locations of the natural gas pipeline to collect pressure data, flow data and acoustic data of the pipeline and valves, respectively, and generate an initial data set:

[0041] Key locations include: near valves. Since valves are the main components where internal leakage may occur, placing sensors around the valves can directly monitor abnormal signals caused by internal leakage, such as 5-10 cm away from the valves. Key locations also include: corners and joints of pipelines, which are prone to stress concentration and structural fatigue, thus causing leakage. Key locations also include: areas with significant pressure changes. By placing sensors to monitor areas with large pressure fluctuations, abnormal conditions can be discovered in a timely manner.

[0042] Assume that the pressure data collected by the pressure sensor is recorded as P(t), where P(t) represents the pressure value at time t;

[0043] The flow data collected by the flow sensor is recorded as F(t), where F(t) represents the flow value at time t;

[0044] The acoustic data collected by the acoustic sensor is recorded as A(t), where A(t) represents the acoustic signal intensity at time t;

[0045] Step 12: Time synchronization processing is performed on the pressure data P(t), the flow data F(t) and the acoustic data A(t) to generate a synchronized data set. The time synchronization processing includes aligning the timestamps of different sensor data.

[0046] Step 13: Build the original running data set D raw , D rawIncludes time-synchronized pressure data P(t), flow data F(t) and acoustic data A(t):

[0047] D raw ={(P(t),F(t),A(t))|t=t1,t2,…,t n};

[0048] Among them, t1, t2, …, t n Indicates different points in time.

[0049] Step 2: The original running data set is transmitted to the data preprocessing module to preprocess the original running data set;

[0050] Preprocessing includes data cleaning, noise reduction, standardization and feature extraction.

[0051] Step 21: Use IoT technology to connect D raw Transmitted from sensor nodes to data preprocessing modules to ensure real-time and integrity of data;

[0052] Step 22: In the data preprocessing module, raw Perform data cleaning to remove outliers and noise data;

[0053] Step 23: Perform noise reduction on the cleaned data, and use a filtering algorithm to remove high-frequency noise components. The filtering algorithm includes but is not limited to wavelet transform or Fourier transform, etc., to obtain smooth pressure data P clean (t), flow data F clean (t) and acoustic data A clean (t);

[0054] Step 24: P clean (t), F clean (t) and A clean (t) Perform standardization to make the data meet the requirements of model input and convert data of different dimensions into the same dimension range, denoted as P norm (t), F norm (t), A norm (t), where:

[0055]

[0056] Among them, μ P , μ F and μ A Respectively represent the mean of pressure, flow and acoustic data, σ P , σ F and σ A represent the standard deviation of pressure, flow and acoustic data, respectively;

[0057] Step 25: Extract features from the standardized data to extract key features that can characterize the valve status and generate a feature data set D feat :

[0058] D feat ={(P norm (t),F norm (t),A norm (t))|t=t1,t2,…,t n}.

[0059] Step 3: Use the Keras framework to build a deep learning model;

[0060] Initial training data set D train It is from F feat The datasets selected and annotated for training deep learning models include historical operation data and known internal leakage data.

[0061] Step 31: Initial training data set D train The historical operation data in the valve include the pressure, flow and acoustic data under normal conditions, and the known valve internal leakage data include the pressure, flow and acoustic data under internal leakage conditions:

[0062] D train ={(P i ,F i ,A i ,y i )|i=1,2,…,N}];

[0063] Among them, P i ,F i ,A i Represent the pressure, flow and acoustic data of the i-th sample, y i Indicates the corresponding label, 0 indicates normal, 1 indicates endoleakage, and N is the number of samples.

[0064] Step 32: Initial training data set D train Perform data cleaning, noise reduction and standardization preprocessing to obtain the preprocessed training data set D train,clean ;D train,clean ={(P train,norm (t),F train,norm (t),A train,norm (t))|t=t1,t2,…,t n};

[0065] Step 33: Use the Keras framework to build an improved deep learning model. The deep learning model combines the characteristics of the natural gas pipeline valve, including:

[0066] Input layer: receives preprocessed pressure, flow and acoustic feature data P train,norm(t) 、F train,norm (t), A train,norm (t);

[0067] Multimodal feature fusion layer: Combining multimodal features, the pressure, flow and acoustic data are fused through the feature fusion network to generate a comprehensive feature representation Z(t):

[0068] Z(t)=Fusion(P train,norm (t),F train,norm (t),A train,norm (t));

[0069] The multimodal feature fusion layer fuses the pressure, flow and acoustic data through the feature fusion network to generate a comprehensive feature representation Z(t), combining multiple different types of sensor data and making full use of the complementary information of various types of data, thereby enhancing the integrity and robustness of the feature representation; traditional methods usually process a single type of data and are easily affected by noise and environmental changes, resulting in unstable detection results, while multimodal feature fusion can more comprehensively reflect the operating status of the valve, improve detection accuracy and robustness, and reduce the possibility of missed detection and false detection.

[0070] Dynamic spatiotemporal feature extraction layer: The dynamic spatiotemporal convolutional network is used to dynamically adjust the convolution kernel size and step size according to different operating states to adapt to different spatiotemporal features and obtain the feature representation H(t):

[0071]

[0072] Among them, W k (t) represents the kth dynamic convolution kernel, σ represents the activation function, * represents the convolution operation, b k represents the bias term, K represents the number of dynamic convolution kernels; K=3.

[0073] The dynamic spatiotemporal feature extraction layer adopts a dynamic spatiotemporal convolutional network, which dynamically adjusts the convolution kernel size and step size according to different operating states, adapts to different spatiotemporal features, and obtains the feature representation H(t). By introducing the dynamic convolution kernel, the model's ability to process complex spatiotemporal features is enhanced; while traditional convolutional neural networks use fixed convolution kernel size and step size, and cannot adapt to changes in different operating states; the dynamic spatiotemporal convolutional network of the present invention can adjust parameters according to the real-time operating state, improves the ability to extract spatiotemporal features, and improves the detection accuracy and adaptability of the model in complex environments.

[0074] Adaptive attention mechanism layer: Adaptive attention mechanism is introduced to dynamically weight the importance of endoleak identification by learning different features, highlighting key moments and important features, and obtaining weighted feature representation H att (t):

[0075]

[0076] Among them, α i (t) represents the attention weight of the i-th time step, e i (t) represents the attention score of the i-th time step, W a 、v a and b a are the weights and biases of the attention mechanism, T represents the total number of time steps;

[0077] Traditional deep learning models usually assign the same importance to all features, which may result in the model being unable to effectively highlight key features when processing complex data; the present invention introduces an adaptive attention mechanism that can dynamically adjust the importance of features according to actual conditions, thereby highlighting the most critical features and moments for endoleak identification, thereby improving the accuracy and robustness of detection.

[0078] It also includes: Building a conditional generative adversarial network layer: Using the conditional generative adversarial network to generate and enhance the endoleak data, expand the diversity of the endoleak data set, and optimize the deep learning model's ability to recognize endoleak features, that is, to represent the weighted feature H att (t) Carry out enhancement and expansion;

[0079] Among them, the generator loss function is:

[0080]

[0081] Discriminator loss function:

[0082]

[0083] Among them, λ is the regularization parameter, MSE is the mean square error, and x real is the real sample, G(z|c) is the generated sample, and D() is the discriminator;

[0084] Traditional models are prone to overfitting and perform poorly on unseen data when data is insufficient. The present invention can generate more endogenous leakage data through conditional generative adversarial networks, enhance the diversity of data sets, thereby improving the performance of the model in data-scarce conditions and enhancing the ability to recognize endogenous leakage features. This not only improves the generalization ability of the model, but also can better adapt to the diverse and complex scenarios in practical applications.

[0085] Fully connected layer: The weighted feature representation is input into the fully connected layer to generate the prediction result of endoleak identification;

[0086] Output layer: Output the predicted value Y(t) of the valve’s internal leakage state, where 0 represents a normal state and 1 represents an internal leakage state.

[0087] Step 4: Use the feature data set to train the deep learning model, adopt the supervised learning method, adjust the deep learning model parameters through the back propagation algorithm, set the convergence condition of minimizing the training error, and stop the iteration when the deep learning model convergence condition is reached and the expected accuracy is achieved;

[0088] Step 41: D feat Input deep learning model;

[0089] Step 42: Initialize the parameters of the deep learning model, including weights and biases. The weights are initialized using uniform distribution or normal distribution, and the biases are initialized to zero:

[0090]

[0091] Among them, U represents uniform distribution, N represents normal distribution, and d is the dimension of the input feature.

[0092] Step 43, define the loss function L(θ), where θ represents the parameters of the model, and the loss function is a semi-supervised learning loss function, which includes labeled data and unlabeled data:

[0093] L(θ)=αL sup (θ)+βL unsup (θ);

[0094] Among them, L sup represents the supervised learning loss with labeled data, L unsup represents the unsupervised learning loss of unlabeled data, α and β are weight parameters used to balance the losses of the two parts;

[0095] Supervised learning loss L with labeled data sup (θ) is defined as the binary cross entropy loss for natural gas pipeline valve internal leakage identification:

[0096]

[0097] Among them, y i is the true label of the i-th sample, 0 means normal, 1 means endoleak, p i is the valve internal leakage probability value predicted by the model, and N is the number of labeled samples;

[0098] The binary cross entropy loss function has good performance when dealing with binary classification problems and can effectively distinguish between the normal state and the internal leakage state; by minimizing the cross entropy loss, the model can fit the labeled data more accurately and improve the accuracy of identifying valve internal leakage; at the same time, the loss function has a stable convergence characteristic in the gradient descent process, which can speed up the training of the model.

[0099] Unsupervised learning loss L for unlabeled data unsup (θ) is defined as the consistency regularization loss, which encourages the model to maintain consistent output for unlabeled data under different perturbations:

[0100]

[0101] Among them, f(x j ; θ) is the jth unlabeled sample, For f(x j ; θ) is the sample after random perturbation, and M is the number of unlabeled samples.

[0102] The consistency regularization loss introduces random perturbations on unlabeled data, encouraging the model to maintain consistency in output when facing perturbed data. This can effectively utilize unlabeled data, improve the generalization ability of the model, and reduce overfitting. By combining the supervised learning loss of labeled data and the consistency regularization loss of unlabeled data, the model can maintain good performance even when data is scarce, further improving the accuracy and robustness of endoleak recognition.

[0103] Step 44: Use the weakly supervised learning method to calculate the gradient of the semi-supervised learning loss function L(θ) with respect to the model parameters through the back propagation algorithm, and use the Adam optimization algorithm to update the model parameters to minimize the comprehensive training error:

[0104]

[0105] Where η is the learning rate, is the gradient of the comprehensive loss function with respect to the model parameters;

[0106] Step 45: During the training process, use the validation data set to evaluate the model, monitor the training error and validation error, and optimize the model performance by adjusting the learning rate and other hyperparameters to prevent overfitting;

[0107] Step 46: When the training error and validation error of the model converge to stable values ​​and reach the expected accuracy index, stop training and obtain the final trained deep learning model M. trained :

[0108]

[0109] Among them, L val represents the loss function on the validation set, ∈ is the convergence threshold;

[0110] Step 47: Use the trained deep learning model M trained The data is saved in the system for use in subsequent steps to conduct real-time valve internal leakage detection and identification.

[0111] The cross-validation method is used to validate the trained deep learning model and evaluate the accuracy, recall rate and F1 value of the deep learning model.

[0112] The pre-processed real-time operation data set is converted into feature data and input into the verified deep learning model. The internal leakage of the valve is detected in real time through the reasoning ability of the deep learning model. The system can distinguish between normal operation and internal leakage and provide diagnostic results.

[0113] When an internal leak is detected, the system automatically triggers an alarm mechanism, notifies maintenance personnel, and generates specific maintenance recommendations based on the diagnostic results of the deep learning model to help quickly locate and repair the problem.

[0114] 1. When valve internal leakage is detected, the system uses the deep learning model M trained The diagnostic results automatically trigger the alarm mechanism, generate alarm signals, and notify maintenance personnel. The alarm signals include valve position, internal leakage type and degree of internal leakage;

[0115] 2. The system sends the alarm signal to the terminal device of the maintenance personnel through the Internet of Things platform;

[0116] 3. Based on the diagnosis results of the deep learning model, the system automatically generates specific maintenance suggestions, including the location of the internal leakage, possible causes and repair steps, to help maintenance personnel quickly locate and repair problems;

[0117] 4. The maintenance suggestion is included in the alarm signal, describing the diagnostic information of the internal leakage and the suggested repair method. The maintenance suggestion format is as follows:

[0118] Location:L valve ;

[0119] LeakTYpe:T leak ;

[0120] Leak Severity: S leak ;

[0121] RecommendedActions:R actions ;

[0122] Among them, L valve Indicates valve position, T leakIndicates the internal leakage type, S leak Indicates the degree of internal leakage, R actions Indicates recommended maintenance actions;

[0123] 5. After receiving the alarm signal and maintenance suggestions, the maintenance personnel will conduct on-site inspection and repair according to the recommended repair steps. The system monitors the valve status in real time through sensors;

[0124] 6. After the maintenance personnel complete the repair, the system records the maintenance process and results and updates the valve maintenance history information in the database;

[0125] 7. The system generates maintenance reports regularly.

[0126] Through big data analysis technology and the predictive power of deep learning models, historical data stored in the cloud is analyzed to provide early warning of potential valve failures and preventive maintenance recommendations;

[0127] Regularly update and optimize deep learning models, and use the latest operating data and fault data to continuously improve the detection accuracy and prediction capabilities of deep learning models.

[0128] Experimental data:

[0129] In a city's natural gas pipeline system, the operating company encountered frequent valve internal leakage problems and low detection efficiency and accuracy. In order to improve the detection efficiency and accuracy and ensure the safe operation of the pipeline system.

[0130] At a natural gas pipeline valve location, the sensor detects abnormal pressure, flow and acoustic signals, and the system transmits these data to the data preprocessing module for processing.

[0131] In order to verify the effectiveness of the method of the present invention, the detection effects of the method of the present invention and the traditional method (data detection method based on a single sensor) in the same scenario are compared, and the specific data are as follows.

[0132] Table 1 Comparison of test data between the method of the present invention and the traditional method

[0133]

[0134] It can be seen from Table 1 that the method of the present invention has significant advantages over the traditional method. First, the detection time of the method of the present invention is greatly reduced. In the detection of the same number of valves, the time is shortened by about 80%, which significantly improves the detection efficiency. Second, the detection accuracy of the method of the present invention is significantly improved, reaching an average of more than 96%, which is much higher than the traditional method of about 78%. In addition, the missed detection rate and false detection rate of the method of the present invention are reduced to less than 3% and about 2%, respectively, which are significantly better than the missed detection rate of 15%-20% and the false detection rate of 5% of the traditional method.

[0135] The method of the present invention not only improves the efficiency and accuracy of natural gas pipeline valve internal leakage detection, but also effectively reduces missed detection and false detection, ensures the safe operation of the pipeline system, and reduces maintenance costs and risks. Practical application results show that the method of the present invention has broad application prospects and significant practical value.

[0136] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A natural gas pipeline valve internal leakage identification method based on a deep learning model, characterized in that: The following steps are involved: Step 1: Collect pressure data, flow data and acoustic data at key locations of the pipeline to construct the original operation data set; Step 2: Preprocess the original running data set to obtain a feature data set; Step 3: Input the labeled feature data set into the deep learning model including input layer, multimodal feature fusion layer, dynamic spatiotemporal feature extraction layer, adaptive attention mechanism layer, fully connected layer and output layer for training.

2. The method for identifying internal leakage of natural gas pipeline valves based on a deep learning model according to claim 1 is characterized in that: Deep learning models include: The multimodal feature fusion layer is used to generate a comprehensive feature representation Z(t) = Fusion(P train,norm (t), F train,norm (t), A train,norm (t)); Output feature representation using dynamic spatiotemporal feature extraction layer Among them, W k (t) represents the kth dynamic convolution kernel, σ represents the activation function, * represents the convolution operation, b k represents the bias term, K represents the number of dynamic convolution kernels; Using adaptive attention layer to output weighted feature representation α i (t) represents the attention weight of the i-th time step; The weighted feature is represented by H att (t) Input the fully connected layer and output the internal leakage state through the output layer.

3. The natural gas pipeline valve internal leakage identification method based on deep learning model according to claim 2 is characterized in that: Using conditional generative adversarial network layers to represent weighted features H att (t) Carry out enhanced expansion.

4. The method for identifying internal leakage of natural gas pipeline valves based on a deep learning model according to claim 2 is characterized in that: Deep learning models also include: Initialize weight and bias parameters; Define the semi-supervised learning loss function, L(θ) = αL sup (θ)+βL unsup (θ), L sup represents the supervised learning loss with labeled data, L unsup represents the unsupervised learning loss of unlabeled data, α and β are weight parameters; A weakly supervised learning method is used to calculate the gradient of the semi-supervised learning loss function with respect to the model parameters through the back-propagation algorithm, and the model parameters are updated using the Adam optimization algorithm to minimize the comprehensive training error. When both the training error and the validation error converge to stable values ​​and meet the expected accuracy, stop training.

5. The method for identifying internal leakage of natural gas pipeline valves based on a deep learning model according to claim 1 is characterized in that: The pressure sensor, flow sensor and acoustic sensor are used to collect pressure data, flow data and acoustic data respectively.

6. The method for identifying internal leakage of natural gas pipeline valves based on a deep learning model according to claim 1 is characterized in that: Critical locations include: near valves, pipe corners, pipe joints, and areas with significant pressure changes.

7. The method for identifying internal leakage of natural gas pipeline valves based on a deep learning model according to claim 1 is characterized in that: Preprocessing includes: Data cleaning, outlier removal, noise data removal, and standardization.

8. The method for identifying internal leakage of natural gas pipeline valves based on a deep learning model according to claim 1 is characterized in that: The deep learning model was validated using the cross-validation method to evaluate accuracy, recall, and F1 value.

9. A natural gas pipeline valve internal leakage identification system based on a deep learning model, characterized in that: include: a memory for storing instructions executable by a processor; A processor, used to execute instructions to implement the method for identifying internal leakage of a natural gas pipeline valve based on a deep learning model as described in any one of claims 1 to 8.

10. A computer readable medium storing computer program code, characterized in that: When the computer program code is executed by a processor, it implements the method for identifying internal leakage of a natural gas pipeline valve based on a deep learning model as described in any one of claims 1 to 8.

Citation Information

Cited By

  • Hydrogen long-distance pipeline sensor optimization method based on artificial intelligence algorithm

    CN119249281A

  • An optimization method for hydrogen long-distance pipeline sensors based on artificial intelligence algorithm

    CN119249281B

  • Drain valve action monitoring method and device, medium and program product

    CN120488096A