Tubular column leakage point monitoring method and system based on DTS and DAS
By combining multimodal data fusion and deep learning models with the three-dimensional geometry and material properties of the tubing, the problem of insufficient accuracy and reliability in tubing leakage monitoring in existing technologies has been solved, achieving high-precision leak location and reliable monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-05
- Publication Date
- 2026-03-10
AI Technical Summary
Existing DTS/DAS-based tubing leak monitoring methods suffer from insufficient accuracy and reliability in accurately locating leak points due to their reliance on a single analytical factor and lack of comprehensiveness.
A multimodal data fusion method is adopted, which combines the three-dimensional geometry and material properties of the tubing to construct a baseline model. A deep learning model is then used to predict leaks. The model includes modules for data acquisition and modeling, data processing, deep learning model construction, and monitoring, to achieve multimodal feature representation and a two-branch multi-scale model structure.
It achieves high-precision location of leaks in the tubing, improves the reliability and accuracy of location, reduces the false alarm rate and missed alarm rate, and has good interpretability.
Smart Images
Figure CN121636934A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of pipe column monitoring, and relates to a pipe column leakage point monitoring method based on DTS and DAS. BACKGROUND
[0002] In recent years, energy pipe column systems play an increasingly important role in urban underground infrastructure construction. Due to long-term operation and the influence of the natural environment, the pipe column is prone to corrosion, aging and other problems, which in turn leads to leakage, bursting and other serious accidents. Therefore, timely detection and accurate positioning of pipe column leakage points are of great significance for ensuring the safety of urban energy supply, reducing economic losses and environmental risks.
[0003] Traditional pipe column monitoring mainly adopts manual patrol, pressure monitoring and other methods, which have low monitoring efficiency and low precision. With the continuous development of distributed optical fiber sensing technology (DOFS), pipe column monitoring technology based on distributed temperature sensing (DTS) and distributed acoustic sensing (DAS) has gradually emerged, showing excellent continuous online monitoring, high precision and long-distance monitoring advantages.
[0004] However, the existing DTS / DAS-based pipe column leakage monitoring methods still have some deficiencies in accurate positioning of leakage points. For example, some of these methods only rely on single modal analysis of temperature or acoustic data, lack of fusion of modal data, and are difficult to fully exploit the implicit multi-dimensional feature information; although some methods attempt to fuse multi-modal data, they mainly use empirical models or shallow machine learning models, which have limited interpretability and generalization ability; in addition, some methods ignore the importance of the structural characteristics of the pipe column itself on the leakage phenomenon, and the accuracy and reliability of the leakage point positioning need to be improved. SUMMARY
[0005] The purpose of the present application is to provide a pipe column leakage point monitoring method based on DTS and DAS, which solves the problem of insufficient accuracy and reliability of leakage point positioning in the prior art DTS / DAS pipe column leakage detection method due to single analysis factor and insufficient comprehensiveness.
[0006] Another purpose of the present application is to provide a pipe column leakage point monitoring system based on DTS and DAS, which solves the problem of low accuracy and reliability of leakage point positioning in the prior art DTS / DAS-based pipe column leakage detection method due to unreasonable model setting and insufficient fusion capability.
[0007] The technical problem solved by the present application is that the DTS and DAS based pipe string leak point monitoring method is implemented in the following steps:
[0008] Step 1, collecting modal data along the pipe string, constructing a reference model according to the modal data; collecting pipe string structure parameters and encoding the pipe string structure parameters to obtain spatial features;
[0009] Step 2, pre-processing the modal data obtained in step 1, extracting time series features from the pre-processed modal data, obtaining modal features according to the time series features; and constructing a multi-modal fusion feature representation according to the spatial features and the modal features;
[0010] Step 3, constructing a double-branch multi-scale fusion leak point prediction model, inputting the multi-modal fusion feature representation into the leak point prediction model, and outputting the pipe string leak point prediction result;
[0011] Step 4, using the monitoring module to monitor the pipe string in real time, evaluating the credibility of the leak point prediction result according to the reference model, and outputting the pipe string leak point position.
[0012] Another technical problem solved by the present application is a DTS and DAS based pipe string leak point monitoring system, which sequentially includes an acquisition modeling module, a data processing module, a deep learning model construction module, and a monitoring module, as follows:
[0013] 1) The acquisition modeling module transmits temperature data and acoustic wave data to the data processing module;
[0014] 2) The data processing module extracts time series features after pre-processing to obtain modal features; encodes the pipeline three-dimensional space information as a feature representation, and aligns and splices the temperature features, acoustic wave features and spatial features according to the position to construct a multi-modal fusion feature representation;
[0015] 3) The deep learning model construction module designs and trains a leak point prediction model, and deploys the obtained leak point position prediction result to the monitoring module;
[0016] 4) The monitoring module analyzes and predicts the updated temperature data, acoustic wave data and pipeline space data, fuses the encoded features and combines them with the spatial features to accurately locate the specific position of the leak source.
[0017] The pipe string leak point monitoring method and system has the beneficial effects that it fuses multi-modal data, combines the three-dimensional geometry and material properties of the pipe string, establishes a reference model and a deep learning based leak point prediction model; at the same time, it introduces innovative technologies such as multi-modal fusion feature representation and double-branch multi-scale model structure, realizes high-precision positioning of the pipe string leak point, and significantly improves the reliability. BRIEF DESCRIPTION OF DRAWINGS
[0018] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;
[0019] Figure 2 This is a schematic diagram of the structure of the computer device constructed according to Embodiment 3 of the present invention. Detailed Implementation
[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0021] The present invention provides a tubing leak detection method based on DTS and DAS, which is implemented according to the following steps:
[0022] Step 1: Collect modal data along the tubing column and construct a baseline model based on the modal data; collect the tubing column structural parameters and encode the tubing column structural parameters to obtain spatial features;
[0023] Step 1: All data collection and modeling processes can be integrated into a single data collection and modeling module.
[0024] Modal data includes temperature data and acoustic data;
[0025] The reference model includes a reference temperature model and a reference acoustic model. Temperature data is collected along the tubing using distributed temperature measurement (DTS) technology, and the reference temperature model is established based on the temperature data. Acoustic data is collected along the tubing using distributed acoustic measurement (DAS) technology, and the reference acoustic model under normal operating conditions is established based on the acoustic data.
[0026] The construction process of the reference acoustic wave model and the reference temperature model is the same. The construction process of the reference temperature model is as follows: real-time acquisition of the distribution of tubing temperature in time and space to obtain the spatiotemporal data matrix of tubing temperature, and singular value decomposition is performed; the singular values are arranged in order of magnitude, and the modal truncation representation is constructed after determining the truncated mode order; after performing spatial modal modeling and temporal modal modeling, the spatial modal model and the temporal modal model are matrix-combined; based on minimizing the residual between the matrix combination and the spatiotemporal data matrix of tubing temperature, the unknown parameters are solved to obtain the optimal decomposed reference temperature model.
[0027] Among them, the structural parameters of the tubular column include the three-dimensional geometric data and material property data of the tubular column, and a three-dimensional geometric model of the tubular column is constructed based on the collected three-dimensional geometric data;
[0028] The acquisition process of the spatial feature is: based on the tubular column three-dimensional geometric model and material attribute data, a multi-attribute weighted graph is constructed; a graph convolutional neural network GCN is designed, the input is the multi-attribute weighted graph, through the layer-by-layer graph convolution and pooling operation, the high-dimensional feature representation of the node is learned, and the local and global structural features of the tubular column are captured; a point cloud network PointNet++ is constructed, the three-dimensional point cloud of the tubular column is taken as the input, through the point-by-point feature extraction, local feature aggregation and global feature learning, the hierarchical feature representation of the point cloud is obtained; the output features of the GCN and the PointNet++ are fused, including the tubular column structural features learned based on the GCN and the tubular column geometric detail features learned based on the PointNet++, the importance of the two types of features is dynamically adjusted through the self-attention mechanism, and the fused features are obtained; the fused features are adaptively pooled and fully connected mapped, and the final spatial feature of the tubular column is obtained, which encodes the three-dimensional geometric structure and material attribute data information.
[0029] Step 2, the modal data obtained in step 1 is preprocessed, the time sequence features are extracted from the preprocessed modal data, and the modal features are obtained according to the time sequence features; and the multi-modal fusion feature representation is constructed according to the spatial features and the modal features obtained in step 1.
[0030] Among them, the modal features include temperature features and acoustic wave features; the temperature features are obtained after the statistical features and time-frequency domain features of the temperature data are extracted from the preprocessed temperature data, and the features are constructed, selected, fused, transformed and standardized; the acoustic wave features are obtained after the statistical features and time-frequency domain features of the acoustic wave data are extracted from the preprocessed acoustic wave data, and the features are constructed, selected, fused, transformed and standardized.
[0031] Among them, the time sequence features include the statistical features and time-frequency domain features of the temperature data and acoustic wave data.
[0032] The preprocessing process is as follows: check the integrity of the temperature data and acoustic wave data and clean up; perform time sequence segmentation; perform detrending operation on the temperature data; perform pre-emphasis processing on the acoustic wave data; perform standardization or normalization processing on the temperature sequence and acoustic wave sequence of each data segment respectively; perform abnormal point monitoring and repair.
[0033] All the data processing processes in the above step can be integrated into a data processing module.
[0034] The construction process of the multi-modal fusion feature representation is as follows: a self-attention mechanism is used to adaptively weight and fuse the spatial features, capturing long-range dependency relationships between different spatial positions; a conditional convolutional neural network is introduced to learn the local patterns and context relationships of the temperature features and the sound wave features, and the convolution kernel parameters are initialized by prior knowledge; a self-attention-gated recurrent network hybrid structure is used to encode the temperature feature and the sound wave feature sequence, capturing long-range dependencies and local state transitions; an adaptive gating mechanism is used to soft-switch and combine the temperature features and the sound wave features, obtaining temperature sound wave fusion features; a multi-layer perceptron combined with a kernel method is introduced to perform nonlinear transformation on the temperature sound wave fusion features; cross-modal self-attention pooling is used on the nonlinear mapping features to obtain the correlation representation between the mapped temperature features and the sound wave features, and the correlation representation is used as the temperature sound wave fusion feature vector after pooling; the temperature sound wave fusion feature vector after pooling is regularized to suppress overfitting.
[0035] Step 3, constructing a double-branch multi-scale fusion leak point prediction model, inputting the multi-modal fusion feature representation into the leak point prediction model, and outputting the leak point prediction result of the pipe column;
[0036] Wherein: the leak point prediction result includes a leak probability prediction and a leak point position prediction; the leak point prediction model includes a temperature branch network, a sound wave branch network, and a fusion branch network; the temperature branch network and the sound wave branch network respectively apply a convolutional neural network to encode the temperature features and the sound wave features; the fusion branch network fuses the two single-modal encoded features and combines them with the spatial features to capture the interaction between different modalities at multiple scales, and finally outputs the leak probability and position prediction, referred to as the leak point prediction result; the entire leak point prediction model is trained end-to-end.
[0037] All the above modeling processes can be integrated into a deep learning model construction module to construct a leak point prediction model.
[0038] Step 4, using the monitoring module to monitor the pipe column in real time, evaluating the credibility of the leak point prediction result according to the benchmark model, and outputting the leak point position of the pipe column.
[0039] The step of evaluating the credibility of the leak point prediction result and outputting the column leak point position comprises: constructing a hierarchical self-attention convolutional encoder to code and model the column space and the multi-source heterogeneous monitoring data, and capturing local and global context dependency relationships; fusing a Monte Carlo assistant and an extreme learning machine to construct a robust benchmark leak point prediction model based on abnormal diffusion; designing a self-attention variational inference module, and using deep variational inference to approximate the lower bound probability distribution of the leak point prediction; introducing a meta-learning meta-path, and adaptively fine-tuning the model parameters according to historical leak point samples; constructing a mask mechanism based on a generative adversarial network, and relieving the false alarm and false alarm defects of uncertain prediction results through adversarial training; using a multi-instance multi-label active learning framework, and adaptively obtaining leak point position labeling samples from artificial experts according to the prediction confidence, so as to continuously improve the model generalization ability; and based on a hierarchical Bayesian neural network, the prediction result of the column leak point position is finally evaluated, and the credible leak point coordinates are output.
[0040] The column leak point monitoring system based on the DTS and the DAS mainly comprises an acquisition modeling module, a data processing module, a deep learning model construction module and a monitoring module, and the following is explained:
[0041] 1) The acquisition modeling module acquires temperature data along the column by using the DTS technology, establishes a benchmark temperature model according to the temperature data, acquires sound wave data along the column by using the DAS technology, establishes a benchmark sound wave model according to the sound wave data, acquires column structure parameters, and obtains spatial features according to the column structure parameters; and the acquisition modeling module transmits the temperature data and the sound wave data to the data processing module;
[0042] 2) The data processing module receives the temperature data and the sound wave data, extracts time sequence features after preprocessing, and obtains modal features; encodes pipeline three-dimensional space information into feature representation, and aligns and splices the temperature features, the sound wave features and the spatial features according to positions to construct a multi-modal fusion feature representation;
[0043] 3) The deep learning model construction module is used for designing and training a double-branch multi-scale fusion leak point prediction model, which comprises a temperature branch network, a sound wave branch network and a fusion branch network. The two single-modal branch networks of temperature and sound wave use a convolutional neural network or a sequence model to code features; the fusion branch network fuses the coded features and combines them with spatial features to capture modal interaction relationships at multiple scales, and outputs leak probability and position prediction; the leak point prediction model is trained in an end-to-end manner, and the obtained leak point position prediction result is deployed to the monitoring module.
[0044] 4) a monitoring module that uses the trained leak prediction model to analyze and predict the updated temperature data, acoustic wave data, and pipeline spatial data, identifies suspected leak points, and through the fusion branch network, fuses the encoded features and combines them with the spatial features to accurately locate the specific location of the leak source.
[0045] In addition, the DTS and DAS-based pipe column leak monitoring system of the present application can also be expanded to connect a visualization display module, which realizes a series of visualization display functions.
[0046] Specifically, the leak monitoring and positioning results of the monitoring module are transmitted to the visualization display module. The visualization display module displays the pipeline structure model, temperature and acoustic wave online monitoring data, leak probability distribution, and accurate positioning location analysis results in a graphical and visual manner. The visualization display module mainly plays a role in the following aspects:
[0047] C1. During the data collection process in step 1, the visualization display module displays the three-dimensional structure model of the pipe column and the arrangement positions of each measuring point, etc. basic information, providing a reference for data collection work.
[0048] C2. In the data preprocessing in step 2, the visualization display module is used to present the original collected temperature, acoustic wave and other modal data, as well as the statistical features, time-frequency domain features, etc. extracted after preprocessing, so as to manually check and analyze the data quality.
[0049] C3. In the deep learning model modeling process in step 3, the visualization display module is used to display the training process of the deep neural network, the change of the loss function curve, etc. to help model debugging and parameter optimization.
[0050] C4. In the actual monitoring and prediction stage in step 4, the visualization display module is used to present the final leak probability prediction results, accurate positioning location, etc. analysis results in a two-dimensional or three-dimensional visual form, helping the operator to quickly and accurately master the running status of the pipe column.
[0051] Based on the above system structure, the present application also provides a computer device comprising a memory and a processor, the memory storing a computer program, and the processor executing the computer program to realize the steps of the above-mentioned DTS and DAS-based pipe column leak monitoring method.
[0052] Based on the above system structure, the present application also provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to realize the steps of the above-mentioned DTS and DAS-based pipe column leak monitoring method.
[0053] It should be noted that the specific limitations of the DTS and DAS-based pipe column leak point monitoring system can be referred to the limitations of the DTS and DAS-based pipe column leak point monitoring method in the above, which will not be repeated here. Each module in the above DTS and DAS-based pipe column leak point monitoring system can be realized by software, hardware and their combination in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor calls to execute the operations corresponding to each of the above modules.
[0054] Embodiment 1
[0055] Reference Figure 1 The DTS and DAS-based pipe column leak point monitoring method of this embodiment 1 is implemented according to the following steps:
[0056] Step 1: Collect modal data and pipe column structure parameters, and construct a reference model according to the modal data; collect pipe column structure parameters, and encode the pipe column structure parameters to obtain spatial features.
[0057] Specifically, the modal data includes temperature data and acoustic wave data, the temperature data is collected along the pipe column by using distributed temperature measurement (DTS) technology, and the acoustic wave data is collected along the pipe column by using distributed acoustic wave measurement (DAS) technology.
[0058] The reference model includes a reference temperature model and a reference acoustic wave model. The reference temperature model under normal working conditions is established according to the temperature data, and the reference acoustic wave model under normal working conditions is established according to the acoustic wave data.
[0059] Further, the construction process of the reference temperature model is:
[0060] 1.1) Collecting temperature spatiotemporal data. Using distributed temperature sensing fiber DTS technology, the distribution of pipe column temperature in time t and space position x is collected in real time, obtaining the spatiotemporal data matrix T(x,t) of pipe column temperature,
[0061] Perform singular value decomposition (SVD) on T(x,t), the expression is as follows:
[0062]
[0063] 1.2) Determine the truncated modal order r, the expression is as follows:
[0064]
[0065] Finally, construct the modal truncated representation T r (x,t), the expression is as follows:
[0066]
[0067] T r (x,t) preserves the target portion of the original temperature field total energy and removes the remaining small singular value modes, achieving the purpose of dimensionality reduction and noise reduction.
[0068] 1.3) After spatial mode modeling and time mode modeling, the spatial mode model and the time mode model are matrixed and combined.
[0069] After spatial mode modeling, for each spatial mode u i (x), a parameterized model or basis function expansion is established, and the expression is as follows:
[0070]
[0071] where a ij is an unknown coefficient that needs to be solved by a fitting algorithm; is a selected basis function family, which can be a known basis function such as polynomial, wavelet, Fourier, etc.; i is the i-th spatial mode; j is the index of the basis function. This formula uses a linear combination of known basis functions i (x) to approximate the complex spatial mode u ij (x), and the approximate expression is obtained by solving the coefficient a . The selection of the basis function will affect the fitting accuracy.
[0072] Then, time mode modeling is performed, and for each time mode v i (t), a time series model such as an autoregressive model AR(p) or an ARMA model is used, and the expression is as follows:
[0073]
[0074] where i is the i-th time mode; p is the autoregressive order AR(p); q is the moving average order MA(q); a k is the autoregressive coefficient, k from 1 to p; b l is the moving average coefficient, l from 0 to q; e(t-l) is the white noise term. This is an ARMA(p,q) autoregressive moving average model, which is widely used in time series modeling. When q=0, it becomes an AR(p) autoregressive model. This model fits the dynamic evolution of the time mode v i (t) through a linear combination of historical values (lag terms) and random disturbance terms at the time v i (t) is located. By solving the coefficients a k and b l , a compact expression of the time mode is obtained.
[0075] The constructed spatial mode and time mode models are matrixed and combined, and the expression is as follows:
[0076]
[0077] Among them, T r (x,t) is a matrix-based combination of spatial and temporal modal models.
[0078] 1.4) Based on T r Minimize the residual between (x,t) and T(x,t), and solve for the parameter a. ij α k β l The baseline temperature model with optimal decomposition is obtained.
[0079] The process of constructing the reference acoustic wave model is the same as that of constructing the reference temperature model, and will not be repeated here.
[0080] It should be noted that in the field of tubing monitoring, the state of the tubing being leak-free and operating safely is referred to as "normal operating condition." Normal operating condition includes the following characteristics: 1) The tubing is intact, with no damage, cracks, leaks, or other abnormalities. 2) The fluid inside the tubing (such as natural gas or crude oil) flows smoothly along the tubing direction without abnormal flow disturbances. 3) The temperature, pressure, and other process parameters inside and outside the tubing are at normal levels within the design operating range. 4) The external environment of the tubing, such as soil and geological conditions, remains normal, with no severe disturbances affecting the tubing. Under normal operating conditions, the temperature distribution along the tubing and the acoustic signals from pipe wall vibration will exhibit relatively stable and regular patterns and characteristics. Therefore, step 1 uses Distributed Temperature Measurement (DTS) and Distributed Acoustic Measurement (DAS) technologies to collect temperature and acoustic data of the tubing under normal operating conditions, establishing a reference temperature model and a reference acoustic model to reflect the "normal pattern" of temperature and acoustic waves during normal operation. Once a leak or other abnormal situation occurs in the tubing string, the fluid flow and temperature distribution inside the string will change, deviating from the pattern under normal operating conditions. By monitoring this deviation, it is possible to determine whether a leak has occurred, thereby further pinpointing the location. Therefore, "normal operating conditions" are relative to abnormal leak situations, and modeling and understanding the data patterns under normal operating conditions is the foundation and prerequisite for monitoring leak anomalies.
[0081] Specifically, the column structure parameters are collected through the acquisition and modeling module. After the column structure parameters are encoded through graph convolutional neural network or point cloud network, spatial features are obtained. The column structure parameters include the three-dimensional geometric data and material property data of the column.
[0082] Furthermore, after the acquisition and modeling module has acquired the three-dimensional geometric data and material property data, it constructs a three-dimensional geometric model of the pipe column based on the three-dimensional geometric data.
[0083] The structural parameters of the tubular column are encoded using a graph convolutional neural network or a point cloud network. Specifically, the three-dimensional geometric model and material property data of the tubular column are encoded using graph convolutional neural networks and point cloud networks. After encoding, spatial features are obtained. The process of obtaining spatial features is as follows:
[0084] A1. Based on the three-dimensional geometric model and material property data of the tubular column, a multi-attribute weighted graph G = (V, E, W) is constructed, where V is the node set, E is the edge set, and W is the edge weight matrix. The edge weight matrix W is used to encode geometric and material information.
[0085] A2. Design a graph convolutional neural network (GCN). The input is a graph G. Through stacked graph convolution and pooling operations, it learns the high-dimensional feature representation of nodes and captures the local and global structural features of the pipeline.
[0086] A3. Construct a point cloud network PointNet++, and convert the 3D point cloud of the pipe P = p i Using |i=1,2,...,N as input, a hierarchical feature representation of the point cloud is obtained through point-by-point feature extraction, local feature aggregation, and global feature learning;
[0087] A4. Further output features of GCN and PointNet++ G and F P By dynamically adjusting the importance of the two types of features through a self-attention mechanism, a fused feature F is obtained. fusion =Attention(F G ,F P ), where F G It is based on the column structure features learned by the Graph Convolutional Neural Network (GCN), F P It is based on the column geometric detail features learned from the PointNet++ point cloud network;
[0088] A5. Regarding the fusion feature F fusion Adaptive pooling and fully connected mapping are performed to obtain the final column space feature vector V. spatial It encodes geometric structure and material property information to prepare for subsequent multimodal fusion.
[0089] Existing methods based on convolutional neural networks or graph neural networks often only capture the topological structure or coarse geometric information of pipes, failing to simultaneously acquire fine local geometric details. This step, by fusing graph convolutional networks and point cloud networks, can simultaneously model the topological structure and fine geometric details of pipes, obtaining a more complete and richer spatial feature representation. This step encodes the material properties of the pipes into the edge weight matrix of the graph, enabling subsequent graph convolutional operations to consider material properties when aggregating node features, thus learning a more intrinsically consistent pipe representation. This approach of integrating prior knowledge into the model provides a new way to improve the model's generalization ability and interpretability. Furthermore, this step, by introducing a self-attention module, enhances the network's ability to model the correlations between local features. The self-attention mechanism can adaptively capture the intrinsic connections between different features within the same local region, obtaining a more compact and richer local feature representation, which helps improve the accuracy of subsequent leak detection.
[0090] Step 2: Preprocess the modal data (i.e., the temperature data and sound wave data collected in Step 1), extract time series features from the preprocessed modal data (temperature data and sound wave data), and obtain modal features based on the time series features; then align and stitch the spatial features and modal features according to their positions to construct a multimodal fusion feature representation.
[0091] Specifically, time series features include statistical features and time-frequency domain features, while modal features include temperature features and acoustic wave features.
[0092] 2.1) The preprocessing process is as follows:
[0093] B1. Check the integrity of temperature and acoustic data, and handle missing and outlier values; for temperature data, fill in a small number of missing values by interpolating nearby points and remove obvious outliers; for acoustic data, remove Gaussian white noise through signal processing techniques such as filtering.
[0094] B2. Time series segmentation: Divide a long time series into multiple overlapping segments according to a fixed window length (e.g., 1 hour) to ensure the continuity and integrity of the data within each segment;
[0095] B3. Temperature data detrending: Perform trend analysis on each temperature data segment, fit and eliminate the annual cycle trend term, and commonly used detrending methods such as wavelet decomposition and moving average can be used;
[0096] B4. Acoustic data pre-emphasis: Acoustic data segments are pre-emphasized using common pre-emphasis methods (such as high-pass filtering, first-order difference calculation, etc.) to increase the energy of high-frequency components;
[0097] B5. Data Standardization / Normalization: Standardize or normalize the temperature sequence and sound wave sequence of each data segment respectively;
[0098] B6. Outlier Detection and Repair: Monitor outliers on the standardized data (e.g., using the 3σ principle), and perform smooth interpolation on very few outliers to avoid excessive disturbance.
[0099] 2.2) Statistical and time-frequency domain features are extracted from the preprocessed temperature data. These features are then constructed, selected, fused, transformed, and standardized to obtain the temperature features. Similarly, statistical and time-frequency domain features are extracted from the preprocessed acoustic wave data. These features are then constructed, selected, fused, transformed, and standardized to obtain the acoustic wave features. The statistical features include mean, standard deviation, maximum, minimum, range, median, mode, quartiles, kurtosis, skewness, coefficient of variation, autocorrelation coefficient, and partial autocorrelation coefficient. The time-frequency domain features include Fourier transform coefficients, dominant frequency and amplitude / phase, wavelet approximation component statistics, wavelet detail component statistics, wavelet energy distribution, and short-time Fourier transform coefficients. Since this step uses existing technology, it will not be elaborated further here.
[0100] 2.3) Construct a multimodal fusion feature representation. The construction process is as follows:
[0101] 2.3.1) A self-attention mechanism is employed to adaptively weight and fuse spatial features, capturing long-range dependencies between different spatial locations. The self-attention mechanism learns the correlation weights between features at different spatial locations, enabling adaptive fusion of spatial features. Compared to simple concatenation or averaging, this method can better uncover long-range dependency patterns in space.
[0102] 2.3.2) A conditional convolutional neural network is introduced to learn the local patterns and contextual relationships of temperature and acoustic wave features. The convolutional kernel parameters are initialized with prior knowledge. The introduction of conditional convolutional kernels based on prior knowledge guides the convolutional network to learn the local patterns and contextual information of temperature and acoustic wave data. Prior knowledge can be obtained from physical laws or domain experience, introducing externally known information into feature encoding and improving the interpretability and generalization of the model.
[0103] 2.3.3) A hybrid structure of self-attention and gated recurrent networks is used to encode temperature and acoustic feature sequences, capturing long-range dependencies and local state transitions. The hybrid structure, which integrates self-attention and gated recurrent units, can simultaneously capture the long-range dependencies and local state transition patterns of the sequences, comprehensively encoding the dynamic behavior of temperature and acoustic data.
[0104] 2.3.4) An adaptive gating mechanism is employed to softly switch and combine temperature and acoustic features to obtain fused temperature and acoustic features. Based on a gating weighting mechanism, the two modal features are adaptively softly switched and combined. Unlike hard splicing, this mechanism empowers the model to select the primary and secondary weights of the two modalities, making it more flexible and able to automatically adapt to changes in data distribution.
[0105] 2.3.5) A multilayer perceptron combined with a kernel method is introduced to perform nonlinear transformation on the temperature-sound wave fusion features. The fused features are then mapped using a nonlinear mapping combining a multilayer perceptron and kernel techniques, improving the separation and expressive power of the fused features. The Gaussian radial basis kernel can automatically learn the local nonlinear structure of the data.
[0106] 2.3.6) Cross-modal self-attention pooling is applied to the nonlinear mapping features to obtain the correlation representation between the mapped temperature features and acoustic wave features, which is then used as the pooled temperature-acoustic wave fusion feature vector. A novel cross-modal self-attention pooling mechanism is designed here to calculate the self-attention weights of the two-modal mapping features in each dimension, obtaining a highly correlated overall feature representation between the two modes, thus fully exploiting the cross-modal correlation information.
[0107] 2.3.7) Regularize the pooled temperature-sound wave fusion feature vector to suppress overfitting. Here, L2 norm normalization of the fusion feature vector helps suppress model overfitting and improves the generalization ability of the fusion feature representation.
[0108] Step 3: Utilize the deep learning model building module to construct a two-branch, multi-scale fusion leak prediction model. Input the multi-modal fusion feature representation into the leak prediction model and output the leak prediction results of the pipeline.
[0109] Specifically, this step designs a two-branch, multi-scale fusion leak prediction model, which includes three parts: a temperature branch network, an acoustic branch network, and a fusion branch network. The temperature branch network and the acoustic branch network respectively apply convolutional neural networks to encode temperature features and acoustic features. The fusion branch network fuses the two single-modal encoded features and combines them with spatial features to capture the interaction between different modalities at multiple scales. Finally, it outputs the leakage probability and location prediction, which is called the leak prediction result.
[0110] The entire leak prediction model is trained end-to-end.
[0111] Step 4: Utilize the monitoring module to monitor the tubing string in real time, evaluate the reliability of the leak prediction results based on the baseline model, and output the location of the leak in the tubing string. The specific process is as follows:
[0112] The reliability of leak prediction results is evaluated based on a benchmark model, and the location of pipe leaks is output. The steps include: constructing a hierarchical self-attention convolutional encoder to encode and model the pipe space and multi-source heterogeneous monitoring data, capturing local and global contextual dependencies; integrating Monte Carlo assistant and extreme learning machine to construct a robust benchmark leak prediction model based on anomaly diffusion; designing a self-attention variational inference module, using deep variational inference to approximate the probability distribution of the lower bound of evidence for leak prediction; introducing meta-learning meta-paths to adaptively fine-tune model parameters based on historical leak examples; constructing a generative adversarial masking mechanism to mitigate the defects of false alarms and false alarms in uncertain prediction results through adversarial training; adopting a multi-instance multi-label active learning framework to adaptively obtain leak location labeled samples from human experts based on prediction confidence, continuously improving the model's generalization ability; and using a hierarchical Bayesian neural network to perform a final confidence evaluation of the pipe leak location prediction results and output reliable leak coordinates.
[0113] In summary, this embodiment 1 integrates multimodal data and combines the three-dimensional geometry and material properties of the tubing to establish a prior benchmark model and a deep learning-based leak prediction model; at the same time, it introduces innovative technologies such as multimodal fusion feature representation and bi-branch multi-scale model structure to achieve high-precision location of tubing leaks.
[0114] To verify the beneficial effects of Embodiment 1, scientific demonstration was conducted through economic benefit calculations and simulation experiments. Embodiment 1 compares and evaluates the performance of the method of this invention with three existing technologies across four dimensions: leak detection accuracy, miss rate, false alarm rate, and model interpretability.
[0115] The existing technologies have three comparative schemes: Comparison scheme 1 is a traditional machine learning-based leak detection method, comparison scheme 2 is a deep learning-based single-modal leak detection, and comparison scheme 3 is a multi-modal fusion leak detection (simple fusion strategy).
[0116] Table 1. Comparison of the Method of the Present Invention with Three Prior Art Methods
[0117] Scheme Leak monitoring accuracy False negative rate False positive rate Interpretability Comparative example 1 72.3％ 21.8％ 16.4％ In Comparative example 2 81.7％ 14.6％ 12.9％ Low Comparative example 3 86.2％ 11.2％ 9.8％ Low The method of the invention 93.6％ 4.7％ 5.1％ High
[0118] As shown in Table 1, the method of the present invention is significantly superior to the three comparative schemes, with the strongest ability to accurately detect leakage points, significantly reducing the risk of leakage and improving safety and reliability, effectively reducing false alarms, and lowering maintenance costs. Based on variational reasoning and other mechanisms, the present invention has good interpretability.
[0119] Example 2
[0120] This embodiment 2 provides a pipe column leak monitoring system based on DTS and DAS, which mainly consists of an acquisition and modeling module, a data processing module, a deep learning model construction module, and a monitoring module.
[0121] The data acquisition and modeling module transmits the acquired temperature and acoustic wave data to the data processing module as input for subsequent feature extraction and fusion. Specifically, the module uses DTS technology to acquire temperature data along the tubing and establishes a baseline temperature model based on the temperature data. Simultaneously, it uses DAS technology to acquire acoustic wave data along the tubing and establishes a baseline acoustic wave model based on the acoustic wave data.
[0122] The data processing module is used to preprocess the temperature data and acoustic data collected by the acquisition and modeling module, extract statistical features and time-frequency domain features respectively, and obtain temperature features and acoustic features; the three-dimensional spatial information of the pipeline is encoded into feature representation, and the temperature features, acoustic features and spatial features are aligned and stitched together according to their positions to construct a multimodal fusion feature input.
[0123] The deep learning model building module is used to design and train a two-branch, multi-scale fusion leak prediction model, comprising three parts: a temperature branch network, an acoustic branch network, and a fusion branch network. The temperature and acoustic branch networks (two single-modal branches) both use convolutional neural networks (or sequence models) for feature encoding. The fusion branch network fuses the encoded features and combines them with spatial features to capture modal interactions at multiple scales, outputting leak probability and location prediction. The leak prediction model undergoes end-to-end training. The leak prediction model then deploys the predicted leak locations to the monitoring module.
[0124] The monitoring module uses a trained leak prediction model to analyze and predict updated temperature data, acoustic data, and pipeline space data, identify suspected leak points, and then uses a fusion branch network to fuse the encoded features and combine them with spatial features to accurately locate the specific location of the leak source.
[0125] Example 3
[0126] Building upon Embodiment 2, Embodiment 3's pipeline leak monitoring system based on DTS and DAS further includes a visualization module. The leak monitoring and location results from the monitoring module are transmitted to the visualization module. The visualization module displays the pipeline structure model, online temperature and acoustic monitoring data, leak probability distribution, and precise location analysis results in a graphical and visual manner. This visualization module plays a crucial role in the following aspects:
[0127] C1. During the data acquisition process in step 1, the visualization module displays the three-dimensional structural model of the tubing, the location of each measuring point, and other basic information, providing a reference for the data acquisition work.
[0128] C2. In the data preprocessing in step 2, the visualization module is used to present the original collected modal data such as temperature and sound waves, as well as the statistical features and time-frequency domain features extracted after preprocessing, so that the data quality can be viewed and analyzed manually.
[0129] C3. In the deep learning model modeling process in step 3, the visualization module is used to show the training process of the deep neural network, the changes in the loss function curve, etc., to help with model debugging and parameter optimization.
[0130] C4. In the actual monitoring and prediction stage in step 4, the visualization module is used to present the final leakage probability prediction results, precise location and other analysis results in an intuitive two-dimensional or three-dimensional visualization form to help operators quickly and accurately grasp the operating status of the tubing string.
[0131] Example 4
[0132] Reference Figure 2 Based on Embodiments 2 and 3, Embodiment 4 of the present invention provides a computer device, which may be a server. The computer device includes a processor, a memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium.
[0133] The computer device's database is used to store motion monitoring data. The computer device's network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in any of the above embodiments of the DTS and DAS-based tubing leak detection method.
[0134] Example 5
[0135] This embodiment 5 provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps in the above embodiments of the pipe leakage monitoring method based on DTS and DAS.
[0136] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the methods described above. Any references to memory, storage, database, or other media used in this embodiment 5 can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.
Claims
1. A method for monitoring leaks in a pipe string based on DTS and DAS, characterized in that, The method is implemented according to the following steps: Step 1, collecting modal data along the pipe string, constructing a baseline model according to the modal data, collecting pipe string structure parameters, and encoding the pipe string structure parameters to obtain spatial features; Step 2, preprocessing the modal data obtained in step 1, extracting time series features from the preprocessed modal data, and obtaining modal features according to the time series features; Then, according to the spatial features and the modal features, a multi-modal fusion feature representation is constructed; Step 3, constructing a double-branch multi-scale fusion leak point prediction model, inputting the multi-modal fusion feature representation into the leak point prediction model, and outputting the leak point prediction result of the pipe string; Step 4, using a monitoring module to monitor the pipe string in real time, evaluating the credibility of the leak point prediction result according to the baseline model, and outputting the leak point position of the pipe string.
2. The DTS and DAS based pipe string leak monitoring method of claim 1, wherein, In step 1, the modal data includes temperature data and acoustic wave data; The baseline model includes a baseline temperature model and a baseline acoustic wave model. Temperature data is collected along the pipe string using distributed temperature measurement technology, and a baseline temperature model is established according to the temperature data. Acoustic wave data is collected along the pipe string using distributed acoustic wave measurement technology, and a baseline acoustic wave model under normal working conditions is established according to the acoustic wave data; The construction process of the baseline acoustic wave model and the baseline temperature model is the same, wherein the construction process of the baseline temperature model is: real-time acquisition of the distribution of pipe string temperature in time and space position, obtaining the space-time data matrix of pipe string temperature, and singular value decomposition; arrange the singular values in size, determine the modal truncation representation after determining the modal truncation order; after spatial modal modeling and time modal modeling, matrix combination of the spatial modal model and the time modal model; based on the matrix combination and the space-time data matrix of the pipe string temperature, the residual error is minimized, and the unknown parameters are solved to obtain the optimal decomposition baseline temperature model; The pipe string structure parameters include three-dimensional geometric data and material attribute data of the pipe string, and a three-dimensional geometric model of the pipe string is constructed according to the collected three-dimensional geometric data.
3. The DTS and DAS based pipe string leak monitoring method of claim 1, wherein, In step 1, the acquisition process of the spatial features is: Based on the three-dimensional geometric model of the pipe string and the material attribute data, a multi-attribute weighted graph is constructed; a graph convolutional neural network (GCN) is designed, the input is a multi-attribute weighted graph, and through layer-by-layer graph convolution and pooling operation, the high-dimensional feature representation of the node is learned to capture the local and global structural features of the pipe string; a point cloud network PointNet++ is constructed, the three-dimensional point cloud of the pipe string is taken as the input, and through point-by-point feature extraction, local feature aggregation and global feature learning, the hierarchical feature representation of the point cloud is obtained; Fusion of the output features of GCN and PointNet++, including pipe string structure features learned based on graph convolutional neural network (GCN) and pipe string geometric detail features learned based on point cloud network (PointNet++), the importance of the two types of features is dynamically adjusted through self-attention mechanism to obtain fusion features; Adaptive pooling and full connection mapping are performed on the fusion features to obtain the final spatial features of the pipe string, which encode three-dimensional geometric structure and material attribute data information.
4. The DTS and DAS based pipe string leak monitoring method of claim 1, wherein, In step 2, the modal features include temperature features and acoustic wave features; The temperature feature is a statistical feature and a time-frequency domain feature of temperature data extracted from preprocessed temperature data, and is obtained after feature construction, selection, fusion, transformation, and standardization processing; The acoustic wave feature is a statistical feature and a time-frequency domain feature of acoustic wave data extracted from preprocessed acoustic wave data, and is obtained after feature construction, selection, fusion, transformation, and standardization processing; The time sequence feature includes statistical features and time-frequency domain features of temperature data and acoustic wave data; The preprocessing process is as follows: checking the integrity of the temperature data and the acoustic wave data and cleaning; time sequence segmentation; detrending the temperature data; pre-emphasizing the acoustic wave data; standardizing or normalizing the temperature sequence and the acoustic wave sequence of each data segment; monitoring and repairing abnormal points.
5. The DTS and DAS based pipe string leak monitoring method of claim 1, wherein, In step 2, the construction process of the multi-modal fusion feature representation is as follows: An adaptive weighting fusion of spatial features is performed by using a self-attention mechanism to capture long-range dependency relationships between different spatial positions; a conditional convolutional neural network is introduced to learn local patterns and context relationships of temperature features and acoustic wave features, and the convolution kernel parameters are initialized by prior knowledge; a self-attention-gated recurrent network hybrid structure is used to encode temperature features and acoustic wave feature sequences to capture long-range dependencies and local state transitions; an adaptive gating mechanism is used to combine temperature features and acoustic wave features to obtain temperature-acoustic wave fusion features; a multi-layer perceptron combined with a kernel method is introduced to perform nonlinear transformation on the temperature-acoustic wave fusion features; Cross-modal self-attention pooling is used on the nonlinear mapping features to obtain the correlation representation between the mapped temperature features and acoustic wave features, and to obtain the pooled temperature-acoustic wave fusion feature vector; the pooled temperature-acoustic wave fusion feature vector is regularized to suppress overfitting.
6. The DTS and DAS based pipe string leak monitoring method of claim 1, wherein, In step 3, the leak point prediction result includes a leak probability prediction and a leak point position prediction; The leak point prediction model includes a temperature branch network, an acoustic wave branch network, and a fusion branch network; the temperature branch network and the acoustic wave branch network respectively apply a convolutional neural network to encode temperature features and acoustic wave features; the fusion branch network combines the two single-modal encoding features and combines them with spatial features to capture the interaction between different modalities at multiple scales, and finally outputs the leak probability and position prediction, which is called the leak point prediction result.
7. The DTS and DAS based pipe string leak monitoring method of claim 1, wherein, In step 4, the specific process is as follows: A hierarchical self-attention convolutional encoder is constructed to encode and model the pipe column space and the multi-source heterogeneous monitoring data to capture local and global context dependencies; a robust benchmark leak point prediction model based on anomaly diffusion is constructed by combining Monte Carlo assistants and extreme learning machines; a self-attention variational inference module is designed to approximate the lower bound probability distribution of the leak point prediction using deep variational inference; a meta-learning meta-path is introduced to adaptively fine-tune the model parameters according to historical leak point samples; a mask mechanism based on generative adversarial is constructed to alleviate the false alarm and false negative defects of uncertain prediction results through adversarial training; Adopting a multi-instance multi-label active learning framework, the leak location labeling samples are adaptively obtained according to the prediction confidence; Based on the hierarchical Bayesian neural network, the final confidence of the prediction result of the pipe column leak location is evaluated, and the reliable leak point coordinates are output.
8. A DTS and DAS based pipe string leak monitoring system characterized by: The acquisition modeling module, the data processing module, the deep learning model construction module and the monitoring module are sequentially included, and the following is explained: 1) The acquisition modeling module transmits the temperature data and the acoustic wave data to the data processing module; 2) The data processing module extracts the time sequence features after preprocessing to obtain the modal features; The three-dimensional space information of the pipeline is encoded into a feature representation, and the temperature features, the acoustic wave features and the spatial features are aligned and spliced according to the positions to construct a multi-modal fusion feature representation; 3) The deep learning model construction module designs and trains the leak prediction model, and the obtained leak location prediction result is deployed to the monitoring module; 4) The monitoring module analyzes and predicts the updated temperature data, acoustic wave data and pipeline space data, combines the encoded features with the spatial features, and accurately locates the specific position of the leak source.
9. The DTS and DAS based pipe string leak monitoring system of claim 8, wherein: The acquisition modeling module uses DTS technology to collect temperature data along the pipe column, establishes a baseline temperature model according to the temperature data, uses DAS technology to collect acoustic wave data along the pipe column, establishes a baseline acoustic wave model according to the acoustic wave data, and collects pipe column structure parameters to obtain spatial features.
10. The DTS and DAS based pipe string leak monitoring system of claim 8, wherein: The leak prediction model includes a temperature branch network, an acoustic wave branch network and a fusion branch network; the two single-modal branch networks of temperature and acoustic wave use convolutional neural networks to encode features; the fusion branch network combines the encoded features with the spatial features, captures the modal interaction relationship at multiple scales, and outputs the leak probability and location prediction.