WO2020133639A1 - 一种管道内检测漏磁数据智能分析系统 - Google Patents

一种管道内检测漏磁数据智能分析系统 Download PDF

Info

Publication number
WO2020133639A1
WO2020133639A1 PCT/CN2019/074907 CN2019074907W WO2020133639A1 WO 2020133639 A1 WO2020133639 A1 WO 2020133639A1 CN 2019074907 W CN2019074907 W CN 2019074907W WO 2020133639 A1 WO2020133639 A1 WO 2020133639A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
defect
magnetic flux
flux leakage
peak
Prior art date
Application number
PCT/CN2019/074907
Other languages
English (en)
French (fr)
Inventor
刘金海
付明芮
卢森骧
张化光
马大中
汪刚
冯健
张鑫博
于歌
魏红秋
Original Assignee
东北大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 东北大学 filed Critical 东北大学
Priority to US16/345,657 priority Critical patent/US11488010B2/en
Publication of WO2020133639A1 publication Critical patent/WO2020133639A1/zh

Links

Images

Classifications

    • FMECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
    • F17STORING OR DISTRIBUTING GASES OR LIQUIDS
    • F17DPIPE-LINE SYSTEMS; PIPE-LINES
    • F17D5/00Protection or supervision of installations
    • F17D5/02Preventing, monitoring, or locating loss
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01NINVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/72Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables
    • G01N27/82Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws
    • G01N27/83Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws by investigating stray magnetic fields
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology

Definitions

  • the invention belongs to the technical field of pipeline detection, and particularly relates to an intelligent analysis system for detecting magnetic flux leakage data in a pipeline.
  • Pipeline transportation is widely used as a continuous, economical, efficient and green transportation method.
  • the national standard stipulates that the design life of the pipeline is 20 years. With the increase of the running time, the damage caused by the pipeline material problem, construction, corrosion and external force will make the pipeline condition deteriorate year by year, and the potential danger will increase sharply. Once a leak occurs, it will not only cause atmospheric pollution, but also easily cause a violent explosion. Therefore, in order to ensure the safety of energy transportation and ecological environment, the pipeline must be regularly inspected and maintained.
  • Non-destructive testing is widely used as an important hand for pipeline safety maintenance.
  • the main methods used for pipeline inspection include magnetic flux leakage inspection, eddy current inspection and ultrasonic inspection.
  • magnetic flux leakage detection is widely used in nearly 90% of pipelines in service. It is a relatively mature and widely used defect detection technology for ferromagnetic materials in foreign developed countries.
  • there are many researches on magnetic flux leakage data including data preprocessing, inspection, size inversion, and data presentation.
  • the current research on magnetic flux leakage data analysis focuses too much on the development of local points, lacks a systematic view on data analysis, and most of the theoretical methods and application technologies lack versatility and portability. Magnetic data analysis is effectively combined, it is difficult to form a set of practical and widely portable data analysis system.
  • the invention invents a data analysis software system for detecting magnetic flux leakage in pipelines from the perspective of face and body; invents a data analysis method from the perspective of artificial intelligence and invents a data preprocessing method based on sparse sampling in the time domain and KNN-softmax, Invented a pipeline connection component detection method based on the combination of selective search and convolutional neural network, invented an abnormal candidate region search and recognition method based on Lagrange multiplication framework and multi-source magnetic flux leakage data fusion, invented a Based on the defect inversion method of random forest, an improved pipeline defect evaluation method based on ASME B31G standard was invented.
  • an intelligent analysis system for in-pipe magnetic flux leakage data detection in which an intelligent analysis system for in-pipe magnetic flux leakage data analysis includes: a data complete set construction module, a discovery module, a quantization module, and a solution Module
  • the original sampled magnetic flux leakage data is connected with the data complete set construction module, and the data complete set construction module is connected with the discovery module through the complete magnetic flux leakage data set, the discovery module is connected with the quantization module, and the quantization module is connected with the solution module;
  • the data complete set construction module is used for data loss reconstruction and noise reduction operations on the original magnetic flux leakage detection data, and adopts a data complete set construction method based on sparse sampling in the time domain and KNN-softmax to construct a complete magnetic flux leakage data set ;
  • the original sampled magnetic flux leakage data is used as multi-source data information, which specifically includes: axial data, radial data, circumferential data, and alpha data;
  • the discovery module is for defect detection, and its process includes component detection and anomaly detection, wherein the component detection completes the detection of welds and flanges of pipeline connection components; the discovery module uses a selective search and convolutional neural network Combined pipeline connection component discovery method to obtain the precise position of the weld; according to the precise position of the weld, the magnetic flux leakage signal of the entire section of the pipeline is u+1 segment respectively, and one of the magnetic flux leakage signals is taken and multiplied based on Lagrangian number Frame and multi-source magnetic flux leakage data fusion anomaly candidate area search and identification method to find out defective magnetic flux leakage signals;
  • Anomaly detection includes: defect, valve instrumentation and metal addition detection, and finally get the defect signal;
  • the quantization module completes the mapping of defect signals to physical characteristics, adopts a defect quantization method based on random forest, and finally gives the size of the defect, that is, length, width and depth;
  • the solution module extracts all defect length columns, depth columns and pipeline attribute parameters in the defect information from the complete magnetic flux leakage data set, adopts an improved pipeline solution based on the ASME B31G standard, and finally gives the evaluation results through the maintenance decision model.
  • the evaluation results include the repair index and repair suggestions for the individual defect locations.
  • conduit attribute parameters comprises a minimum yield strength of SMYS, a minimum tensile strength SMTS, nominal diameter D d, and the thickness t a maximum allowable operating pressure MAOP;
  • the data complete set construction module adopts a data complete set construction method based on sparse sampling in the time domain and KNN-softmax to obtain a complete magnetic flux leakage data set, which specifically includes the following steps:
  • Step 1.1 Collect raw magnetic flux leakage detection data directly from the subsea pipeline magnetic flux leakage detector, and perform secondary baseline correction on the data.
  • the original sampled magnetic flux leakage data is used as multi-source data information, including: axial data, radial Data, circumferential data, alpha data;
  • Step 1.1.1 Perform a baseline correction on the original magnetic flux leakage detection data, expressed as:
  • k c is the number of mileage counting points; Is the original value of the position of the j a channel at the i a mileage counting point; Is the corrected value of the j a channel at the i a mileage counting point; s is the median of all channels, and n a is the number of magnetic flux leakage internal detector channels.
  • Step 1.1.2 Remove the excess value ⁇ T a from the data, and assign the position value of the excess value to the median s of all channels, expressed as:
  • Step 1.1.3 Perform secondary baseline correction on the data beyond the limit:
  • k c is the number of mileage counting points; It is the primary correction value of the j a channel at the i a mileage counting point; It is the value after the second correction of the j a channel at the position of the i a mileage counting point; s′ is the median of all the channels after one correction.
  • Step 1.2 Perform quasi-time-domain sparse sampling anomaly detection processing on the data after the second baseline correction
  • Step 1.2.1 Perform anomalous signal time-domain modeling on the data after the secondary baseline correction, that is, correspond to the sampling points as time information;
  • Step 1.2.1.1 mathematical modeling of the abnormal part, the modeling results are as follows:
  • p(t)′ represents the pipeline magnetic flux leakage detection voltage bulge compensation signal
  • f represents the signal sampling rate
  • t represents the sampling time
  • t 1 and t 2 represent the sampling time interval
  • a represents the power pipeline
  • n represents the system fluctuation amplitude Coefficient, f(t)' frequency of voltage waveform change
  • Step 1.2.1.2 Set the amount of change in abnormal data for magnetic flux leakage detection with the interval as the collection unit, take k e collected data as an interval, consider the variance of the pipeline system voltage data collected in each interval as the data change, and determine the leakage Magnetic data voltage signal fluctuation degree, the specific calculation method is:
  • I c represents the sampling points within a given interval
  • It represents the average value of the voltage data of the pipeline system collected in this interval
  • ⁇ f 0 represents the degree of fluctuation of the voltage signal of the magnetic flux leakage data.
  • Step 1.2.1.3 Calculate the amount of voltage state change The formula is as follows:
  • Step 1.2.2 Judgment of abnormal signal: if Then the data at this time is regarded as an abnormality caused by external disturbance, which is an abnormal part.
  • Step 1.2.3 Manual extraction of training sample features Manually extract test sample features Manually extract data features to be interpolated i b , j b , k d are the number of features.
  • Step 1.3 Perform the missing interpolation process based on KNN-logistic regression on the magnetic flux leakage data of the submarine pipeline.
  • Step 1.3.1 Training and testing KNN and softmax regression models.
  • Step 1.3.1.1 divides the feature sample data T into two parts, a part of the feature sample data X Train is used to train the KNN model, and another part of the feature sample data T Test is used to test the KNN model.
  • Step 1.3.1.2 Input X Train into the KNN model, determine a K value, and train the KNN model.
  • Step 1.3.1.4 Feature sample data classified into each category
  • the data set corresponding to the feature sample is Correct with Normalized separately with Expressed as:
  • Step 1.3.1.5 Add a softmax regression model at each type of node, the assumption function is as follows:
  • x is the sample input value and y is the sample output value;
  • is the training model parameter;
  • k f is the vector dimension;
  • i e is the i e category of the classification;
  • p(y i e
  • x) represents the category i e estimated probability value.
  • Step 1.3.1.6 Set the training sample set at each node Input into softmax regression model to get the output value after interpolation
  • the loss function J( ⁇ ) is:
  • x is the sample input value
  • y is the sample output value
  • is the training model parameter
  • k f is the vector dimension
  • i e is the i e category of the classification
  • j e is the j e sample input in the classification
  • M d is the number of samples
  • 1 ⁇ is an indicative function, if the braces are true, the expression value is 1.
  • Step 1.3.3 Input the characteristics and data set of the data to be interpolated into the trained model to implement the interpolation of the missing data to obtain a complete magnetic flux leakage data set.
  • the original sampled magnetic flux leakage data is used as multi-source data information, so Get a complete multi-source magnetic flux leakage data set.
  • the discovery module adopts a pipeline connection component discovery method based on selective search combined with a convolutional neural network to obtain the precise position of the weld seam, which specifically includes the following steps:
  • Step 2.1 Extract the pipeline magnetic flux leakage signal data: from the complete magnetic flux leakage data set, take the entire section of the pipeline magnetic flux leakage signal matrix D and divide it into n g segments of the pipeline magnetic flux leakage signal matrix in equal proportions Each split magnetic flux leakage signal matrix is composed of Data composition;
  • Step 2.2 Color map of magnetic flux leakage signal conversion: Set the upper limit of signal amplitude A top and the lower amplitude of signal amplitude A floor , and then the pipeline magnetic flux leakage signal matrix Convert to pipeline color graph matrix
  • Step 2.2.1 Set the upper limit of signal amplitude A top and the lower limit of signal amplitude A floor .
  • Step 2.2.2 According to the following formula, the pipeline magnetic flux leakage signal matrix Convert to a gray matrix between 0-255
  • d ij is the component of the magnetic flux leakage signal matrix D, and gray ij is the component of the gray matrix Gray;
  • Step 2.2.3 Set the gray matrix according to the following formula Convert to include 3D color matrix
  • r ij is the element of matrix R.
  • g ij is the element of matrix G.
  • b ij is the element of matrix B.
  • Step 2.3 Selective search: For each segment of the pipeline color map C k , use selective search to extract m c candidate regions
  • Step 2.3.3 Calculate the similarity sim ⁇ r ka , r kb ⁇ of all adjacent regions r ka , r kb according to the following formula.
  • Step 2.3.4 Repeat step 2.3.3 until the similarity of all adjacent regions is calculated, and update the similarity collection Sim according to the following formula.
  • Step 2.3.6 Repeat Step 2.3.5 until Sim is empty and get m c merged regions These areas are candidate areas.
  • Step 2.4 Convolutional neural network: use the convolutional neural network to distinguish the extracted candidate regions, and record the weld position Loc 1 , Loc 2 ,...,Loc w and its score Soc 1 determined by the convolutional neural network ,Soc 2 ,...,Soc w .
  • Step 2.5 Non-maximum suppression: According to the weld position Loc 1 ,Loc 2 ,...,Loc w and its score Soc 1 ,Soc 2 ,...,Soc w in step 2.4, suppression according to non-maximum The algorithm obtains the precise position of the weld seam L 1 , L 2 , ..., Lu .
  • the magnetic flux leakage signals of the entire pipeline are respectively u+1 segments, and one of the magnetic flux leakage signals is taken.
  • the discovery module uses an abnormal candidate based on the fusion of Lagrangian number multiplication framework and multi-source magnetic flux leakage data.
  • the area search and identification method to find the defective magnetic flux leakage signal includes the following steps:
  • Step 3.1 Establish a data reconstruction framework based on Lagrange multiplication
  • 1 represents the 1 norm of the matrix
  • * represents the kernel norm of the matrix
  • is the weight parameter.
  • Step 3.1.2 Convert the constrained optimization model to an unconstrained optimization model:
  • l represents the Lagrange function
  • ⁇ > represents the inner product of the matrix
  • is the penalty factor
  • Y is the Lagrangian multiplication matrix
  • Step 3.1.3 Iterative optimization.
  • the optimization model of matrix A is:
  • soft threshold operator For the convenience of calculation, the problem of kernel norm minimization can be solved by soft threshold operator.
  • - ⁇ ) + , where y + max(y ,0), the operator can be used as follows:
  • USV T is the singular value decomposition of matrix Z.
  • Correct U ⁇ R m ⁇ r , V ⁇ R r ⁇ n . r is the rank of the matrix.
  • Step 3.1.4 Set the iteration cutoff conditions, the cutoff conditions are:
  • S is the weight matrix, and the use of the S weight matrix can greatly reduce the iteration time. Thereby increasing the detection speed.
  • Step 3.2 Search for pipeline anomaly candidate regions based on multi-data fusion
  • Step 3.2.1 Perform anomaly region search on the single-axis data under the data reconstruction framework based on Lagrangian multiplication, and the three-axis anomaly regions are respectively ⁇ X , ⁇ Y , ⁇ Z.
  • Step 3.2.2 Establish a three-axis fusion optimization framework:
  • Step 3.2.3 Use non-maximum suppression algorithm to eliminate the overlap, while considering the diversity of candidate regions generated, the window forms that are close to each other are combined, and the largest periphery of the two is used as the periphery of the new window.
  • the merge criterion is: adjacent windows If the lateral distance of the center of the volume is less than the minimum value of the lateral length of the adjacent window.
  • Step 3.3 Recognition of pipeline magnetic flux leakage anomalies based on an evolvable model.
  • Step 3.3.1 Extract abnormal samples from the complete magnetic flux leakage data set and establish an abnormal recognition model based on Convolutional Neural Network (CNN).
  • CNN Convolutional Neural Network
  • Step 3.3.2 Add a new label to those samples with wrong recognition as a new classification, go to step 3.3.1, re-establish the abnormal recognition model, re-classify, and find the defective magnetic flux leakage signal.
  • the quantization module uses a random forest-based defect quantization method to obtain the defect size, which specifically includes the following steps:
  • Step 4.1 Collect data; detect the magnetic flux leakage signal of the defect and extract the characteristic of the magnetic flux leakage signal to obtain the characteristic value of the magnetic flux leakage signal, specifically:
  • the peak and valley positions and peak and valley values of the axial maximum channel magnetic flux leakage signal are found.
  • 10 relevant features of the waveform are extracted, respectively: single Peak-defect peak, single-peak maximum peak-valley difference, double-peak-valley width, double-peak defect signal left-peak-valley difference and right-peak-valley difference, double-peak defect signal peak-to-peak spacing, axial special point spacing, area characteristics, surface energy characteristics , Defect volume, defect body energy.
  • the valley width of the defect signal can reflect the axial distribution of the defect signal.
  • the combination of the peak-to-peak spacing of the defect signal and the peak-to-valley value can roughly determine the shape of the abnormal data curve, which is helpful for quantitative analysis of the length and depth of the defect.
  • the special point extraction method is: set the ratio m_RateA for the length, find the threshold value according to X+(YX)*m_RateA, where X is the average value of the valley, Y is the maximum peak value.
  • X is the average value of the valley
  • Y is the maximum peak value.
  • the two points closest to the threshold in the axial flux leakage signal of the maximum channel are the special points.
  • the distance between the special points is the key feature of the defect length.
  • Sa represents the defect waveform area
  • x(t) represents the defect signal data point
  • min[x(t)] represents the minimum valley value of the defect
  • N 1 represents the position of the left valley of the defect
  • N 2 represents the position of the right valley of the defect.
  • S e is the defect plane wave energy
  • Defect volume The volume of the defect is the sum of the defect areas in the range of the defect channel, which is expressed as:
  • V a represents the volume of the defect
  • n 1 represents the initial channel to a particular point position determining signal
  • n 2 represents the channel Zhouxiang Xin terminated by special position determination
  • S a (t) represents a single axial channel defect area.
  • Defect body energy is to sum the energy of the defect surface within the defect range, and its expression is:
  • V e represents the energy of the defect body
  • S e (t) represents the energy of the signal surface of a single axial defect.
  • Step 4.2 Take the characteristic value of the magnetic flux leakage signal of the defect as a sample; manually measure the defect size as a label, and the defect size includes the depth, width and length of the defect; manually select the initial training set and test set.
  • Step 4.3 Train the network; input the training set into the initial random forest network.
  • Step 4.4 Adjust the network; test the results of the random forest regression network through the test set, and adjust the network to obtain the final network through parameter adjustment.
  • Step 4.4.1 MFL signals from the original defect using a sample wherein M h ⁇ N h dimensions Bootstraping random sampling with replacement defect samples selected m e, m e ⁇ M h, a total of T c samples, generating T c training sets;
  • Step 4.4.2 For T c training sets, train T c regression tree models respectively.
  • Step 4.4.3 For a single regression tree model, select n e features in the magnetic flux defect signal feature set, where n e ⁇ N, and then according to the information gain ratio each time the split
  • H A (D) represents the entropy of feature A
  • g(D, A) represents its information gain.
  • the feature of selecting the information gain ratio to the maximum is split. Initially set the parameter max_features to None, which does not limit the number of features selected by the network;
  • Step 4.4.4 Each tree has been split in this way.
  • C(T) represents the model's prediction error of the defect size, that is, the degree of fit.
  • represents the model complexity, and ⁇ is used to adjust the complexity of the regression tree.
  • the prediction error of the loss function is taken as the percentage of the wall thickness at the 90% position of the POF.
  • Step 4.4.5 Model parameter optimization, using CVGridSearch grid search and K-fold cross-validation to find the optimal parameters.
  • the optimal parameters include the random forest frame parameters, the out-of-bag sample evaluation score e oob and the maximum number of iterations, and the tree model parameters Maximum feature number, maximum depth, minimum number of samples required for internal node subdivision and minimum number of samples for leaf nodes.
  • Step 4.4.6 The generated multiple decision trees form a random forest. For the regression problem network established by the defect feature samples, the average predicted value of multiple trees determines the final predicted defect size.
  • Step 4.5 Input the data to be tested in the test set into the random forest network adjusted according to step 4.4, and output the predicted defect size. If the data to be tested is the depth of the defect size, the output is the depth of the predicted defect size; The measured data is the width of the defect size, and the output is the width of the predicted defect size. If the data to be measured is the length of the defect size, the output is the length of the predicted defect size; where, according to the international oil pipeline POF standard, its predicted depth reflects In order to sort the values at the 80% position by the absolute value of the error, the formula is: among them Design depth and predicted depth respectively.
  • the solution module uses an improved pipeline solution based on the ASME B31G standard, imports the maintenance decision model, and outputs the evaluation results.
  • the specific steps include the following steps:
  • Step 5.1 Extract all defect length columns, depth columns and pipeline attribute parameters in the defect information from the complete magnetic flux leakage data set; the pipeline attribute parameters include minimum yield strength SMYS, minimum tensile strength SMTS, nominal outer diameter D d , wall thickness t a and the maximum allowable operating pressure MAOP;
  • Step 5.2 Calculate the flow stress value
  • SMYS is the minimum yield strength of the pipe, the unit is Mpa;
  • SMTS is the minimum tensile strength, the unit is Mpa.
  • d is the depth of the defect in mm
  • t a is the wall thickness of the pipe in mm
  • D d is the nominal outer diameter in mm.
  • Step 5.5 Calculate the maintenance index among them Among them, P is the maximum allowable design pressure; if the maintenance index ERF is less than 1, it means that the defect is acceptable, and if it is greater than or equal to 1, it is not acceptable. At this time, it should be repaired or replaced.
  • Step 5.6 Import the maintenance decision model, perform qualitative and quantitative analysis based on expert experience and life prediction model, then assess the severity of pipeline corrosion, formulate maintenance rules, and output the assessment results according to the maintenance rules, including: maintenance index and maintenance recommendations;
  • Rule 1 The maximum depth of wall thickness loss at the defect is greater than or equal to 80%, which is major corrosion. Maintenance suggestion: it should be repaired or replaced immediately.
  • Rule 2 The ERF at the defect is greater than or equal to 1, which is severe corrosion. Repair suggestion: it should be repaired immediately ,
  • Rule 3 The ERF value of the defect is greater than or equal to 0.95 and less than 1.0, which is general corrosion. Maintenance suggestions: you can observe for 1-3 months.
  • Rule 4 The maximum depth of the defect is greater than or equal to 20% and less than 40% is slight corrosion. Maintenance suggestion: It can be observed regularly without treatment.
  • the data complete set construction module of the present invention proposes a secondary baseline correction algorithm, which reduces the influence of abnormal data on the overall base value and improves the accuracy of baseline correction Sex.
  • the algorithm of adding logistic regression to each box of KNN is used to implement the interpolation of the missing data. This method is suitable for different types of data missing types, and it has strong anti-interference ability against the uncertainty of actual engineering data.
  • the selective search algorithm is introduced to generate candidate regions, which improves the speed and accuracy of generating candidate regions; the use of convolutional neural network calculations against candidate regions to classify missing regions , Increase the robustness of the weld detection algorithm to signal noise, and improve the classification accuracy;
  • the present invention uses multi-source magnetic flux leakage data for reconstruction, and analyzes the deviation between the reconstructed data and the source data to achieve anomaly detection.
  • a novel weight matrix is used in conditional calculation in. The experimental results show that this method has a good effect on abnormal detection.
  • the quantization module it is different from the general feature extraction method according to the characteristics of unstable magnetic leakage signal mutation.
  • the present invention proposes a method based on magnetic flux leakage signal waveform and statistical feature extraction, which enhances the model recognition effect; using the offshore oil pipeline POF standard to customize the random forest iteration loss function, which makes the algorithm highly adaptable in the field and quantifies the defects High precision.
  • the method of the invention has been applied to the inversion of actual engineering pipelines, and the effect of quantifying defect size is good.
  • the present invention is based on actual engineering applications. Compared with the original ASME B31G method, this method improves the calculation of flow stress to increase the failure pressure and reduce conservatism. B31G is too conservative, so frequent repairs will not be incurred due to ASME B31G being too conservative.
  • the present invention proposes an intelligent analysis system and method for detecting magnetic flux leakage data in pipelines. Compared with general magnetic flux leakage data analysis methods, the present invention proposes an intelligent analysis process for detecting magnetic flux leakage data in pipelines.
  • the sequence of the process includes: data complete set construction module, discovery module, quantization module and solution module. This process realizes the preprocessing of the magnetic flux leakage data detected in the original pipeline, the detection of the connected components and the abnormality detection.
  • the abnormality detection includes: defects, valve instrumentation and metal addition, the inversion of the defect size and the final maintenance decision.
  • FIG. 1 is a flowchart of an operation process of an intelligent analysis system for detecting magnetic flux leakage data in a pipeline according to an embodiment of the present invention
  • FIG. 2 is a block diagram of an intelligent analysis system for detecting magnetic flux leakage data in a pipeline according to an embodiment of the present invention
  • FIG. 3 is a flowchart of a method for constructing a complete data set based on sparse sampling in the time domain and KNN-softmax according to an embodiment of the present invention
  • FIG. 4 is a flowchart of a pipeline connection component discovery method based on a combination of selective search and a convolutional neural network according to an embodiment of the present invention
  • FIG. 5 is a schematic diagram of anomaly region search based on Lagrange multiplication according to an embodiment of the present invention.
  • FIG. 6 is a schematic diagram of an abnormal candidate region recommendation and recognition framework based on multi-source magnetic flux data fusion according to an embodiment of the present invention
  • FIG. 8 is a schematic diagram of data before and after baseline correction according to an embodiment of the present invention. wherein, FIG. 8(a) is a schematic diagram of data before baseline correction, and FIG. 8(b) is a schematic diagram of data after baseline correction;
  • FIG. 9 is a schematic diagram of a complete data set obtained before and after interpolation of the KNN-softmax algorithm according to an embodiment of the present invention; wherein, FIG. 9(a) is a schematic diagram of a complete data set without interpolation, and FIG. 9(b) is a KNN- A schematic diagram of the complete data set obtained after interpolation by the softmax algorithm;
  • FIG. 10 is a schematic diagram of a simulation result of a pipeline connection component discovery according to an embodiment of the present invention.
  • FIG. 11 is a schematic diagram of a simulation result of finding a defective magnetic flux leakage signal according to an embodiment of the present invention.
  • FIG. 12 is a histogram of defect quantification performance comparison of an embodiment of the present invention.
  • FIG. 13 is a scatter diagram of defect inversion residuals according to an embodiment of the present invention.
  • the invention provides an intelligent analysis software system for magnetic flux leakage data in pipelines, proposes an internal detection magnetic flux leakage data analysis system from the perspective of non-destructive testing and evaluation, and invents a kind of sparse sampling based on time-like domain and KNN-softmax from the perspective of intelligence Data complete set construction method, a pipeline connection component discovery method based on the combination of selective search and convolutional neural network, and an anomaly candidate region search and recognition method based on Lagrange multiplication framework and multi-source magnetic flux leakage data fusion 1.
  • a defect quantification method based on random forest and an improved pipeline solution based on ASME B31G standard. Realize the safe operation and maintenance of the pipeline.
  • FIG. 2 it is a block diagram of the magnetic flux leakage data intelligent analysis software system of the present invention.
  • the entire system includes four modules: a data complete set construction module, a discovery module, a quantization module, and a solution module.
  • the data complete set construction module implements data anomaly detection and reconstruction to construct a complete data set;
  • the discovery module includes component detection and anomaly detection, and its main goal is to identify defects;
  • the quantization module implements the mapping from signals to physical attributes to derive defects The length, width and depth of the solution;
  • the solution module is a comprehensive defect detection, size inversion results and knowledge model of pipeline properties and historical data, and finally gives a maintenance strategy.
  • the intelligent analysis system for magnetic flux leakage data in pipelines proposed by the present invention is specifically implemented as follows:
  • the data complete set construction module adopts a data complete set construction method based on sparse sampling in the time domain and KNN-softmax to obtain a complete magnetic flux leakage data set.
  • FIG. 3 it is a flowchart of data preprocessing based on sparse sampling in the time domain and KNN-softmax.
  • the baseline is corrected twice, then the data is modeled in a time-like manner and the data is anomaly identified.
  • the KNN-softmax regression model is used for data interpolation, and finally a complete magnetic flux leakage data set is constructed.
  • the specific steps of data preprocessing based on sparse sampling in the time domain and KNN-softmax are as follows:
  • Step 1.1 Collect raw magnetic flux leakage detection data directly from the subsea pipeline magnetic flux leakage detector, and perform secondary baseline correction on the data.
  • the original sampled magnetic flux leakage data is used as multi-source data information, including: axial data, radial Data, weekly data, alpha data;.
  • Step 1.1.1 Perform a baseline correction on the original magnetic flux leakage detection data, expressed as:
  • k c is the number of mileage counting points; Is the original value of the position of the j a channel at the i a mileage counting point; Is the corrected value of the j a channel at the i a mileage counting point; s is the median of all channels, and n a is the number of magnetic flux leakage internal detector channels.
  • Step 1.1.2 Remove the excess value ⁇ T a from the data, and assign the position value of the excess value to the median s of all channels, expressed as:
  • Step 1.1.3 Perform secondary baseline correction on the data beyond the limit:
  • k c is the number of mileage counting points; It is the primary correction value of the j a channel at the i a mileage counting point; It is the value after the second correction of the j a channel at the position of the i a mileage counting point; s′ is the median of all the channels after one correction.
  • Step 1.2 Perform quasi-time-domain sparse sampling anomaly detection processing on the data after the second baseline correction
  • Step 1.2.1 Perform anomalous signal time-domain modeling on the data after the secondary baseline correction, that is, correspond to the sampling points as time information;
  • Step 1.2.1.1 mathematical modeling of the abnormal part, the modeling results are as follows:
  • p(t)′ represents the pipeline magnetic flux leakage detection voltage bulge compensation signal
  • f represents the signal sampling rate
  • t represents the sampling time
  • t 1 and t 2 represent the sampling time interval
  • a represents the power pipeline
  • n represents the system fluctuation amplitude Coefficient, f(t)' frequency of voltage waveform change
  • I c represents the sampling points within a given interval
  • It represents the average value of the voltage data of the pipeline system collected in this interval
  • ⁇ f 0 represents the degree of fluctuation of the voltage signal of the magnetic flux leakage data.
  • Step 1.2.1.3 Calculate the amount of voltage state change The formula is as follows:
  • Step 1.2.2 Judgment of abnormal signal: if Then the data at this time is regarded as an abnormality caused by external disturbance, which is an abnormal part.
  • test sample features T′ (X′ 1 ,X′ 2 ,...,X′ 7 ,X′ 8 ), also extract 8 features, namely the left valley value, right valley value, valley width, peak value of the data , Left peak valley difference, right peak valley difference, differential left peak and differential right peak.
  • T′′ (X′′ 1 ,X′′ 2 ,...,X′′ 7 ,X′′ 8 ), and also extract 8 features, namely the left valley value, right valley value, valley width of the data , Peak value, left peak valley difference, right peak valley difference, differential left peak and differential right peak.
  • Step 1.3 Perform the missing interpolation process based on KNN-logistic regression on the magnetic flux leakage data of the submarine pipeline.
  • Step 1.3.1 Training and testing KNN and softmax regression models.
  • Step 1.3.1.1 divides the feature sample data T into two parts, a part of the feature sample data X Train is used to train the KNN model, and another part of the feature sample data T Test is used to test the KNN model.
  • Step 1.3.1.2 Input X Train into the KNN model, set the initial value of K to 5, and train the KNN model.
  • Step 1.3.1.4 Feature sample data classified into each category
  • the data set corresponding to the feature sample is Correct with Normalized separately with Expressed as:
  • Step 1.3.1.5 Add a softmax regression model at each type of node, the assumption function is as follows:
  • x is the sample input value and y is the sample output value;
  • is the training model parameter;
  • k f is the vector dimension;
  • i e is the i e category of the classification;
  • p(y i e
  • x) represents the category i e estimated probability value.
  • Step 1.3.1.6 Set the training sample set at each node Input into softmax regression model to get the output value after interpolation
  • the loss function J( ⁇ ) is:
  • x is the sample input value
  • y is the sample output value
  • is the training model parameter
  • k f is the vector dimension
  • i e is the i e category of the classification
  • j e is the j e sample input in the classification
  • M d is the number of samples
  • 1 ⁇ is an indicative function, if the braces are true, the expression value is 1.
  • Step 1.3.3 Input the characteristics and data set of the data to be interpolated into the trained model to implement the interpolation of the missing data to obtain a complete magnetic flux leakage data set.
  • the original sampled magnetic flux leakage data is used as multi-source data information, so Get a complete multi-source magnetic flux leakage data set.
  • Figure 8(a) is a schematic diagram of data before baseline correction. From Figure 8(a), it can be seen that the base value of the data without baseline correction is very different. After adding the offset, the data of each channel The distribution is uneven; Figure 8(b) is a schematic diagram of the data after baseline correction. It can be seen from Figure 8(b) that the base values of the data after baseline correction are equal. After adding the offset, the data distribution of each channel is even, which reduces the subsequent Data processing errors.
  • Figure 9(a) is a schematic diagram of a data set that contains missing data without interpolation
  • Figure 9(b) is a schematic diagram of a complete data set obtained after interpolation by the KNN-softmax algorithm, as can be seen from Figure 9(b) No matter at the defect position or at the smooth position, the algorithm can complete the interpolation of the missing data.
  • the discovery module adopts a pipeline connection component discovery method based on selective search combined with a convolutional neural network to obtain the precise position of the weld seam, which specifically includes the following steps:
  • Step 2.1 Extract the pipeline magnetic flux leakage signal data: from the complete magnetic flux leakage data set, take the entire section of the pipeline magnetic flux leakage signal matrix D and divide it into n g segments of the pipeline magnetic flux leakage signal matrix in equal proportions
  • Each split magnetic flux leakage signal matrix is composed of It consists of data; the magnetic flux leakage signal matrix D of the entire pipeline of size M ⁇ N collected by the magnetic flux leakage detector.
  • the magnetic flux leakage signal matrix D 1 ,D 2 ,...,D 10 of the size M 1 ⁇ N, M 2 ⁇ N,...M 10 ⁇ N are equally divided.
  • M 1 +M 2 +...+M 10 M.
  • Step 2.2 Color map of magnetic flux leakage signal conversion: Set the upper limit of signal amplitude A top and the lower amplitude of signal amplitude A floor , and then the pipeline magnetic flux leakage signal matrix Convert to pipeline color graph matrix
  • Step 2.2.1 Set the upper limit of signal amplitude A top and the lower limit of signal amplitude A floor .
  • Step 2.2.2 According to the following formula, the pipeline magnetic flux leakage signal matrix Convert to a gray matrix between 0-255
  • d ij is the component of the magnetic flux leakage signal matrix D, and gray ij is the component of the gray matrix Gray;
  • Step 2.2.3 Set the gray matrix according to the following formula Convert to include 3D color matrix
  • c 255
  • r ij is the element of matrix R.
  • g ij is the element of matrix G.
  • b ij is the element of matrix B.
  • Step 2.3 Selective search: For each segment of the pipeline color map C k , use selective search to extract m c candidate regions
  • Step 2.3.3 Calculate the similarity sim ⁇ r ka , r kb ⁇ of all adjacent regions r ka , r kb according to the following formula.
  • Step 2.3.4 Repeat step 2.3.3 until the similarity of all adjacent regions is calculated, and update the similarity collection Sim according to the following formula.
  • Step 2.3.6 Repeat step 2.3.5 until Sim is empty. Get m c merged regions These areas are candidate areas.
  • Step 2.4 Convolutional neural network: candidate region identification:
  • Step 2.4.1 To construct a 72 ⁇ 72 convolutional neural network, the middle layer of the convolutional neural network includes 4 convolutional layers, 4 downsampling layers, and 1 fully connected layer. Among them, each convolutional layer is followed by a down-sampling layer used to obtain the local weighted average as the second feature extraction.
  • Step 2.4.2 Extract P N l ⁇ N l weld seam color maps from the historical data as samples of the convolutional neural network. 80% of the random samples are used as training samples, and the remaining 20% are used as test samples.
  • Step 2.4.3 Train the network repeatedly 500 times, of which the highest test success rate is the final network Net.
  • Step 2.4.4 Input the candidate regions r k1, r k2, ... r km into the trained convolutional neural network for discrimination. For the area judged to be a weld, record its position Loc and network score Soc. Finally, w positions Loc 1 , Loc 2 ,...,Loc w and their scores Soc 1 ,Soc 2 ,...,Soc w are obtained .
  • Step 2.5 Non-maximum suppression: According to the above weld position Loc 1 , Loc 2 ,...,Loc w and its score Soc 1 ,Soc 2 ,...,Soc w ,according to the non-maximum suppression algorithm, Obtain the precise position of the weld seam L 1 , L 2 ,..., Lu .
  • step 2 The simulation result of step 2 is shown in FIG. 10: the pipeline component discovery method proposed by the present invention: the accuracy rate is 95.3% and the recall rate is 97.94%, compared with the traditional method using a threshold: the accuracy rate is 91.5% and the recall rate is 95.51 %. It can be seen that the method proposed by the present invention has better performance.
  • the magnetic flux leakage signals of the entire section of the pipeline are respectively u+1 segments, and one of the magnetic flux leakage signals is taken.
  • the discovery module adopts a fusion based on Lagrange number multiplication framework and multi-source magnetic flux leakage data.
  • the method of searching and identifying abnormal candidate regions to find the defective magnetic flux leakage signal specifically includes the following steps:
  • Step 3.1 Establish a data reconstruction framework based on Lagrange multiplication
  • the constrained optimization model is transformed into an unconstrained optimization model, and finally the reconstruction matrix is obtained by alternating iterations, thus The error matrix of the reconstructed matrix and the observation matrix is obtained, and the abnormal region is obtained through an appropriate threshold, and finally the region is regularized.
  • Specific steps are as follows:
  • Step 3.1.2 Convert the constrained optimization model to an unconstrained optimization model
  • Step 3.1.3 Iterative optimization.
  • the optimization model of matrix A is:
  • soft threshold operator For the convenience of calculation, the problem of kernel norm minimization can be solved by soft threshold operator.
  • - ⁇ ) + , where y + max(y ,0), the operator can be used as follows:
  • Step 3.1.4 Set iteration cut-off conditions.
  • the cut-off conditions are: Among them, S is a weight matrix, and the utility of the S weight matrix can greatly reduce the iteration time. Thereby increasing the detection speed.
  • the setting of the S matrix of the present invention is as follows:
  • Step 3.2 Search for pipeline anomaly candidate regions based on multi-data fusion
  • FIG. 6 it is a framework for recommending and identifying anomalous candidate regions based on the fusion of multi-source magnetic flux leakage data in the present invention.
  • anomalous regions are recommended for multi-source data respectively, and then through the region optimization framework, from the boundary It is optimized from the perspective of the region, and finally the abnormal candidate region is obtained, input to the recognition model, and finally classified. Specific steps are as follows:
  • Step 3.2.1 Respectively search for abnormal areas under unidirectional data under the framework of data reconstruction based on Lagrange multiplication. It is found that the three-axis abnormal areas are ⁇ X , ⁇ Y , ⁇ Z.
  • Step 3.2.2 Establish a three-axis fusion optimization framework
  • Step 3.2.3 Use non-maximum suppression algorithm to eliminate the overlap, while considering the diversity of candidate regions generated, the window forms that are close to each other are merged, the largest periphery of the two is used as the periphery of the new window, and the merge criterion is: adjacent windows If the lateral distance of the center of the volume is less than the minimum value of the lateral length of the adjacent window.
  • Step 3.3 Recognition of pipeline magnetic flux leakage anomalies based on an evolvable model.
  • Step 3.3.1 Extract abnormal samples from the complete magnetic flux leakage data set and establish an abnormal recognition model based on Convolutional Neural Network (CNN).
  • CNN Convolutional Neural Network
  • Step 3.3.2 For those samples with wrong recognition, add new labels and re-enter the model for training. With the increasing number of data changes, the recognition model continues to evolve.
  • step 3 The simulation result in step 3 is shown in FIG. 11: the pipeline abnormality discovery method proposed by the present invention: the accuracy rate is 95.73% and the recall rate is 93.86%; the uniaxial data abnormality discovery accuracy rate is 93.07% and the recall rate is 89.73%; In the traditional method based on feature extraction: the accuracy rate is 88.98% and the recall rate is 81.93%. It can be seen that this method has better performance.
  • the quantization module uses a random forest-based defect quantization method to obtain the defect size, which specifically includes the following steps:
  • Step 4.1 Collect data; detect the magnetic flux leakage signal of the defect and extract the characteristic of the magnetic flux leakage signal to obtain the characteristic value of the magnetic flux leakage signal, specifically:
  • the peak and valley positions and peak and valley values of the axial maximum channel magnetic flux leakage signal are found.
  • 10 relevant features of the waveform are extracted, respectively: single Peak-defect peak, single-peak maximum peak-valley difference, double-peak-valley width, double-peak defect signal left-peak-valley difference and right-peak-valley difference, double-peak defect signal peak-to-peak spacing, axial special point spacing, area characteristics, surface energy characteristics , Defect volume, defect body energy.
  • the valley width of the defect signal can reflect the axial distribution of the defect signal.
  • the combination of the peak-to-peak spacing of the defect signal and the peak-to-valley value can roughly determine the shape of the abnormal data curve, which is helpful for quantitative analysis of the length and depth of the defect.
  • the special point extraction method is: set the ratio m_RateA for the length, find the threshold value according to X+(YX)*m_RateA, where X is the average value of the valley, Y is the maximum peak value.
  • X is the average value of the valley
  • Y is the maximum peak value.
  • the two points closest to the threshold in the axial flux leakage signal of the maximum channel are the special points.
  • the distance between the special points is the key feature of the defect length.
  • Sa represents the defect waveform area
  • x(t) represents the defect signal data point
  • min[x(t)] represents the minimum valley value of the defect
  • N 1 represents the position of the left valley of the defect
  • N 2 represents the position of the right valley of the defect.
  • S e is the defect plane wave energy
  • Defect volume The volume of the defect is the sum of the defect areas in the range of the defect channel, which is expressed as:
  • V a represents the volume of the defect
  • n 1 represents the initial channel to a particular point position determining signal
  • n 2 represents the channel Zhouxiang Xin terminated by special position determination
  • S a (t) represents a single axial channel defect area.
  • Defect body energy is to sum the energy of the defect surface within the defect range, and its expression is:
  • V e represents the energy of the defect body
  • S e (t) represents the energy of the signal surface of a single axial defect.
  • Step 4.2 Take the characteristic value of the magnetic flux leakage signal of the defect as a sample; manually measure the defect size as a label, and the defect size includes the depth, width and length of the defect; manually select the initial training set and test set.
  • Step 4.3 Train the network; input the training set into the initial random forest network.
  • Step 4.4 Adjust the network; check the results of the random forest regression network through the test set, and adjust the network to obtain the final network by adjusting the parameters.
  • Specific practice: Initially set the parameter n f n f /3 and set the maximum feature number max_features to None.
  • Step 4.4.1 MFL signals from the original defect using a sample wherein M h ⁇ N h dimensions Bootstraping random sampling with replacement defect samples selected m e, m e ⁇ M h, a total of T c samples, generating T c training sets;
  • Step 4.4.2 For T c training sets, train T c regression tree models respectively.
  • Step 4.4.3 For a single regression tree model, select n e features in the magnetic flux defect signal feature set, where n e ⁇ N, and then according to the information gain ratio each time the split
  • H A (D) represents the entropy of feature A
  • g(D, A) represents its information gain.
  • the feature of selecting the information gain ratio to the maximum is split. Initially set the parameter max_features to None, which does not limit the number of features selected by the network;
  • Step 4.4.4 Each tree has been split in this way.
  • C(T) represents the model's prediction error of the defect size, that is, the degree of fit.
  • represents the model complexity, and ⁇ is used to adjust the complexity of the regression tree.
  • the prediction error of the loss function is taken as the percentage of the wall thickness at the 90% position of the POF. Initially, the maximum tree depth max_depth is set to 5.
  • Step 4.4.5 Model parameter optimization, using CVGridSearch grid search and K-fold cross-validation to find the optimal parameters. It includes the random forest frame parameters, the out-of-bag sample evaluation score e oob and the maximum number of iterations, as well as the tree model parameters maximum feature number max_features, maximum depth max_depth, minimum number of samples required for internal node subdivision and minimum number of leaves node samples.
  • Step 4.4.6 The generated multiple decision trees form a random forest. For the regression problem network established from defect feature samples, the average predicted value of multiple trees determines the final predicted defect size information.
  • Step 4.5 Input the data to be tested into the random forest network adjusted according to step 4.4, and output the predicted defect size. If the data to be tested is the depth of the defect size, the output is the depth of the predicted defect size; if the data to be tested Is the width in the defect size, the output is the width of the predicted defect size, and if the data to be measured is the length of the defect size, the output is the length of the predicted defect size; where, according to the international oil pipeline POF standard, the predicted depth reflects the The absolute value of the error ranks the value at the 80th percentile, the formula is: The formula is: among them Design depth and predicted depth respectively. Taking the loss function of iteration n p generation no longer reduced as a parameter optimization termination condition, the network finally outputs the maximum number of iterations n_estimators is 172.
  • step 4 compares the performance of the present invention with the traditional defect inversion algorithm, as shown in Table 1:
  • Table 1 and FIG. 12 reflect the algorithm of the present invention according to the international offshore oil pipeline POF standard.
  • the absolute error of the length of the defect is within 10mm
  • the width is within 15mm
  • the absolute error of the depth accounts for the wall thickness (9.5mm)
  • the percentage is within 10.
  • the accuracy is higher and the variance is smaller, which meets the accuracy requirements of industrial defect inversion.
  • Experimental results verify that the algorithm has good generalization ability and robustness.
  • the solution module uses an improved pipeline solution based on the ASME B31G standard, imports the maintenance decision model, and outputs the evaluation results, as shown in Figure 7, specifically including the following steps:
  • Step 5.1 Extract all defect length columns, depth columns and pipeline attribute parameters in the defect information from the complete magnetic flux leakage data set; the pipeline attribute parameters include minimum yield strength SMYS, minimum tensile strength SMTS, nominal outer diameter D d , wall thickness t a and the maximum allowable operating pressure MAOP.
  • Step 5.2 Calculate the flow stress value
  • SMYS is the minimum yield strength of the pipe, the unit is Mpa;
  • SMTS is the minimum tensile strength, the unit is Mpa.
  • d is the depth of the defect in mm
  • t a is the wall thickness of the pipe in mm
  • D d is the nominal outer diameter in mm.
  • Step 5.5 Calculate the maintenance index among them Among them, P is the maximum allowable design pressure; if the maintenance index ERF is less than 1, it means that the defect is acceptable, and if it is greater than or equal to 1, it is not acceptable. At this time, it should be repaired or replaced.
  • Step 5.6 Import the maintenance decision model, perform qualitative and quantitative analysis based on expert experience and life prediction model, then assess the severity of pipeline corrosion, formulate maintenance rules, and output the assessment results according to the maintenance rules, including: maintenance index and maintenance recommendations;
  • Rule 1 The maximum depth of wall thickness loss at the defect is greater than or equal to 80%, which is major corrosion. Maintenance suggestion: it should be repaired or replaced immediately.
  • Rule 2 The ERF at the defect is greater than or equal to 1, which is severe corrosion. Repair suggestion: it should be repaired immediately ,
  • Rule 3 The ERF value of the defect is greater than or equal to 0.95 and less than 1.0, which is general corrosion. Maintenance suggestions: you can observe for 1-3 months.
  • Rule 4 The maximum depth of the defect is greater than or equal to 20% and less than 40% is slight corrosion. Maintenance suggestion: It can be observed regularly without treatment.
  • step 5 The simulation results of step 5 are shown in the curve shown in Fig. 14 after the basic parameter information of the pipeline is brought into ASME B31G 1991 and the improved standard.
  • the improved formula reduces the conservativeness by changing the value of the flow stress.
  • the ASME B31G 1991 standard is too conservative, and the actual inspection process often increases maintenance or tube replacement, resulting in economics. The above waste is applicable to the old pipeline. Due to the reduced conservativeness of the improved formula, the cost due to frequent maintenance is reduced.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Chemical & Material Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Computation (AREA)
  • Chemical Kinetics & Catalysis (AREA)
  • Electrochemistry (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Mechanical Engineering (AREA)
  • Investigating Or Analyzing Materials By The Use Of Magnetic Means (AREA)

Abstract

一种管道内检测漏磁数据智能分析系统,包括:数据完备集构建模块、发现模块、量化模块以及解决方案模块;在数据完备集构建模块中采用基于类时域稀疏采样和KNN-softmax的数据完备集构建方法,得到完备漏磁数据集;在发现模块中采用基于选择性搜索与卷积神经网络相结合的管道连接组件发现方法,得到焊缝的精确位置;在发现模型中采用基于拉格朗日数乘框架和多源漏磁数据融合的异常候选区域搜索与识别方法,找出有缺陷的漏磁信号;在量化模块中采用基于随机森林的缺陷量化方法,得到缺陷尺寸;在解决方案模块中采用一种基于ASME B31G标准改进的管道解决方案,输出评估结果。这种管道内检测漏磁数据智能分析系统实现了预处理,连接组件检测和异常检测,缺陷尺寸反演以及最终维修决策。

Description

一种管道内检测漏磁数据智能分析系统 技术领域
本发明属于管道检测技术领域,具体特别涉及一种管道内检测漏磁数据智能分析系统。
背景技术
管道运输作为一种连续、经济、高效、绿色的运输手段而被广泛应用。国标规定的管道设计寿命为20年,随着运行时间的增长,因管道材质问题、施工、腐蚀和外力作用造成的损伤,会使管道状况逐年恶化,潜在危险激增。一旦发生泄漏,不但会造成大气污染,也极易引发剧烈爆炸。因此,为了确保能源运输和生态环境安全,必须对管道定期进行安检和维护。
无损检测(Non-destructive testing,NDT)作为管道安全维护的一种重要手被广泛应用。目前,用于管道检测的主要方法包漏磁检测、涡流检测和超声检测。其中,漏磁检测被广泛应用于接近90%的在役管道中,是国外发达国家技术相对成熟、应用最广泛的一种针对铁磁性材料的缺陷检测技术。目前针对漏磁数据的分析研究很多,包括数据预处理、检测、尺寸反演、数据呈现等。然而,现有对漏磁数据分析的研究过于注重局部点的发展,缺乏对数据分析的系统观,绝大多数理论方法和应用技术缺乏通用性和可移植性,并未将智能化技术与漏磁数据分析有效的结合在一起,很难形成一套切实可行的、可广泛移植的数据分析系统。
本发明从面和体的角度,发明一种管道内检测漏磁数据分析软件系统;从人工智能角度发明数据分析方法,发明一种基于类时域稀疏采样和KNN-softmax的数据预处理方法、发明一种基于选择性搜索与卷积神经网络相结合的管道连接组件检测方法、发明一种基于拉格朗日数乘框架和多源漏磁数据融合的异常候选区域搜索与识别方法,发明一种基于随机森林的缺陷反演方法,发明一种基于ASME B31G标准改进的管道缺陷评估方法。
发明内容
基于以上技术问题,本发明提供一种管道内检测漏磁数据智能分析系统,其中,一种管道内检测漏磁数据智能分析系统,包括:数据完备集构建模块、发现模块、量化模块以及解决方案模块;
原始采样漏磁数据与数据完备集构建模块相连接,数据完备集构建模块通过完备漏磁数据集与发现模块相连接,发现模块与量化模块相连接,量化模块与解决方案模块相连接;
所述数据完备集构建模块用于对原始漏磁内检测数据进行数据缺失重构以及降噪操作,采用基于类时域稀疏采样和KNN-softmax的数据完备集构建方法,构建完备漏磁数据集;
所述数据完备集构建模块中将原始采样漏磁数据作为多源数据信息,具体包括:轴向数据、径向数据、周向数据、α向数据;
所述发现模块是进行缺陷检测,其过程包括组件检测和异常检测,其中,组件检测完成对管道连接组件焊缝和法兰的检测;所述发现模块,采用基于选择性搜索与卷积神经网络相结合的管道连接组件发现方法,得到焊缝的精确位置;根据焊缝的精确位置,将整段管道漏磁信号分别u+1段,取其中一段漏磁信号,采用基于拉格朗日数乘框架和多源漏磁数据融合的异常候选区域搜索与识别方法,找出有缺陷的漏磁信号;
异常检测包括:缺陷、阀门仪表和金属增加检测,并最终得到缺陷信号;
所述量化模块完成缺陷信号到物理特性的映射,采用基于随机森林的缺陷量化方法,最终给出缺陷的尺寸,即长度、宽度和深度;
所述解决方案模块从完备漏磁数据集中提取缺陷信息中所有的缺陷长度列、深度列以及管道属性参数,采用基于ASME B31G标准改进的管道解决方案,通过维修决策模型,最终给出评估结果,评估结果包括单独缺陷位置的维修指数、维修建议。
其中管道属性参数包括最小屈服强度SMYS、最小拉伸强度SMTS、公称外径D d、壁厚t a和最大允许操作压力MAOP;
所述数据完备集构建模块,采用基于类时域稀疏采样和KNN-softmax的数据完备集构建方法,得到完备漏磁数据集,具体包括如下步骤:
步骤1.1:从海底管道漏磁检测仪中直接采集原始漏磁检测数据,并且对数据进行二次基线校正,其中,原始采样漏磁数据作为多源数据信息,具体包括:轴向数据、径向数据、周向数据、α向数据;
步骤1.1.1:对原始漏磁检测数据进行一次基线校正,表示为:
Figure PCTCN2019074907-appb-000001
其中,k c为里程计数点数量;
Figure PCTCN2019074907-appb-000002
为第j a通道在第i a里程计数点位置的原始值;
Figure PCTCN2019074907-appb-000003
为第j a通道在第i a里程计数点位置的校正后的值;s为所有通道的中值,n a为漏磁内检测器通道数。
步骤1.1.2:去除数据中的超限值±T a,将超限值的位置值赋予所有通道的中值s,表示为:
Figure PCTCN2019074907-appb-000004
步骤1.1.3:对去除超限值的数据进行二次基线校正:
Figure PCTCN2019074907-appb-000005
其中,k c为里程计数点数量;
Figure PCTCN2019074907-appb-000006
为第j a通道在第i a里程计数点位置的一次校正值;
Figure PCTCN2019074907-appb-000007
为 第j a通道在第i a里程计数点位置的二次校正后的值;s′为一次校正后所有通道的中值。
步骤1.2:对二次基线校正后的数据进行类时域稀疏采样异常检测处理;
步骤1.2.1:对二次基线校正后的数据进行异常信号类时域建模,即将采样点对应为时间信息;
步骤1.2.1.1:对异常部分进行数学建模,建模结果所示为:
f(t)'=p(t)'*sin(2πnft)
其中,
Figure PCTCN2019074907-appb-000008
其中,p(t)′表示管道漏磁检测电压凸起补偿信号,f表示信号采样率,t表示采样时间,t 1,t 2表示采样时间间隔,a表示电力管道,n为系统波动幅值系数,f(t)'电压波形变化频率;
步骤1.2.1.2:以区间为采集单位设定漏磁检测异常数据变化量,以k e个采集数据为一个区间,将各区间内采集的管道系统电压数据的方差视为数据变化量,判定漏磁数据电压信号波动程度,具体计算方法为:
Figure PCTCN2019074907-appb-000009
即:
Figure PCTCN2019074907-appb-000010
其中,
Figure PCTCN2019074907-appb-000011
表示给定区间内的第i c个取样点,
Figure PCTCN2019074907-appb-000012
表示该区间内采集管道系统电压数据的平均值,Δf 0表示漏磁数据电压信号波动程度。
步骤1.2.1.3:计算电压状态变化量
Figure PCTCN2019074907-appb-000013
公式如下:
Figure PCTCN2019074907-appb-000014
步骤1.2.2:异常信号判别:若
Figure PCTCN2019074907-appb-000015
则此时的数据认为是外界扰动产生的异常,即为异常部分。
步骤1.2.3:人工提取训练样本特征
Figure PCTCN2019074907-appb-000016
人工提取测试样本特征
Figure PCTCN2019074907-appb-000017
人工提取待插补数据特征
Figure PCTCN2019074907-appb-000018
i b,j b,k d为特征数目。
步骤1.3:对海底管道漏磁数据进行基于KNN-逻辑回归的缺失插补处理。
步骤1.3.1:训练测试KNN、softmax回归模型。
步骤1.3.1.1将特征样本数据T分为两部分,一部分特征样本数据X Train用于训练KNN模型,另一部分特征样本数据T Test用于测试KNN模型。
步骤1.3.1.2:将X Train输入到KNN模型中,确定一个K值,训练KNN模型。
步骤1.3.1.3:将T Test输入到训练完成的KNN模型中进行分类,计算判别错误率,若错误率小于阈值,则采用V折交叉法,改变训练和测试样本继续训练,否则令K=K+1继续训练模型,当K大于阈值M时停止训练。
步骤1.3.1.4:分到每类中的特征样本数据
Figure PCTCN2019074907-appb-000019
特征样本对应的数据集为
Figure PCTCN2019074907-appb-000020
Figure PCTCN2019074907-appb-000021
Figure PCTCN2019074907-appb-000022
分别进行归一化处理得到
Figure PCTCN2019074907-appb-000023
Figure PCTCN2019074907-appb-000024
表示为:
Figure PCTCN2019074907-appb-000025
Figure PCTCN2019074907-appb-000026
是特征样本数据的平均值;
Figure PCTCN2019074907-appb-000027
特征样本对应数据的平均值;
步骤1.3.1.5:在每一类的节点处添加一个softmax回归模型,其假设函数为如式表示为:
Figure PCTCN2019074907-appb-000028
其中,x为样本输入值,y为样本输出值;θ为训练模型参数;k f为向量维数;i e为分类的第i e个类别;p(y=i e|x)表示对类别i e估算的概率值。
步骤1.3.1.6:将每个节点处中的训练样本集
Figure PCTCN2019074907-appb-000029
输入到softmax回归模型中,得到插补之后的输出值
Figure PCTCN2019074907-appb-000030
其损失函数J(θ)为:
Figure PCTCN2019074907-appb-000031
其中,x为样本输入值;y为样本输出值;θ为训练模型参数;k f为向量维数;i e为分类的第i e个类别;j e为该分类中第j e个样本输入;m d为样本个数;1{·}为示性函数,若大括号 里为真值,表达式值为1。
步骤1.3.2:计算预测结果的损失函数,设定阈值P。若J(θ)>P,回到步骤1.3.2.2,令K=K+1继续训练模型。直至J(θ)≤P,当K大于阈值M时停止训练,输出插补之后的输出值y (i)'。
步骤1.3.3:将待插补数据特征及数据集输入到训练完成的模型中,实现对缺失数据的插补,得到完备漏磁数据集,因原始采样漏磁数据作为多源数据信息,故得到完备多源漏磁数据集。
所述发现模块,采用基于选择性搜索与卷积神经网络相结合的管道连接组件发现方法,得到焊缝的精确位置,具体包括如下步骤:
步骤2.1:提取管道漏磁信号数据:从完备漏磁数据集中,取将整段管道漏磁信号矩阵D,等比例分割成n g段管道漏磁信号矩阵
Figure PCTCN2019074907-appb-000032
每一个分割后的漏磁信号矩阵都由
Figure PCTCN2019074907-appb-000033
个数据组成;
步骤2.2:漏磁信号转换色彩图:设置信号幅值上限A top和信号幅值下线A floor,并据此将管道漏磁信号矩阵
Figure PCTCN2019074907-appb-000034
转换为管道彩色图矩阵
Figure PCTCN2019074907-appb-000035
步骤2.2.1:设置信号幅值上限A top和信号幅值下线A floor
步骤2.2.2:根据如下公式将管道漏磁信号矩阵
Figure PCTCN2019074907-appb-000036
转换为0-255之间的灰度矩阵
Figure PCTCN2019074907-appb-000037
Figure PCTCN2019074907-appb-000038
其中
Figure PCTCN2019074907-appb-000039
d ij为漏磁信号矩阵D组成元素,gray ij为灰度矩阵Gray组成元素;
步骤2.2.3:根据如下公式将灰度矩阵
Figure PCTCN2019074907-appb-000040
转换为包含
Figure PCTCN2019074907-appb-000041
Figure PCTCN2019074907-appb-000042
的三维彩色矩阵
Figure PCTCN2019074907-appb-000043
Figure PCTCN2019074907-appb-000044
其中,r ij为矩阵R组成元素。g ij为矩阵G组成元素。b ij为矩阵B组成元素。
步骤2.3:选择性搜索:对每一段管道彩色图C k,利用选择性搜索,提取出m c个候选区域
Figure PCTCN2019074907-appb-000045
步骤2.3.1:每一段管道彩色图C k,利用切分方法得到候选的区域集合R k={r k1,r k2,...,r kw}。
步骤2.3.2:初始化相似度集合Sim=φ。
步骤2.3.3:根据如下公式计算所有相临区域r ka,r kb的相似度sim{r ka,r kb}。
Figure PCTCN2019074907-appb-000046
步骤2.3.4:重复步骤2.3.3直到所有相邻区域的相似度都计算出来,根据如下公式更新相似度合集Sim。
Sim=Sim∪sim(r ka,r kb)
步骤2.3.5:从Sim找到最大的相似度sim{r kc,r kd}=max(Sim),并根据此得到合并区域:
r ke=r kc∪r kd
从Sim中剔除出sim{r kc,r kd}。
步骤2.3.6:重复步骤2.3.5,直到Sim为空,得到m c个合并后区域
Figure PCTCN2019074907-appb-000047
这些区域就是候选区域。
步骤2.4:卷积神经网络:利用卷积神经网络对提取出的候选区域进行判别,模记录卷积神经网络判断的焊缝位置Loc 1,Loc 2,...,Loc w及其评分Soc 1,Soc 2,...,Soc w
步骤2.5:非极大值抑制:根据步骤2.4中焊缝位置Loc 1,Loc 2,...,Loc w及其评分Soc 1,Soc 2,...,Soc w,根据非极大值抑制算法得到焊缝的精确位置L 1,L 2,...,L u
根据焊缝的精确位置,将整段管道漏磁信号分别u+1段,取其中一段漏磁信号,所述发现模块,采用基于拉格朗日数乘框架和多源漏磁数据融合的异常候选区域搜索与识别方法,找出有缺陷的漏磁信号,具体包括如下步骤:
步骤3.1:建立基于拉格朗日数乘的数据重构框架;
步骤3.1.1:建立数据重构模型
Figure PCTCN2019074907-appb-000048
subject to P=A+E,其中P是观测矩阵,E是误差矩阵,A是重构后的低秩矩阵。||·|| 1表示矩阵的1范数,||·|| *表示矩阵的核范数,λ是权重参数。
步骤3.1.2:将有约束优化化模型为无约束优化模型:
Figure PCTCN2019074907-appb-000049
其中,l表示拉格朗日函数,<·>表示矩阵的内积,μ是惩罚因子。Y是拉格朗日数乘矩阵。该无约束模型最小化问题通过如下迭代过程解决:
Figure PCTCN2019074907-appb-000050
步骤3.1.3:迭代优化。矩阵A的优化模型为:
Figure PCTCN2019074907-appb-000051
为了计算方便,核范数最小化问题可以通过soft阈值算子解决,soft阈值的计算公式为(x,τ)=sgn(x)(|x|-τ) +,其中y +=max(y,0),该算子可以被用如下优化过程:
Figure PCTCN2019074907-appb-000052
其中,USV T是矩阵Z的奇异值分解。对
Figure PCTCN2019074907-appb-000053
U∈R m×r,V∈R r×n。r为矩阵的秩。
因此,矩阵A的优化问题转化为
Figure PCTCN2019074907-appb-000054
同样的,矩阵E的优化问题转化为
Figure PCTCN2019074907-appb-000055
步骤3.1.4:设置迭代截止条件,截止条件为:
Figure PCTCN2019074907-appb-000056
其中,S为权重矩阵,S权重矩阵的使用可以大大降低迭代时间。从而提高检测速度。
步骤3.2:基于多数据融合的管道异常候选区域搜索;
步骤3.2.1:分别对单轴数据在基于拉格朗日数乘的数据重构框架下进行异常区域搜索,得出三轴异常区域分别为Ο XYZ
步骤3.2.2:建立三轴融合优化框架:
Figure PCTCN2019074907-appb-000057
步骤3.2.3:用非极大值抑制算法消除重叠,同时考虑候选区域生成的多样性,距离近的窗体合并,以二者的最大外围为新窗体外围,合并准则为:相邻窗体中心横向距离如果小于相邻窗体横向长度的最小值。
步骤3.3基于可进化模型的管道漏磁异常识别。
步骤3.3.1:从完备漏磁数据集中提取异常样本,建立基于卷积神经网络(Convolutional Neural Network,CNN)的异常识别模型。
步骤3.3.2:针对那些识别错误的样本,添加新的标签,作为新的分类,转到步骤3.3.1,重新建立异常识别模型,重新进行分类,找出有缺陷的漏磁信号。
所述量化模块,采用基于随机森林的缺陷量化方法,得到缺陷尺寸,具体包括如下步骤:
步骤4.1:收集数据;对缺陷漏磁信号进行检测,并将其漏磁信号进行特征提取,得到缺陷漏磁信号特征值,具体为:
根据轴向最大通道漏磁信号上的极小值点,找到轴向最大通道漏磁信号峰谷位置及峰谷值,判断为单双峰缺陷之后,提取波形相关10个特征,分别为:单峰缺陷峰值,单峰最大峰谷差,双峰谷宽度,双峰缺陷信号左峰谷差和右峰谷差,双峰缺陷信号峰峰间距,轴向特殊点间距,面积特征,面能量特征,缺陷体积,缺陷体能量。
10个特征具体描述如下:
A.单峰缺陷峰值:Y v则为缺陷最小谷值,Y p-v为最大峰谷差。由于缺陷漏磁信号受到内检测器检测环境等多种因素影响,数据的基准线波动较大。取缺陷数据的峰谷差值作为特征量可以很好地消除信号基线的影响,可以提高缺陷定量分析的可靠性。
B.单峰最大峰谷差:其表达式为:Y p-v=Y p-Y v,式中Y p是单峰缺陷峰值,Y v则为缺陷最小谷值,Y p-v为最大峰谷差。由于缺陷漏磁信号受到内检测器检测环境等多种因素影响,数据的基准线波动较大。取缺陷数据的峰谷差值作为特征量可以很好地消除信号基线的影响,可以提高缺陷定量分析的可靠性。
C.双峰谷宽度:用公式表示为:X v-v=X vr-X vl,其中,X v-v表示缺陷轴向信号谷宽度,X vr是缺陷右谷位置,X vl为缺陷左谷位置。缺陷信号的谷宽度能够反映出缺陷信号在轴向上的分布情况。
D.双峰缺陷信号左峰谷差和右峰谷差:用公式表示为:Y lp-lv=Y lp-Y lv,Y rp-rv=Y rp-Y rv,式中,Y lv是漏磁信号的左谷值,Y rv是漏磁信号右谷值,Y lp为双峰信号左峰值,Y rp为双峰信号右峰值,Y lp-lv为左峰谷差,Y rp-rv为右峰谷差。
E.双峰缺陷信号峰峰间距:其表达式为:X p-p=X pr-X pl,式中,X pr是右峰位置,X pl为左锋位置,X p-p是信号峰峰间距。缺陷信号的峰峰间距与峰谷值的结合能够大致的确定异常数据曲线的形状,有助于对缺陷长度和深度进行定量分析。
F.轴向特殊点间距:为求缺陷长度的关键特征量,特殊点提取方法为:设置求长的比例m_RateA,根据X+(Y-X)*m_RateA求出阈值,其中,X为谷值平均值,Y为最大峰值,轴向最大通道漏磁信号中和阈值最接近的两个点即为特殊点,特殊点的间距为求缺陷长度的关键特征量。
G.面积特征:以数值较低的谷值为基线,取两个谷之间的数据曲线与基线之间覆盖的面积,用公式表示为:
Figure PCTCN2019074907-appb-000058
其中,S a表示缺陷波形面积;x(t)表示缺陷信号数据点;min[x(t)]表示缺陷最小谷值;N 1表示缺陷左谷位置;N 2表示缺陷右谷位置。
H.面能量特征:求取两个谷之间的数据曲线的能量,用公式表示为:
Figure PCTCN2019074907-appb-000059
式中:S e为缺陷波形面能量。
I.缺陷体积:缺陷的体积就是在缺陷通道范围内对缺陷面积求和,用公式表示为:
Figure PCTCN2019074907-appb-000060
式中,V a表示缺陷体积;n 1表示向信号特殊点位置确定的起始通道;n 2表示周向信号特殊点位置确定的终止通道;S a(t)表示单条通道轴向缺陷面积。
J.缺陷体能量:缺陷体能量就是在缺陷范围内对缺陷面能量进行求和,其表达式为:
Figure PCTCN2019074907-appb-000061
式中:V e表示缺陷体能量;S e(t)表示单条轴向缺陷信号面能量。
步骤4.2:以缺陷漏磁信号特征值作为样本;人工测量缺陷尺寸作为标签,缺陷尺寸包括缺陷的深度、宽度和长度;人工选择初始训练集和测试集。
步骤4.3:训练网络;将训练集输入初始随机森林网络中。
步骤4.4:调整网络;通过测试集检验随机森林回归网络结果,并通过调参调整网络得到最终网络。
步骤4.4.1:从原始漏磁信号特征缺陷样本M h×N h维中使用Bootstraping方法随机有放回采样选出m e个缺陷样本,m e≤M h,共进行T c次采样,生成T c个训练集;
步骤4.4.2:对于T c个训练集,分别训练T c个回归树模型。
步骤4.4.3:对于单个回归树模型,在漏磁缺陷信号特征集中,选择n e个特征,其中,n e≤N,然后每次分裂时根据信息增益比
Figure PCTCN2019074907-appb-000062
公式中H A(D)表示特征A的熵,g(D,A)表示其信息增益。选择信息增益比最大值的特征进行分裂。初始将参数最大特征数max_features设置为None,即不限制网络选取特征数;
步骤4.4.4:每棵树都一直这样分裂下去,在分裂过程中为防止过拟合,考虑回归树的复杂度,对回归树进行剪枝。剪枝通过极小化损失函数C α(T)=C(T)+α|T|,其中C(T)代表模型对缺陷尺寸预测误差,即拟合程度。|T|代表模型复杂度,α用于调节回归树复杂度。使用国际海油运输管道POF标准,将损失函数的预测误差取为POF 90%位置占壁厚的百分数。
步骤4.4.5:模型调参优化,采用CVGridSearch网格搜索及K折交叉验证寻找最优参数,最优参数包括随机森林框架参数,袋外样本评估分数e oob和最大迭代次数,以及树模型参数最大特征数,最大深度,内部节点再划分所需最小样本数和叶子节点最少样本数。
步骤4.4.6:将生成的多棵决策树组成随机森林,对于缺陷特征样本建立的回归问题网络,由多棵树预测值的均值决定最终预测缺陷尺寸大小。
步骤4.5:将测试集中待测数据输入到按照步骤4.4调整的随机森林网络中,输出预测缺陷尺寸,此时若待测数据为缺陷尺寸中的深度,则输出为预测缺陷尺寸的深度;若待测数据为缺陷尺寸中的宽度,则输出为预测缺陷尺寸的宽度,若待测数据为缺陷尺寸的长度,则输出为预测缺陷尺寸的长度;其中,按照国际输油管道POF标准,其预测深度反映了按误差绝对值排序百分之80位置的值,公式为:
Figure PCTCN2019074907-appb-000063
其中
Figure PCTCN2019074907-appb-000064
分别为设计深度与 预测深度。
所述解决方案模块,采用基于ASME B31G标准改进的管道解决方案,导入维修决策模型,输出评估结果,具体包括如下步骤:
步骤5.1:从完备漏磁数据集中提取缺陷信息中所有的缺陷长度列、深度列以及管道属性参数;其中管道属性参数包括最小屈服强度SMYS、最小拉伸强度SMTS、公称外径D d、壁厚t a和最大允许操作压力MAOP;
步骤5.2:计算流变应力值
Figure PCTCN2019074907-appb-000065
其中SMYS是管材的最小屈服强度,单位为Mpa;SMTS是最小拉伸强度,单位为Mpa。
步骤5.3:计算管道预测失效压力
Figure PCTCN2019074907-appb-000066
当z<=20时,长度伸缩系数
Figure PCTCN2019074907-appb-000067
当z>20时,长度伸缩系数L 0=(ηz+λ a),
Figure PCTCN2019074907-appb-000068
腐蚀区金属损失面积
Figure PCTCN2019074907-appb-000069
原始面积A area0=t aL。其中d为缺陷深度,单位为mm;t a为管道壁厚,单位为mm;D d为公称外径,单位为mm。
步骤5.4:计算管道最大失效压力
Figure PCTCN2019074907-appb-000070
整理得:
Figure PCTCN2019074907-appb-000071
当z<=20时,θ a=2/3,当z>20时,θ a=1。
步骤5.5:计算维修指数
Figure PCTCN2019074907-appb-000072
其中
Figure PCTCN2019074907-appb-000073
其中P为最大允许设计压力;如果维修指数ERF小于1表示缺陷可接受,大于等于1不可接受,此时应该维修或换管。
步骤5.6:导入维修决策模型,基于专家经验及寿命预测模型进行定性和定量分析然后评估管道腐蚀的严重程度,制定维修规则,根据维修规则,输出评估结果,包括:维修指数及维修建议;其中,规则一:缺陷处的壁厚损失最大深度大于等于80%属于重大腐蚀,维修建议:应该立即维修或换管,规则二:缺陷处的ERF大于等于1,属于严重腐蚀,维修建议:应该马上维修,规则三:缺陷处的ERF值大于等于0.95且小于1.0,属于一般腐蚀,维修建议:可以观察1-3个月,规则四:缺陷处最大深度大于等于20%且小于40%属于轻微腐蚀,维修建议:可以定期观察,不做处理。
有益技术效果:
(1)本发明的数据完备集构建模块,相比于一般的基线校正算法,提出了一种二次基线 校正算法,该方法降低了异常数据对整体基值的影响,提高了基线校正的准确性。同时采用在KNN每个盒子中添加逻辑回归的算法实现对缺失数据的插补,该方法适用于不同种类的数据缺失类型中,同时针对实际工程数据的不确定性具有很强的抗干扰能力。
(2)在发现模块中,区别于一般的焊缝检测方法,引入选择性搜索算法生成候选区域,提高了生成候选区的速度与准确性;利用卷积神经网络算反对候选区域进行将缺分类,增加了焊缝检测算法对于信号噪声的鲁棒性,提高了分类准确性;
(3)在发现模块中,本发明采用多源漏磁数据进行重构,通过分析重构数据与源数据的偏差实现异常检测,同时为了提升算法速度,一种新颖的权重矩阵应用在条件计算中。实验结果表明该方法异常检出效果良好。
(4)在量化模块中,区别于一般的特征提取方法,根据漏磁信号突变不平稳的特点。本发明提出了一种基于漏磁信号波形及统计特征提取方法,增强了模型识别效果;运用海油管道POF标准自定义了随机森林迭代损失函数,使算法在本领域适应性高,缺陷量化结果精度高。本发明方法已经应用于实际工程管道反演,量化缺陷尺寸效果良好。
(5)在所述解决方案模块,本发明基于实际工程应用,相对于原始的ASME B31G方法,本次的方法改进了流变应力的计算从而增大了失效压力,降低了保守性,而ASME B31G保守性过高,因此不会因为ASME B31G的保守性过高而进行频繁的维修产生大量的维修费用。
(6)本发明提出了一种管道内检测漏磁数据智能分析系统及方法,相比于一般的漏磁数据分析方法,本发明从整体角度提出了一种管道内检测漏磁数据智能分析流程,该流程顺序包括:数据完备集构建模块、发现模块、量化模块以及解决方案模块。该流程实现了对原始管道内检测漏磁数据的预处理,连接组件的检测和异常检测,异常检测包括:缺陷、阀门仪表和金属增加,缺陷尺寸的反演以及最终的维修决策。
附图说明
图1为本发明实施例的一种管道内检测漏磁数据智能分析系统运行过程流程图;
图2为本发明实施例的一种管道内检测漏磁数据智能分析系统框图;
图3为本发明实施例的基于类时域稀疏采样和KNN-softmax的数据完备集构建方法流程图;
图4为本发明实施例的基于选择性搜索与卷积神经网络相结合的管道连接组件发现方法流程图;
图5为本发明实施例的基于拉格朗日数乘的异常区域搜索示意图;
图6为本发明实施例的基于多源漏磁数据融合的异常候选区域推荐和识别框架示意图;
图7为本发明实施例的一种基于ASME B31G标准改进的管道解决方案流程图;
图8为本发明实施例的基线校正前后的数据示意图;其中,图8(a)为基线校正前的数据示意图,图8(b)为基线校正后的数据示意图;
图9本发明实施例的进行KNN-softmax算法插补前后得到的完备数据集的示意图;其中,图9(a)为未进行插补的完备数据集示意图,图9(b)为进行KNN-softmax算法插补后得到的完备数据集的示意图;
图10为本发明实施例的管道连接组件发现仿真结果示意图;
图11为本发明实施例的找出有缺陷的漏磁信号仿真结果示意图;
图12为本发明实施例的缺陷量化性能对比柱状图;
图13为本发明实施例的缺陷反演残差散点图;
图14为本发明实施例的比较ASME B31G 1991及改进的标准剩余强度评价曲线。
具体实施方式
下面将结合附图和实施例,对本发明作进一步描述。
本发明提供一种管道内检测漏磁数据智能分析软件系统,从无损检测评估整体角度提出内检测漏磁数据分析系统,并从智能角度发明了一种基于类时域稀疏采样和KNN-softmax的数据完备集构建方法、一种基于选择性搜索与卷积神经网络相结合的管道连接组件发现方法、一种基于拉格朗日数乘框架和多源漏磁数据融合的异常候选区域搜索与识别方法、一种基于随机森林的缺陷量化方法和一种基于ASME B31G标准改进的管道解决方案。实现管道的安全运维。
如图2所示,为本发明的漏磁数据智能分析软件系统框图,整个系统包括4个模块:数据完备集构建模块、发现模块、量化模块以及解决方案模块。其中数据完备集构建模块实现数据的异常检测以及重构,构建完备数据集;发现模块包括组件检测和异常检测,其主要目标是识别缺陷;量化模块实现从信号到物理属性的映射,得出缺陷的长度、宽度和深度;解决方案模块是综合缺陷检测、尺寸反演结果以及管道属性和历史数据知识模型,最终给出维修策略。
本发明提出的管道内检测漏磁数据智能分析系统,如图1所示,具体实施如下:
所述数据完备集构建模块,采用一种基于类时域稀疏采样和KNN-softmax的数据完备集构建方法,得到完备漏磁数据集。
如图3所示,为本发明基于类时域稀疏采样和KNN-softmax的数据预处理流程图。首先对数据进行两次基线校正,然后对数据进行类时域建模并对数据进行异常识别。最后针对数据异常部分,运用KNN-softmax回归模型进行数据插补,最终构建完备漏磁数据集。基于类时域稀疏采样和KNN-softmax的数据预处理的具体步骤如下:
步骤1.1:从海底管道漏磁检测仪中直接采集原始漏磁检测数据,并且对数据进行二次基线校正,其中,原始采样漏磁数据作为多源数据信息,具体包括:轴向数据、径向数据、周向数据、α向数据;。
步骤1.1.1:对原始漏磁检测数据进行一次基线校正,表示为:
Figure PCTCN2019074907-appb-000074
其中,k c为里程计数点数量;
Figure PCTCN2019074907-appb-000075
为第j a通道在第i a里程计数点位置的原始值;
Figure PCTCN2019074907-appb-000076
为第j a通道在第i a里程计数点位置的校正后的值;s为所有通道的中值,n a为漏磁内检测器通道数。
步骤1.1.2:去除数据中的超限值±T a,将超限值的位置值赋予所有通道的中值s,表示为:
Figure PCTCN2019074907-appb-000077
步骤1.1.3:对去除超限值的数据进行二次基线校正:
Figure PCTCN2019074907-appb-000078
其中,k c为里程计数点数量;
Figure PCTCN2019074907-appb-000079
为第j a通道在第i a里程计数点位置的一次校正值;
Figure PCTCN2019074907-appb-000080
为第j a通道在第i a里程计数点位置的二次校正后的值;s′为一次校正后所有通道的中值。
步骤1.2:对二次基线校正后的数据进行类时域稀疏采样异常检测处理;
步骤1.2.1:对二次基线校正后的数据进行异常信号类时域建模,即将采样点对应为时间信息;
步骤1.2.1.1:对异常部分进行数学建模,建模结果所示为:
f(t)'=p(t)'*sin(2πnft)
其中,
Figure PCTCN2019074907-appb-000081
其中,p(t)′表示管道漏磁检测电压凸起补偿信号,f表示信号采样率,t表示采样时间,t 1,t 2表示采样时间间隔,a表示电力管道,n为系统波动幅值系数,f(t)'电压波形变化频率;
步骤1.2.1.2:以区间为采集单位设定漏磁检测异常数据变化量,以k e=100个采集数据为一个区间,将各区间内采集的管道系统电压数据的方差视为数据变化量,判定漏磁数据电压信号波动程度,具体计算方法为:
Figure PCTCN2019074907-appb-000082
即:
Figure PCTCN2019074907-appb-000083
其中,
Figure PCTCN2019074907-appb-000084
表示给定区间内的第i c个取样点,
Figure PCTCN2019074907-appb-000085
表示该区间内采集管道系统电压数据的平均值,Δf 0表示漏磁数据电压信号波动程度。
步骤1.2.1.3:计算电压状态变化量
Figure PCTCN2019074907-appb-000086
公式如下:
Figure PCTCN2019074907-appb-000087
步骤1.2.2:异常信号判别:若
Figure PCTCN2019074907-appb-000088
则此时的数据认为是外界扰动产生的异常,即为异常部分。
步骤1.2.3:人工提取训练样本特征T=(X 1,X 2,…,X 7,X 8),共提取8个特征,分别是数据的左谷值、右谷值、谷宽度、峰值、左峰谷差、右峰谷差、微分左峰值和微分右峰值。
人工提取测试样本特征T′=(X′ 1,X′ 2,…,X′ 7,X′ 8),同样提取8个特征,分别是数据的左谷值、右谷值、谷宽度、峰值、左峰谷差、右峰谷差、微分左峰值和微分右峰值。
人工提取待插补数据特征T″=(X″ 1,X″ 2,…,X″ 7,X″ 8),同样提取8个特征,分别是数据的左谷值、右谷值、谷宽度、峰值、左峰谷差、右峰谷差、微分左峰值和微分右峰值。
步骤1.3:对海底管道漏磁数据进行基于KNN-逻辑回归的缺失插补处理。
步骤1.3.1:训练测试KNN、softmax回归模型。
步骤1.3.1.1将特征样本数据T分为两部分,一部分特征样本数据X Train用于训练KNN模型,另一部分特征样本数据T Test用于测试KNN模型。
步骤1.3.1.2:将X Train输入到KNN模型中,K值初始值设定为5,训练KNN模型。
步骤1.3.1.3:将T Test输入到训练完成的KNN模型中进行分类,计算判别错误率,若错误率小于阈值,则采用V折交叉法,改变训练和测试样本继续训练,否则令K=K+1继续训练模型,当K大于阈值M时停止训练。
步骤1.3.1.4:分到每类中的特征样本数据
Figure PCTCN2019074907-appb-000089
特征样本对应的数据集为
Figure PCTCN2019074907-appb-000090
Figure PCTCN2019074907-appb-000091
Figure PCTCN2019074907-appb-000092
分别进行归一化处理得到
Figure PCTCN2019074907-appb-000093
Figure PCTCN2019074907-appb-000094
表示为:
Figure PCTCN2019074907-appb-000095
Figure PCTCN2019074907-appb-000096
是特征样本数据的平均值;
Figure PCTCN2019074907-appb-000097
特征样本对应数据的平均值;
步骤1.3.1.5:在每一类的节点处添加一个softmax回归模型,其假设函数为如式表示为:
Figure PCTCN2019074907-appb-000098
其中,x为样本输入值,y为样本输出值;θ为训练模型参数;k f为向量维数;i e为分类的第i e个类别;p(y=i e|x)表示对类别i e估算的概率值。
步骤1.3.1.6:将每个节点处中的训练样本集
Figure PCTCN2019074907-appb-000099
输入到softmax回归模型中,得到插补之后的输出值
Figure PCTCN2019074907-appb-000100
其损失函数J(θ)为:
Figure PCTCN2019074907-appb-000101
其中,x为样本输入值;y为样本输出值;θ为训练模型参数;k f为向量维数;i e为分类的第i e个类别;j e为该分类中第j e个样本输入;m d为样本个数;1{·}为示性函数,若大括号里为真值,表达式值为1。
步骤1.3.2:计算预测结果的损失函数,设定阈值P=0.5。若J(θ)>P,回到步骤1.3.2.2,令K=K+1继续训练模型。直至J(θ)≤P,当K大于阈值M时停止训练,输出插补之后的输出值y (i)'。
步骤1.3.3:将待插补数据特征及数据集输入到训练完成的模型中,实现对缺失数据的插补,得到完备漏磁数据集,因原始采样漏磁数据作为多源数据信息,故得到完备多源漏磁数据集。
步骤1的仿真结果:图8(a)为基线校正前的数据示意图,从图8(a)中可知,未进行基线校正的数据基值差距很大,加入偏移量后每个通道的数据分布不均匀;图8(b)为基线 校正后的数据示意图,由图8(b)可知进行基线校正后的数据基值相等,加入偏移量后每个通道数据分布均匀,从而降低了后续数据处理的误差。
图9(a)为未进行数据插补的,含有缺失的数据集的示意图,图9(b)为进行KNN-softmax算法插补后得到的完备数据集的示意图,由图9(b)可知,无论在缺陷位置还是在平滑位置,该算法都能将缺失的数据插补完成。
所述发现模块,采用基于选择性搜索与卷积神经网络相结合的管道连接组件发现方法,得到焊缝的精确位置,具体包括如下步骤:
如图4所示,为本发明基于选择性搜索与卷积神经网络相结合的管道连接组件检测流程,首选将漏磁信号转换为彩色图,然后利用祖安泽兴搜索获得候选区域并利用卷积神经网路进行识别,最后采用非极大值抑制的方法去除区域重叠,过得最终组件位置,具体步骤如下:
步骤2.1:提取管道漏磁信号数据:从完备漏磁数据集中,取将整段管道漏磁信号矩阵D,等比例分割成n g段管道漏磁信号矩阵
Figure PCTCN2019074907-appb-000102
每一个分割后的漏磁信号矩阵都由
Figure PCTCN2019074907-appb-000103
个数据组成;漏磁内检测器采集得到的大小M×N整段管道漏磁信号矩阵D。等比例分割成大小为M 1×N,M 2×N,...M 10×N的漏磁信号矩阵D 1,D 2,...,D 10。其中M 1=M 2=...=M 10,M 1+M 2+...+M 10=M。
步骤2.2:漏磁信号转换色彩图:设置信号幅值上限A top和信号幅值下线A floor,并据此将管道漏磁信号矩阵
Figure PCTCN2019074907-appb-000104
转换为管道彩色图矩阵
Figure PCTCN2019074907-appb-000105
步骤2.2.1:设置信号幅值上限A top和信号幅值下线A floor
步骤2.2.2:根据如下公式将管道漏磁信号矩阵
Figure PCTCN2019074907-appb-000106
转换为0-255之间的灰度矩阵
Figure PCTCN2019074907-appb-000107
Figure PCTCN2019074907-appb-000108
其中
Figure PCTCN2019074907-appb-000109
d ij为漏磁信号矩阵D组成元素,gray ij为灰度矩阵Gray组成元素;
步骤2.2.3:根据如下公式将灰度矩阵
Figure PCTCN2019074907-appb-000110
转换为包含
Figure PCTCN2019074907-appb-000111
Figure PCTCN2019074907-appb-000112
的三维彩色矩阵
Figure PCTCN2019074907-appb-000113
Figure PCTCN2019074907-appb-000114
其中,c=255,r ij为矩阵R组成元素。g ij为矩阵G组成元素。b ij为矩阵B组成元素。
步骤2.3:选择性搜索:对每一段管道彩色图C k,利用选择性搜索,提取出m c个候选区域
Figure PCTCN2019074907-appb-000115
步骤2.3.1:每一段管道彩色图C k,利用切分方法得到候选的区域集合R k={r k1,r k2,...,r kw}。
步骤2.3.2:初始化相似度集合Sim=φ。
步骤2.3.3:根据如下公式计算所有相临区域r ka,r kb的相似度sim{r ka,r kb}。
Figure PCTCN2019074907-appb-000116
步骤2.3.4:重复步骤2.3.3直到所有相邻区域的相似度都计算出来,根据如下公式更新相似度合集Sim。
Sim=Sim∪sim(r ka,r kb)
步骤2.3.5:从Sim找到最大的相似度sim{r kc,r kd}=max(Sim),并根据此得到合并区域r ke=r kc∪r kd。从Sim中剔除出sim{r kc,r kd}。
步骤2.3.6:重复步骤2.3.5,直到Sim为空。得到m c个合并后区域
Figure PCTCN2019074907-appb-000117
这些区域就是候选区域。
步骤2.4:卷积神经网络:候选区域识别:
步骤2.4.1:构建输入为72×72的卷积神经网络的,卷积神经网络的中间层包括4个卷积层,4个降采样层和1个全连接层。其中,每个卷积层后面都紧跟着一个用来求局部加权平均的降采样层作为二次特征提取。
步骤2.4.2:从历史数据中提取P个N l×N l的焊缝彩色图作为卷积神经网络的样本。随机取样本中80%作为训练样本,剩余20%作为测试样本。
步骤2.4.3:将网络反复训练500次,其中测试成功率最高的作为最终网络Net。
步骤2.4.4:将候选区域r k1,r k2,...r km分别输入训练好的卷积神经网络进行判别。对于判断是焊缝的区域,记录下其位置Loc与网络评分Soc。最终得到w个位置Loc 1,Loc 2,...,Loc w及其评 分Soc 1,Soc 2,...,Soc w
步骤2.5:非极大值抑制:根据上述焊缝位置Loc 1,Loc 2,...,Loc w及其评分Soc 1,Soc 2,...,Soc w,根据非极大值抑制算法,得到焊缝的精确位置L 1,L 2,...,L u
步骤2的仿真结果,如图10所示:本发明提出的管道组件发现方法:准确率为95.3%召回率为97.94%,相比于采用阈值的传统方法:准确率为91.5%召回率为95.51%。可以看出本发明提出的方法性能更好。
根据焊缝的精确位置,将整段管道漏磁信号分别u+1段,取其中一段漏磁信号,所述发现模块,采用一种基于拉格朗日数乘框架和多源漏磁数据融合的异常候选区域搜索与识别方法,找出有缺陷的漏磁信号,如图6所示,具体包括如下步骤:
步骤3.1:建立基于拉格朗日数乘的数据重构框架;
如图5所示,为本发明基于拉格朗日数乘的异常区域搜索流程,通过拉格朗日数乘算法将有约束优化模型化为无约束优化模型,最终通过交替迭代获得重构矩阵,从而得到重构矩阵与观测矩阵的误差矩阵,通过合适的阈值获得异常区域,最终将区域规则化。具体步骤如下:
步骤3.1.1:建立数据重构模型
Figure PCTCN2019074907-appb-000118
subject to D=A+E。
步骤3.1.2:将有约束优化化模型为无约束优化模型
Figure PCTCN2019074907-appb-000119
该无约束模型最小化问题可以通过如下迭代过程解决:
Figure PCTCN2019074907-appb-000120
步骤3.1.3:迭代优化。矩阵A的优化模型为:
Figure PCTCN2019074907-appb-000121
为了计算方便,核范数最小化问题可以通过soft阈值算子解决,soft阈值的计算公式为(x,τ)=sgn(x)(|x|-τ) +,其中y +=max(y,0),该算子可以被用如下优化过程:
Figure PCTCN2019074907-appb-000122
因此,矩阵A的优化问题转化为
Figure PCTCN2019074907-appb-000123
同样的,矩阵E的优化问题转化为
Figure PCTCN2019074907-appb-000124
步骤3.1.4:设置迭代截止条件。截止条件为:
Figure PCTCN2019074907-appb-000125
其中S为权重矩阵,S权重矩阵的实用可以大大降低迭代时间。从而提高检测速度。本发明S矩阵的设置如下:
Figure PCTCN2019074907-appb-000126
步骤3.2:基于多数据融合的管道异常候选区域搜索;
如图6所示,为本发明基于多源漏磁数据融合的异常候选区域推荐和识别框架,在上述数据重构框架下对多源数据分别进行异常区域推荐,然后通过区域优化框架,从边界和区域角度进行优化,最终得到异常候选区域,输入到识别模型,最终进行分类。具体步骤如下:
步骤3.2.1:分别对单轴数据在基于拉格朗日数乘的数据重构框架下进行异常区域搜索。得出三轴异常区域分别为Ο XYZ
步骤3.2.2:建立三轴融合优化框架
Figure PCTCN2019074907-appb-000127
步骤3.2.3:用非极大值抑制算法消除重叠,同时考虑候选区域生成的多样性,距离近的窗体合并,以二者的最大外围为新窗体外围,合并准则为:相邻窗体中心横向距离如果小于相邻窗体横向长度的最小值。
步骤3.3基于可进化模型的管道漏磁异常识别。
步骤3.3.1:从完备漏磁数据集中提取异常样本,建立基于卷积神经网络(Convolutional Neural Network,CNN)的异常识别模型。
步骤3.3.2:针对那些识别错误的样本,添加新的标签,重新输入模型中进行训练,随着变迁数据数据的不断增多,识别模型不断进化。
步骤3中仿真结果,如图11所示:本发明提出的管道异常发现方法:准确率为95.73%召回率为93.86%;单轴数据异常发现准确率为93.07%召回率为89.73%;相比于传统基于特征提取的方法:准确率为88.98%召回率为81.93%。可以看出,本方法性能更好。
所述量化模块,采用一种基于随机森林的缺陷量化方法,得到缺陷尺寸,具体包括如下 步骤:
步骤4.1:收集数据;对缺陷漏磁信号进行检测,并将其漏磁信号进行特征提取,得到缺陷漏磁信号特征值,具体为:
根据轴向最大通道漏磁信号上的极小值点,找到轴向最大通道漏磁信号峰谷位置及峰谷值,判断为单双峰缺陷之后,提取波形相关10个特征,分别为:单峰缺陷峰值,单峰最大峰谷差,双峰谷宽度,双峰缺陷信号左峰谷差和右峰谷差,双峰缺陷信号峰峰间距,轴向特殊点间距,面积特征,面能量特征,缺陷体积,缺陷体能量。
10个特征具体描述如下:
A.单峰缺陷峰值:Y v则为缺陷最小谷值,Y p-v为最大峰谷差。由于缺陷漏磁信号受到内检测器检测环境等多种因素影响,数据的基准线波动较大。取缺陷数据的峰谷差值作为特征量可以很好地消除信号基线的影响,可以提高缺陷定量分析的可靠性。
B.单峰最大峰谷差:其表达式为:Y p-v=Y p-Y v,式中Y p是单峰缺陷峰值,Y v则为缺陷最小谷值,Y p-v为最大峰谷差。由于缺陷漏磁信号受到内检测器检测环境等多种因素影响,数据的基准线波动较大。取缺陷数据的峰谷差值作为特征量可以很好地消除信号基线的影响,可以提高缺陷定量分析的可靠性。
C.双峰谷宽度:用公式表示为:X v-v=X vr-X vl,其中,X v-v表示缺陷轴向信号谷宽度,X vr是缺陷右谷位置,X vl为缺陷左谷位置。缺陷信号的谷宽度能够反映出缺陷信号在轴向上的分布情况。
D.双峰缺陷信号左峰谷差和右峰谷差:用公式表示为:Y lp-lv=Y lp-Y lv,Y rp-rv=Y rp-Y rv,式中,Y lv是漏磁信号的左谷值,Y rv是漏磁信号右谷值,Y lp为双峰信号左峰值,Y rp为双峰信号右峰值,Y lp-lv为左峰谷差,Y rp-rv为右峰谷差。
E.双峰缺陷信号峰峰间距:其表达式为:X p-p=X pr-X pl,式中,X pr是右峰位置,X pl为左锋位置,X p-p是信号峰峰间距。缺陷信号的峰峰间距与峰谷值的结合能够大致的确定异常数据曲线的形状,有助于对缺陷长度和深度进行定量分析。
F.轴向特殊点间距:为求缺陷长度的关键特征量,特殊点提取方法为:设置求长的比例m_RateA,根据X+(Y-X)*m_RateA求出阈值,其中,X为谷值平均值,Y为最大峰值,轴向最大通道漏磁信号中和阈值最接近的两个点即为特殊点,特殊点的间距为求缺陷长度的关键特征量。
G.面积特征:以数值较低的谷值为基线,取两个谷之间的数据曲线与基线之间覆盖的面积,用公式表示为:
Figure PCTCN2019074907-appb-000128
其中,S a表示缺陷波形面积;x(t)表示缺陷信号数据点;min[x(t)]表示缺陷最小谷值;N 1表示缺陷左谷位置;N 2表示缺陷右谷位置。
H.面能量特征:求取两个谷之间的数据曲线的能量,用公式表示为:
Figure PCTCN2019074907-appb-000129
式中:S e为缺陷波形面能量。
I.缺陷体积:缺陷的体积就是在缺陷通道范围内对缺陷面积求和,用公式表示为:
Figure PCTCN2019074907-appb-000130
式中,V a表示缺陷体积;n 1表示向信号特殊点位置确定的起始通道;n 2表示周向信号特殊点位置确定的终止通道;S a(t)表示单条通道轴向缺陷面积。
J.缺陷体能量:缺陷体能量就是在缺陷范围内对缺陷面能量进行求和,其表达式为:
Figure PCTCN2019074907-appb-000131
式中:V e表示缺陷体能量;S e(t)表示单条轴向缺陷信号面能量。
步骤4.2:以缺陷漏磁信号特征值作为样本;人工测量缺陷尺寸作为标签,缺陷尺寸包括缺陷的深度、宽度和长度;人工选择初始训练集和测试集。
步骤4.3:训练网络;将训练集输入初始随机森林网络中。
步骤4.4:调整网络;通过测试集检验随机森林回归网络结果,并通过调参调整网络得到 最终网络,具体实践:输入M h=666N h=6,设置参数m f=sqrt(),T f=56。具体实践:初始将参数设置n f=n f/3,最大特征数max_features设置为None。
步骤4.4.1:从原始漏磁信号特征缺陷样本M h×N h维中使用Bootstraping方法随机有放回采样选出m e个缺陷样本,m e≤M h,共进行T c次采样,生成T c个训练集;
步骤4.4.2:对于T c个训练集,分别训练T c个回归树模型。
步骤4.4.3:对于单个回归树模型,在漏磁缺陷信号特征集中,选择n e个特征,其中,n e≤N,然后每次分裂时根据信息增益比
Figure PCTCN2019074907-appb-000132
公式中H A(D)表示特征A的熵,g(D,A)表示其信息增益。选择信息增益比最大值的特征进行分裂。初始将参数最大特征数max_features设置为None,即不限制网络选取特征数;
步骤4.4.4:每棵树都一直这样分裂下去,在分裂过程中为防止过拟合,考虑回归树的复杂度,对回归树进行剪枝。剪枝通过极小化损失函数C α(T)=C(T)+α|T|,其中C(T)代表模型对缺陷尺寸预测误差,即拟合程度。|T|代表模型复杂度,α用于调节回归树复杂度。使用国际海油运输管道POF标准,将损失函数的预测误差取为POF 90%位置占壁厚的百分数。初始将最大树深max_depth设置为5。
步骤4.4.5:模型调参优化,采用CVGridSearch网格搜索及K折交叉验证寻找最优参数。包括随机森林框架参数,袋外样本评估分数e oob和最大迭代次数,以及树模型参数最大特征数max_features,最大深度max_depth,内部节点再划分所需最小样本数和叶子节点最少样本数。
步骤4.4.6:将生成的多棵决策树组成随机森林。对于从缺陷特征样本建立的回归问题网络,由多棵树预测值的均值决定最终预测缺陷尺寸大小信息。
步骤4.5:将待测数据输入到按照步骤4.4调整的随机森林网络中,输出预测缺陷尺寸,此时若待测数据为缺陷尺寸中的深度,则输出为预测缺陷尺寸的深度;若待测数据为缺陷尺寸中的宽度,则输出为预测缺陷尺寸的宽度,若待测数据为缺陷尺寸的长度,则输出为预测缺陷尺寸的长度;其中,按照国际输油管道POF标准,其预测深度反映了按误差绝对值排序百分之80位置的值,公式为:,公式为:
Figure PCTCN2019074907-appb-000133
其中
Figure PCTCN2019074907-appb-000134
分别为设计深度与预测深度。将迭代n p代内损失函数不再减少作为参数寻优终止条件,网络最终输出最大迭代次数n_estimators为172。
步骤4的仿真结果,本发明与传统缺陷反演算法性能对比,如表1所示:
表1缺陷反演算法性能:
Figure PCTCN2019074907-appb-000135
Figure PCTCN2019074907-appb-000136
表1和图12反映出本发明算法按照国际海油管道POF标准,在80%置信度时,缺陷的长度绝对误差在10mm之内,宽度15mm之内,深度绝对误差占壁厚(9.5mm)百分比在10之内。相比较于传统随机森林算法精度更高,且方差较小,达到工业缺陷反演精度要求。实验结果验证了算法泛化能力和鲁棒性较好。
从图13中可以看出缺陷尺寸反演精度和稳定性量化结果较好。并且预测点并无大幅度偏离真实点实例,这对工业故障检测十分重要,因为如果一个缺陷尺寸预测偏差大,会导致后续修复措施损失严重。
所述解决方案模块,采用一种基于ASME B31G标准改进的管道解决方案,导入维修决策模型,输出评估结果,如图7所示,具体包括如下步骤:
步骤5.1:从完备漏磁数据集中提取缺陷信息中所有的缺陷长度列、深度列以及管道属性参数;其中管道属性参数包括最小屈服强度SMYS、最小拉伸强度SMTS、公称外径D d、壁厚t a和最大允许操作压力MAOP。
步骤5.2:计算流变应力值
Figure PCTCN2019074907-appb-000137
其中SMYS是管材的最小屈服强度,单位为Mpa;SMTS是最小拉伸强度,单位为Mpa。
步骤5.3:计算管道预测失效压力
Figure PCTCN2019074907-appb-000138
当z<=20时,长度伸缩系数
Figure PCTCN2019074907-appb-000139
当z>20时,长度伸缩系数L 0=(ηz+λ a),
Figure PCTCN2019074907-appb-000140
腐蚀区金属损失面积
Figure PCTCN2019074907-appb-000141
原始面积A area0=t aL。其中d为缺陷深度,单位为mm;t a为管道壁厚,单位为mm;D d为公称外径,单位为mm。
步骤5.4:计算管道最大失效压力
Figure PCTCN2019074907-appb-000142
整理得:
Figure PCTCN2019074907-appb-000143
当z<=20时,θ a=2/3,当z>20时,θ a=1。
步骤5.5:计算维修指数
Figure PCTCN2019074907-appb-000144
其中
Figure PCTCN2019074907-appb-000145
其中P为最大允许设计压力;如果维修指数ERF小于1表示缺陷可接受,大于等于1不可接受,此时应该维修或换管。
步骤5.6:导入维修决策模型,基于专家经验及寿命预测模型进行定性和定量分析然后评 估管道腐蚀的严重程度,制定维修规则,根据维修规则,输出评估结果,包括:维修指数及维修建议;其中,规则一:缺陷处的壁厚损失最大深度大于等于80%属于重大腐蚀,维修建议:应该立即维修或换管,规则二:缺陷处的ERF大于等于1,属于严重腐蚀,维修建议:应该马上维修,规则三:缺陷处的ERF值大于等于0.95且小于1.0,属于一般腐蚀,维修建议:可以观察1-3个月,规则四:缺陷处最大深度大于等于20%且小于40%属于轻微腐蚀,维修建议:可以定期观察,不做处理。
步骤5仿真结果,如图14所示的曲线是分别将管道基本参数信息带入ASME B31G 1991及改进后的标准后绘制的,改进后的公式通过改变流变应力的值降低了保守性。图中曲线是维修指数ERF=1的曲线,如果缺陷位于曲线上方说明该缺陷腐蚀比较严重应该立即维修,ASME B31G 1991标准保守性过高,实际检测过程中经常增加维修或换管力度,造成经济上的浪费,适用于老管道,由于改进后的公式保守性降低,进而减少了因经常维修而产生的费用。

Claims (7)

  1. 一种管道内检测漏磁数据智能分析系统,其特征在于,包括:数据完备集构建模块、发现模块、量化模块以及解决方案模块;
    原始采样漏磁数据与数据完备集构建模块相连接,数据完备集构建模块通过完备漏磁数据集与发现模块相连接,发现模块与量化模块相连接,量化模块与解决方案模块相连接;
    所述数据完备集构建模块用于对原始漏磁内检测数据进行数据缺失重构以及降噪操作,采用基于类时域稀疏采样和KNN-softmax的数据完备集构建方法,构建完备漏磁数据集;
    所述数据完备集构建模块中将原始采样漏磁数据作为多源数据信息,具体包括:轴向数据、径向数据、周向数据、α向数据;
    所述发现模块是进行缺陷检测,其过程包括组件检测和异常检测,其中,组件检测完成对管道连接组件焊缝和法兰的检测;所述发现模块,采用基于选择性搜索与卷积神经网络相结合的管道连接组件发现方法,得到焊缝的精确位置;根据焊缝的精确位置,将整段管道漏磁信号分别u+1段,取其中一段漏磁信号,采用基于拉格朗日数乘框架和多源漏磁数据融合的异常候选区域搜索与识别方法,找出有缺陷的漏磁信号;
    异常检测包括:缺陷、阀门仪表和金属增加检测,并最终得到缺陷信号;
    所述量化模块完成缺陷信号到物理特性的映射,采用基于随机森林的缺陷量化方法,最终给出缺陷的尺寸,即长度、宽度和深度;
    所述解决方案模块从完备漏磁数据集中提取缺陷信息中所有的缺陷长度列、深度列以及管道属性参数,采用基于ASME B31G标准改进的管道解决方案,通过维修决策模型,最终给出评估结果,评估结果包括单独缺陷位置的维修指数、维修建议。
  2. 根据权利要求1所述的管道内检测漏磁数据智能分析系统,其特征在于,所述管道属性参数包括:最小屈服强度SMYS、最小拉伸强度SMTS、公称外径D d、壁厚t a和最大允许操作压力MAOP。
  3. 根据权利要求1所述的管道内检测漏磁数据智能分析系统,其特征在于,所述数据完备集构建模块,采用基于类时域稀疏采样和KNN-softmax的数据完备集构建方法,得到完备漏磁数据集,具体包括如下步骤:
    步骤1.1:从海底管道漏磁检测仪中直接采集原始漏磁检测数据,并且对数据进行二次基线校正,其中,原始采样漏磁数据作为多源数据信息,具体包括:轴向数据、径向数据、周向数据、α向数据;
    步骤1.1.1:对原始漏磁检测数据进行一次基线校正,表示为:
    Figure PCTCN2019074907-appb-100001
    其中,k c为里程计数点数量;
    Figure PCTCN2019074907-appb-100002
    为第j a通道在第i a里程计数点位置的原始值;
    Figure PCTCN2019074907-appb-100003
    为第j a通道在第i a里程计数点位置的校正后的值;s为所有通道的中值,n a为漏磁内检测器通道数;
    步骤1.1.2:去除数据中的超限值±T a,将超限值的位置值赋予所有通道的中值s,表示为:
    Figure PCTCN2019074907-appb-100004
    步骤1.1.3:对去除超限值的数据进行二次基线校正:
    Figure PCTCN2019074907-appb-100005
    其中,k c为里程计数点数量;
    Figure PCTCN2019074907-appb-100006
    为第j a通道在第i a里程计数点位置的一次校正值;
    Figure PCTCN2019074907-appb-100007
    为第j a通道在第i a里程计数点位置的二次校正后的值;s′为一次校正后所有通道的中值;
    步骤1.2:对二次基线校正后的数据进行类时域稀疏采样异常检测处理;
    步骤1.2.1:对二次基线校正后的数据进行异常信号类时域建模,即将采样点对应为时间信息;
    步骤1.2.1.1:对异常部分进行数学建模,建模结果所示为:
    f(t)'=p(t)'*sin(2πnft)
    其中,
    Figure PCTCN2019074907-appb-100008
    其中,p(t)′表示管道漏磁检测电压凸起补偿信号,f表示信号采样率,t表示采样时间,t 1,t 2表示采样时间间隔,a表示电力管道,n为系统波动幅值系数,f(t)'电压波形变化频率;
    步骤1.2.1.2:以区间为采集单位设定漏磁检测异常数据变化量,以k e个采集数据为一个区间,将各区间内采集的管道系统电压数据的方差视为数据变化量,判定漏磁数据电压信号波动程度,具体计算方法为:
    Figure PCTCN2019074907-appb-100009
    即:
    Figure PCTCN2019074907-appb-100010
    其中,
    Figure PCTCN2019074907-appb-100011
    表示给定区间内的第i c个取样点,
    Figure PCTCN2019074907-appb-100012
    表示该区间内采集管道系统电压数据的平 均值,Δf 0表示漏磁数据电压信号波动程度;
    步骤1.2.1.3:计算电压状态变化量
    Figure PCTCN2019074907-appb-100013
    公式如下:
    Figure PCTCN2019074907-appb-100014
    步骤1.2.2:异常信号判别:若
    Figure PCTCN2019074907-appb-100015
    则此时的数据认为是外界扰动产生的异常,即为异常部分;
    步骤1.2.3:人工提取训练样本特征
    Figure PCTCN2019074907-appb-100016
    人工提取测试样本特征
    Figure PCTCN2019074907-appb-100017
    人工提取待插补数据特征
    Figure PCTCN2019074907-appb-100018
    i b,j b,k d为特征数目;
    步骤1.3:对海底管道漏磁数据进行基于KNN-逻辑回归的缺失插补处理;
    步骤1.3.1:训练测试KNN、softmax回归模型;
    步骤1.3.1.1将特征样本数据T分为两部分,一部分特征样本数据X Train用于训练KNN模型,另一部分特征样本数据T Test用于测试KNN模型;
    步骤1.3.1.2:将X Train输入到KNN模型中,确定一个K值,训练KNN模型;
    步骤1.3.1.3:将T Test输入到训练完成的KNN模型中进行分类,计算判别错误率,若错误率小于阈值,则采用V折交叉法,改变训练和测试样本继续训练,否则令K=K+1继续训练模型,当K大于阈值M时停止训练;
    步骤1.3.1.4:分到每类中的特征样本数据
    Figure PCTCN2019074907-appb-100019
    特征样本对应的数据集为
    Figure PCTCN2019074907-appb-100020
    Figure PCTCN2019074907-appb-100021
    Figure PCTCN2019074907-appb-100022
    分别进行归一化处理得到
    Figure PCTCN2019074907-appb-100023
    Figure PCTCN2019074907-appb-100024
    表示为:
    Figure PCTCN2019074907-appb-100025
    Figure PCTCN2019074907-appb-100026
    是特征样本数据的平均值;
    Figure PCTCN2019074907-appb-100027
    特征样本对应数据的平均值;
    步骤1.3.1.5:在每一类的节点处添加一个softmax回归模型,其假设函数为如式表示为:
    Figure PCTCN2019074907-appb-100028
    其中,x为样本输入值,y为样本输出值;θ为训练模型参数;k f为向量维数;i e为分类的第i e个类别;p(y=i e|x)表示对类别i e估算的概率值;
    步骤1.3.1.6:将每个节点处中的训练样本集
    Figure PCTCN2019074907-appb-100029
    输入到softmax回归模型中,得到插补之后的输出值
    Figure PCTCN2019074907-appb-100030
    其损失函数J(θ)为:
    Figure PCTCN2019074907-appb-100031
    其中,x为样本输入值;y为样本输出值;θ为训练模型参数;k f为向量维数;i e为分类的第i e个类别;j e为该分类中第j e个样本输入;m d为样本个数;1{·}为示性函数,若大括号里为真值,表达式值为1;
    步骤1.3.2:计算预测结果的损失函数,设定阈值P,若J(θ)>P,回到步骤1.3.2.2,令K=K+1继续训练模型,直至J(θ)≤P,当K大于阈值M时停止训练,输出插补之后的输出值y (i)';
    步骤1.3.3:将待插补数据特征及数据集输入到训练完成的模型中,实现对缺失数据的插补,得到完备漏磁数据集,因原始采样漏磁数据作为多源数据信息,故得到完备多源漏磁数据集。
  4. 根据权利要求1所述的管道内检测漏磁数据智能分析系统,其特征在于,所述发现模块,采用基于选择性搜索与卷积神经网络相结合的管道连接组件发现方法,得到焊缝的精确位置,具体包括如下步骤:
    步骤2.1:提取管道漏磁信号数据:从完备漏磁数据集中,取将整段管道漏磁信号矩阵D,等比例分割成n g段管道漏磁信号矩阵
    Figure PCTCN2019074907-appb-100032
    每一个分割后的漏磁信号矩阵都由
    Figure PCTCN2019074907-appb-100033
    个数据组成;
    步骤2.2:漏磁信号转换色彩图:设置信号幅值上限A top和信号幅值下线A floor,并据此将管道漏磁信号矩阵
    Figure PCTCN2019074907-appb-100034
    转换为管道彩色图矩阵
    Figure PCTCN2019074907-appb-100035
    步骤2.2.1:设置信号幅值上限A top和信号幅值下线A floor
    步骤2.2.2:根据如下公式将管道漏磁信号矩阵
    Figure PCTCN2019074907-appb-100036
    转换为0-255之间的灰度矩阵
    Figure PCTCN2019074907-appb-100037
    Figure PCTCN2019074907-appb-100038
    其中,
    Figure PCTCN2019074907-appb-100039
    d ij为漏磁信号矩阵D组成元素,gray ij为灰度矩阵Gray组成元素;
    步骤2.2.3:根据如下公式将灰度矩阵
    Figure PCTCN2019074907-appb-100040
    转换为包含
    Figure PCTCN2019074907-appb-100041
    Figure PCTCN2019074907-appb-100042
    的三维彩色矩阵
    Figure PCTCN2019074907-appb-100043
    Figure PCTCN2019074907-appb-100044
    其中,r ij为矩阵R组成元素,g ij为矩阵G组成元素,b ij为矩阵B组成元素;
    步骤2.3:选择性搜索:对每一段管道彩色图C k,利用选择性搜索,提取出m c个候选区域
    Figure PCTCN2019074907-appb-100045
    步骤2.3.1:每一段管道彩色图C k,利用切分方法得到候选的区域集合R k={r k1,r k2,...,r kw};
    步骤2.3.2:初始化相似度集合Sim=φ;
    步骤2.3.3:根据如下公式计算所有相临区域r ka,r kb的相似度sim{r ka,r kb};
    Figure PCTCN2019074907-appb-100046
    步骤2.3.4:重复步骤2.3.3直到所有相邻区域的相似度都计算出来,根据如下公式更新相似度合集Sim;
    Sim=Sim∪sim(r ka,r kb)
    步骤2.3.5:从Sim找到最大的相似度sim{r kc,r kd}=max(Sim),并根据此得到合并区域:
    r ke=r kc∪r kd
    从Sim中剔除出sim{r kc,r kd};
    步骤2.3.6:重复步骤2.3.5,直到Sim为空,得到m c个合并后区域
    Figure PCTCN2019074907-appb-100047
    这些区域就是候选区域;
    步骤2.4:卷积神经网络:利用卷积神经网络对提取出的候选区域进行判别,模记录卷积神经网络判断的焊缝位置Loc 1,Loc 2,...,Loc w及其评分Soc 1,Soc 2,...,Soc w
    步骤2.5:非极大值抑制:根据步骤2.4中焊缝位置Loc 1,Loc 2,...,Loc w及其评分Soc 1,Soc 2,...,Soc w,根据非极大值抑制算法得到焊缝的精确位置L 1,L 2,...,L u
  5. 根据权利要求1所述的管道内检测漏磁数据智能分析系统,其特征在于,根据焊缝的精确位置,将整段管道漏磁信号分别u+1段,取其中一段漏磁信号,所述发现模块,采用基于拉格朗日数乘框架和多源漏磁数据融合的异常候选区域搜索与识别方法,找出有缺陷的漏磁信号,具体包括如下步骤:
    步骤3.1:建立基于拉格朗日数乘的数据重构框架;
    步骤3.1.1:建立数据重构模型
    Figure PCTCN2019074907-appb-100048
    subject to P=A+E,其中P是观测矩阵,E是误差矩阵,A是重构后的低秩矩阵,||·|| 1表示矩阵的1范数,||·|| *表示矩阵的核范数,λ是权重参数;
    步骤3.1.2:将有约束优化化模型为无约束优化模型:
    Figure PCTCN2019074907-appb-100049
    其中,l表示拉格朗日函数,<·>表示矩阵的内积,μ是惩罚因子,Y是拉格朗日数乘矩阵,该无约束模型最小化问题通过如下迭代过程解决:
    Figure PCTCN2019074907-appb-100050
    步骤3.1.3:迭代优化,矩阵A的优化模型为:
    Figure PCTCN2019074907-appb-100051
    为了计算方便,核范数最小化问题可以通过soft阈值算子解决,soft阈值的计算公式为(x,τ)=sgn(x)(|x|-τ) +,其中y +=max(y,0),该算子可以被用如下优化过程:
    Figure PCTCN2019074907-appb-100052
    其中,USV T是矩阵Z的奇异值分解,对
    Figure PCTCN2019074907-appb-100053
    U∈R m×r,V∈R r×n,r为矩阵的秩;
    因此,矩阵A的优化问题转化为
    Figure PCTCN2019074907-appb-100054
    同样的,矩阵E的优化问 题转化为
    Figure PCTCN2019074907-appb-100055
    步骤3.1.4:设置迭代截止条件,截止条件为:
    Figure PCTCN2019074907-appb-100056
    其中,S为权重矩阵,S权重矩阵的使用可以大大降低迭代时间,从而提高检测速度;
    步骤3.2:基于多数据融合的管道异常候选区域搜索;
    步骤3.2.1:分别对单轴数据在基于拉格朗日数乘的数据重构框架下进行异常区域搜索,得出三轴异常区域分别为Ο XYZ
    步骤3.2.2:建立三轴融合优化框架:
    Figure PCTCN2019074907-appb-100057
    步骤3.2.3:用非极大值抑制算法消除重叠,同时考虑候选区域生成的多样性,距离近的窗体合并,以二者的最大外围为新窗体外围,合并准则为:相邻窗体中心横向距离如果小于相邻窗体横向长度的最小值;
    步骤3.3基于可进化模型的管道漏磁异常识别;
    步骤3.3.1:从完备漏磁数据集中提取异常样本,建立基于卷积神经网络的异常识别模型;
    步骤3.3.2:针对那些识别错误的样本,添加新的标签,作为新的分类,转到步骤3.3.1,重新建立异常识别模型,重新进行分类,找出有缺陷的漏磁信号。
  6. 根据权利要求1所述的管道内检测漏磁数据智能分析系统,其特征在于,所述量化模块,采用基于随机森林的缺陷量化方法,得到缺陷尺寸,具体包括如下步骤:
    步骤4.1:收集数据;对缺陷漏磁信号进行检测,并将其漏磁信号进行特征提取,得到缺陷漏磁信号特征值,具体为:
    根据轴向最大通道漏磁信号上的极小值点,找到轴向最大通道漏磁信号峰谷位置及峰谷值,判断为单双峰缺陷之后,提取波形相关10个特征,分别为:单峰缺陷峰值,单峰最大峰谷差,双峰谷宽度,双峰缺陷信号左峰谷差和右峰谷差,双峰缺陷信号峰峰间距,轴向特殊点间距,面积特征,面能量特征,缺陷体积,缺陷体能量;
    10个特征具体描述如下:
    A.单峰缺陷峰值:Y v则为缺陷最小谷值,Y p-v为最大峰谷差,由于缺陷漏磁信号受到内 检测器检测环境等多种因素影响,数据的基准线波动较大,取缺陷数据的峰谷差值作为特征量可以很好地消除信号基线的影响,可以提高缺陷定量分析的可靠性;
    B.单峰最大峰谷差:其表达式为:Y p-v=Y p-Y v,式中Y p是单峰缺陷峰值,Y v则为缺陷最小谷值,Y p-v为最大峰谷差,由于缺陷漏磁信号受到内检测器检测环境等多种因素影响,数据的基准线波动较大,取缺陷数据的峰谷差值作为特征量可以很好地消除信号基线的影响,可以提高缺陷定量分析的可靠性;
    C.双峰谷宽度:用公式表示为:X v-v=X vr-X vl,其中,X v-v表示缺陷轴向信号谷宽度,X vr是缺陷右谷位置,X vl为缺陷左谷位置,缺陷信号的谷宽度能够反映出缺陷信号在轴向上的分布情况;
    D.双峰缺陷信号左峰谷差和右峰谷差:用公式表示为:Y lp-lv=Y lp-Y lv,Y rp-rv=Y rp-Y rv,式中,Y lv是漏磁信号的左谷值,Y rv是漏磁信号右谷值,Y lp为双峰信号左峰值,Y rp为双峰信号右峰值,Y lp-lv为左峰谷差,Y rp-rv为右峰谷差;
    E.双峰缺陷信号峰峰间距:其表达式为:X p-p=X pr-X pl,式中,X pr是右峰位置,X pl为左锋位置,X p-p是信号峰峰间距,缺陷信号的峰峰间距与峰谷值的结合能够大致的确定异常数据曲线的形状,有助于对缺陷长度和深度进行定量分析;
    F.轴向特殊点间距:为求缺陷长度的关键特征量,特殊点提取方法为:设置求长的比例m_RateA,根据X+(Y-X)*m_RateA求出阈值,其中,X为谷值平均值,Y为最大峰值,轴向最大通道漏磁信号中和阈值最接近的两个点即为特殊点,特殊点的间距为求缺陷长度的关键特征量;
    G.面积特征:以数值较低的谷值为基线,取两个谷之间的数据曲线与基线之间覆盖的面积,用公式表示为:
    Figure PCTCN2019074907-appb-100058
    其中,S a表示缺陷波形面积;x(t)表示缺陷信号数据点;min[x(t)]表示缺陷最小谷值;N 1表示缺陷左谷位置;N 2表示缺陷右谷位置;
    H.面能量特征:求取两个谷之间的数据曲线的能量,用公式表示为:
    Figure PCTCN2019074907-appb-100059
    式中:S e为缺陷波形面能量;
    I.缺陷体积:缺陷的体积就是在缺陷通道范围内对缺陷面积求和,用公式表示为:
    Figure PCTCN2019074907-appb-100060
    式中,V a表示缺陷体积;n 1表示向信号特殊点位置确定的起始通道;n 2表示周向信号特殊点位置确定的终止通道;S a(t)表示单条通道轴向缺陷面积;
    J.缺陷体能量:缺陷体能量就是在缺陷范围内对缺陷面能量进行求和,其表达式为:
    Figure PCTCN2019074907-appb-100061
    式中:V e表示缺陷体能量;S e(t)表示单条轴向缺陷信号面能量;
    步骤4.2:以缺陷漏磁信号特征值作为样本;人工测量缺陷尺寸作为标签,缺陷尺寸包括缺陷的深度、宽度和长度;人工选择初始训练集和测试集;
    步骤4.3:训练网络;将训练集输入初始随机森林网络中;
    步骤4.4:调整网络;通过测试集检验随机森林回归网络结果,并通过调参调整网络得到最终网络;
    步骤4.4.1:从原始漏磁信号特征缺陷样本M h×N h维中使用Bootstraping方法随机有放回采样选出m e个缺陷样本,m e≤M h,共进行T c次采样,生成T c个训练集;
    步骤4.4.2:对于T c个训练集,分别训练T c个回归树模型;
    步骤4.4.3:对于单个回归树模型,在漏磁缺陷信号特征集中,选择n e个特征,其中,n e≤N,然后每次分裂时根据信息增益比
    Figure PCTCN2019074907-appb-100062
    公式中H A(D)表示特征A的熵,g(D,A)表示其信息增益,选择信息增益比最大值的特征进行分裂,初始将参数最大特征数max_features设置为None,即不限制网络选取特征数;
    步骤4.4.4:每棵树都一直这样分裂下去,在分裂过程中为防止过拟合,考虑回归树的复杂度,对回归树进行剪枝,剪枝通过极小化损失函数C α(T)=C(T)+α|T|,其中C(T)代表模型对缺陷尺寸预测误差,即拟合程度,|T|代表模型复杂度,α用于调节回归树复杂度,使用国际海 油运输管道POF标准,将损失函数的预测误差取为POF 90%位置占壁厚的百分数;
    步骤4.4.5:模型调参优化,采用CVGridSearch网格搜索及K折交叉验证寻找最优参数,最优参数包括随机森林框架参数,袋外样本评估分数e oob和最大迭代次数,以及树模型参数最大特征数,最大深度,内部节点再划分所需最小样本数和叶子节点最少样本数;
    步骤4.4.6:将生成的多棵决策树组成随机森林,对于缺陷特征样本建立的回归问题网络,由多棵树预测值的均值决定最终预测缺陷尺寸大小;
    步骤4.5:将测试集中待测数据输入到按照步骤4.4调整的随机森林网络中,输出预测缺陷尺寸,此时若待测数据为缺陷尺寸中的深度,则输出为预测缺陷尺寸的深度;若待测数据为缺陷尺寸中的宽度,则输出为预测缺陷尺寸的宽度,若待测数据为缺陷尺寸的长度,则输出为预测缺陷尺寸的长度;其中,按照国际输油管道POF标准,其预测深度反映了按误差绝对值排序百分之80位置的值,公式为:
    Figure PCTCN2019074907-appb-100063
    其中
    Figure PCTCN2019074907-appb-100064
    分别为设计深度与预测深度。
  7. 根据权利要求1所述的管道内检测漏磁数据智能分析系统,其特征在于,所述解决方案模块,采用基于ASME B31G标准改进的管道解决方案,导入维修决策模型,输出评估结果,具体包括如下步骤:
    步骤5.1:从完备漏磁数据集中提取缺陷信息中所有的缺陷长度列、深度列以及管道属性参数;其中管道属性参数包括最小屈服强度SMYS、最小拉伸强度SMTS、公称外径D d、壁厚t a和最大允许操作压力MAOP;
    步骤5.2:计算流变应力值
    Figure PCTCN2019074907-appb-100065
    其中SMYS是管材的最小屈服强度,单位为Mpa;SMTS是最小拉伸强度,单位为Mpa;
    步骤5.3:计算管道预测失效压力
    Figure PCTCN2019074907-appb-100066
    当z<=20时,长度伸缩系数
    Figure PCTCN2019074907-appb-100067
    当z>20时,长度伸缩系数L 0=(ηz+λ a),
    Figure PCTCN2019074907-appb-100068
    腐蚀区金属损失面积
    Figure PCTCN2019074907-appb-100069
    原始面积A area0=t aL,其中d为缺陷深度,单位为mm;t a为管道壁厚,单位为mm;D d为公称外径,单位为mm;
    步骤5.4:计算管道最大失效压力
    Figure PCTCN2019074907-appb-100070
    整理得:
    Figure PCTCN2019074907-appb-100071
    当z<=20 时,θ a=2/3,当z>20时,θ a=1;
    步骤5.5:计算维修指数
    Figure PCTCN2019074907-appb-100072
    其中
    Figure PCTCN2019074907-appb-100073
    其中P为最大允许设计压力;如果维修指数ERF小于1表示缺陷可接受,大于等于1不可接受,此时应该维修或换管;
    步骤5.6:导入维修决策模型,基于专家经验及寿命预测模型进行定性和定量分析然后评估管道腐蚀的严重程度,制定维修规则,根据维修规则,输出评估结果,包括:维修指数及维修建议;其中,规则一:缺陷处的壁厚损失最大深度大于等于80%属于重大腐蚀,维修建议:应该立即维修或换管,规则二:缺陷处的ERF大于等于1,属于严重腐蚀,维修建议:应该马上维修,规则三:缺陷处的ERF值大于等于0.95且小于1.0,属于一般腐蚀,维修建议:可以观察1-3个月,规则四:缺陷处最大深度大于等于20%且小于40%属于轻微腐蚀,维修建议:可以定期观察,不做处理。
PCT/CN2019/074907 2018-12-29 2019-02-13 一种管道内检测漏磁数据智能分析系统 WO2020133639A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/345,657 US11488010B2 (en) 2018-12-29 2019-02-13 Intelligent analysis system using magnetic flux leakage data in pipeline inner inspection

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811633698.5A CN109783906B (zh) 2018-12-29 2018-12-29 一种管道内检测漏磁数据智能分析系统及方法
CN201811633698.5 2018-12-29

Publications (1)

Publication Number Publication Date
WO2020133639A1 true WO2020133639A1 (zh) 2020-07-02

Family

ID=66498905

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/074907 WO2020133639A1 (zh) 2018-12-29 2019-02-13 一种管道内检测漏磁数据智能分析系统

Country Status (2)

Country Link
CN (1) CN109783906B (zh)
WO (1) WO2020133639A1 (zh)

Cited By (39)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112130100A (zh) * 2020-09-09 2020-12-25 中国石油天然气股份有限公司 管道端面剩磁信号确定方法、装置、计算机设备及介质
CN112150410A (zh) * 2020-08-24 2020-12-29 浙江工商大学 一种焊缝缺陷的自动检测方法及其系统
CN112304255A (zh) * 2020-12-14 2021-02-02 核工业理化工程研究院 长管类组件端部跳动值检测方法及检测装置
CN112329590A (zh) * 2020-10-30 2021-02-05 中海石油(中国)有限公司 一种管道组件检测系统及检测方法
CN112336310A (zh) * 2020-11-04 2021-02-09 吾征智能技术(北京)有限公司 一种基于fcbf和svm融合的心脏疾病诊断系统
CN112417738A (zh) * 2020-12-08 2021-02-26 江苏科技大学 一种含随机坑点蚀缺陷的球形耐压壳的数值计算方法
CN112881514A (zh) * 2021-01-22 2021-06-01 北京埃彼咨能源科技有限公司 一种油气管道用漏磁成像内检测仪
CN113237951A (zh) * 2021-05-11 2021-08-10 重庆大学 一种基于形状上下文动态时间规整的金属板疲劳损伤超声导波检测方法
CN113391239A (zh) * 2021-06-10 2021-09-14 国网四川省电力公司营销服务中心 一种基于边缘计算的互感器异常监测方法与系统
CN113390956A (zh) * 2021-06-18 2021-09-14 西安建筑科技大学 双磁敏元件探头及基于探头的漏磁检测缺陷定量评价方法
CN113506440A (zh) * 2021-09-08 2021-10-15 四川国蓝中天环境科技集团有限公司 一种拉格朗日坐标系下多源数据融合的交通状态估计方法
CN113723210A (zh) * 2021-08-05 2021-11-30 中冶南方城市建设工程技术有限公司 一种应用于智慧园区给水管泄露边缘设备检测方法和系统
CN114002327A (zh) * 2021-11-05 2022-02-01 湘潭大学 一种用于钢绞线完整程度的检测方法
CN114371009A (zh) * 2022-02-28 2022-04-19 西安理工大学 基于改进随机森林的高速列车轴承故障诊断方法
CN114370984A (zh) * 2022-01-18 2022-04-19 中水东北勘测设计研究有限责任公司 一种长距离供水管道水压试验渗漏位置的智能识别方法
CN114518203A (zh) * 2022-01-26 2022-05-20 中交隧道工程局有限公司 一种盾构隧道管片渗漏点识别方法
CN114811452A (zh) * 2022-04-08 2022-07-29 国网山东省电力公司建设公司 一种变电站埋地管道泄漏监测定位方法、装置和系统
CN114926692A (zh) * 2022-06-01 2022-08-19 国家石油天然气管网集团有限公司 管道斜接角度检测方法及装置
CN114994172A (zh) * 2022-05-06 2022-09-02 北京工业大学 一种基于贝叶斯理论的超声c扫描路径优化方法
CN115099151A (zh) * 2022-07-01 2022-09-23 西南石油大学 基于gwo-bp修正机理模型的管道腐蚀速率预测方法
CN115096987A (zh) * 2022-06-24 2022-09-23 合肥工业大学 一种基于漏磁信号特征的管道缺陷量化方法
CN115180364A (zh) * 2022-08-01 2022-10-14 宁夏广天夏科技股份有限公司 基于gmi磁传感器的矿用输送带异物监测装置及方法
CN115654381A (zh) * 2022-10-24 2023-01-31 电子科技大学 一种基于图神经网络的供水管道漏损检测方法
CN116011183A (zh) * 2022-12-08 2023-04-25 中国石油大学(北京) 在役油气管道检测方法、装置、设备和存储介质
CN116384253A (zh) * 2023-04-20 2023-07-04 中国石油大学(北京) 管道漏磁检测缺陷深度预测模型建立、预测方法及装置
CN116429911A (zh) * 2023-06-13 2023-07-14 中国科学院合肥物质科学研究院 一种基于缺陷脉冲信号与图像相融合的智能化识别方法
CN116628435A (zh) * 2023-07-21 2023-08-22 山东高速股份有限公司 一种路网交通流量数据修复方法、装置、设备及介质
CN116658741A (zh) * 2023-07-28 2023-08-29 东北大学 一种管道不停输内修复装置
CN116721704A (zh) * 2023-08-03 2023-09-08 北京星云医学检验实验室有限公司 一种分级防护的生物信息数据库的更新方法及系统
CN116825243A (zh) * 2023-05-09 2023-09-29 安徽工程大学 一种基于多源数据的热障涂层服役寿命预测方法及系统
CN116956754A (zh) * 2023-09-21 2023-10-27 中石化经纬有限公司 一种结合深度学习的裂缝型漏失压力计算方法
CN117454236A (zh) * 2023-06-30 2024-01-26 哈尔滨工业大学 基于重叠信号的航天继电器多余物检测与材质识别方法
CN117490810A (zh) * 2023-12-29 2024-02-02 浙江黄氏建设科技股份有限公司 一种基于定位信息的水表远程监控方法及监控系统
CN117571815A (zh) * 2023-11-15 2024-02-20 中磁数智(北京)科技有限公司 一种供热直埋管道的预留支管位置的弱磁检测方法
CN117607243A (zh) * 2023-11-23 2024-02-27 中磁数智(北京)科技有限公司 一种用于交叉管道焊缝的磁记忆检测系统和方法
CN117635543A (zh) * 2023-11-09 2024-03-01 西安交通大学 一种金属内部缺陷正向追踪和故障反向定位方法及系统
CN117908497A (zh) * 2024-03-18 2024-04-19 山东汉峰新材料科技有限公司 基于物联网的甘宝素产线安全监测系统
CN117909886A (zh) * 2024-03-18 2024-04-19 南京海关工业产品检测中心 一种基于优化随机森林模型的锯齿棉品级分类方法及系统
CN117952983A (zh) * 2024-03-27 2024-04-30 中电科大数据研究院有限公司 一种基于人工智能的智能制造生产过程监控方法和系统

Families Citing this family (31)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110146589B (zh) * 2019-06-17 2022-12-30 东北大学 一种管道漏磁数据的高清可视化方法
TWI738016B (zh) * 2019-06-26 2021-09-01 元翎精密工業股份有限公司 液態氣體充填之智能化學習機制與自適應調節系統
CN110516589B (zh) * 2019-08-26 2023-06-02 东北大学 一种管道漏磁数据的边界精确识别方法
CN110826589A (zh) * 2019-09-10 2020-02-21 国网浙江省电力有限公司杭州供电公司 用于电力系统内基于knn算法的公司分类方法
US11415531B2 (en) * 2019-10-11 2022-08-16 KLA Corp. Statistical learning-based mode selection for multi-mode inspection
CN110929376B (zh) * 2019-10-22 2023-10-20 东北大学 一种基于gan的管道漏磁检测数据缺失的重构方法
CN110807257A (zh) * 2019-11-04 2020-02-18 中国人民解放军国防科技大学 航空发动机剩余寿命预测方法
CN110889830B (zh) * 2019-11-13 2023-03-24 河南科技大学 基于三维漏磁彩色成像的钢丝绳损伤检测方法及检测系统
CN110930377B (zh) * 2019-11-18 2022-05-13 福州大学 一种基于多任务学习的排水管道异常类型自动检测方法
CN110942450A (zh) * 2019-11-19 2020-03-31 武汉大学 一种基于深度学习的多生产流水线实时缺陷检测方法
CN110992363B (zh) * 2019-12-27 2023-05-16 中国石油大学(华东) 一种基于生成对抗网络的三维漏磁检测缺陷重构方法
CN111352617B (zh) * 2020-03-16 2023-03-31 山东省物化探勘查院 一种基于Fortran语言的磁法数据辅助整理方法
CN116194759A (zh) * 2020-05-08 2023-05-30 香港商维视拍智能检测有限公司 用于具人工智能动力的检验及预测分析的系统及方法
CN111831872A (zh) * 2020-06-09 2020-10-27 中海油能源发展装备技术有限公司 一种小缺陷特征增强的漏磁数据自适应灰度可视化方法
CN111861985B (zh) * 2020-06-09 2024-04-16 中海油能源发展装备技术有限公司 一种基于自适应模糊神经网络的漏磁缺陷深度识别方法
CN112052554B (zh) * 2020-07-23 2024-04-30 中国石油天然气集团有限公司 一种建立管道埋藏型缺陷自身高度预测模型的方法
CN112016600A (zh) * 2020-08-14 2020-12-01 中国石油大学(北京) 管道异常识别方法、装置及系统
CN111999377B (zh) * 2020-08-31 2023-04-07 河北工业大学 漏磁检测表征缺陷宽度的方法
CN112329588B (zh) * 2020-10-30 2024-01-05 中海石油(中国)有限公司 一种基于Faster R-CNN的管道故障检测方法
CN112345626B (zh) * 2020-10-30 2022-07-29 东北大学 一种基于异构场信号的管道缺陷智能反演方法
CN113076817B (zh) * 2021-03-17 2022-11-04 上海展湾信息科技有限公司 焊缝气孔缺陷实时检测方法及系统
CN113256409A (zh) * 2021-07-12 2021-08-13 广州思迈特软件有限公司 基于机器学习的银行零售客户流失预测方法
CN113899809B (zh) * 2021-08-20 2024-02-27 中海石油技术检测有限公司 基于cnn分类和rnn预测的管道内检测器定位方法
CN114199992B (zh) * 2021-12-02 2024-04-05 国家石油天然气管网集团有限公司 一种储油罐罐壁腐蚀检测方法及系统
CN114354740B (zh) * 2022-03-09 2022-05-31 成都熊谷油气科技有限公司 一种管道检测系统
CN115130496B (zh) * 2022-05-24 2023-06-16 北京化工大学 基于Bagging和RM-LOF集成单分类器的管道压力信号异常检测方法
CN115081485B (zh) * 2022-07-04 2023-04-07 中特检深燃安全技术服务(深圳)有限公司 一种基于ai的漏磁内检测数据自动分析方法
CN116822255B (zh) * 2023-08-29 2023-12-29 国网信息通信产业集团有限公司 基于自主可控软硬件平台的轻量化ai模型构建方法
CN117056864B (zh) * 2023-10-11 2024-01-23 广东力创信息技术有限公司 一种管道漏损预警方法及系统
CN117195664B (zh) * 2023-11-07 2024-01-12 北京市科学技术研究院 一种地下管线监测预警方法和系统
CN117761148A (zh) * 2024-02-22 2024-03-26 西南石油大学 一种基于振动信号与漏磁信号融合的噪声消除方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002122571A (ja) * 2000-10-12 2002-04-26 Kenzo Miya 欠陥検査方法と欠陥検査装置
CN102122351A (zh) * 2011-03-01 2011-07-13 哈尔滨工程大学 一种基于rbf神经网络的管道缺陷智能识别方法
CN106247171A (zh) * 2015-06-12 2016-12-21 宁波市鄞州磁泰电子科技有限公司 管道缺陷检测方法、管道缺陷检测装置和管道缺陷检测设备
CN106870957A (zh) * 2017-03-21 2017-06-20 东北大学 一种管道缺陷漏磁信号的特征提取方法
CN106950276A (zh) * 2017-03-21 2017-07-14 东北大学 一种基于卷积神经网络的管道缺陷深度的反演方法
CN107842713A (zh) * 2017-11-03 2018-03-27 东北大学 基于knn‑svr的海底管道漏磁数据缺失插补方法

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103196356B (zh) * 2013-04-07 2015-10-21 克拉玛依市金牛工程建设有限责任公司 基于支持向量机的油管缺陷定量识别方法
US20180196005A1 (en) * 2017-01-06 2018-07-12 Baker Hughes, A Ge Company, Llc Pipe inspection tool using colocated sensors

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002122571A (ja) * 2000-10-12 2002-04-26 Kenzo Miya 欠陥検査方法と欠陥検査装置
CN102122351A (zh) * 2011-03-01 2011-07-13 哈尔滨工程大学 一种基于rbf神经网络的管道缺陷智能识别方法
CN106247171A (zh) * 2015-06-12 2016-12-21 宁波市鄞州磁泰电子科技有限公司 管道缺陷检测方法、管道缺陷检测装置和管道缺陷检测设备
CN106870957A (zh) * 2017-03-21 2017-06-20 东北大学 一种管道缺陷漏磁信号的特征提取方法
CN106950276A (zh) * 2017-03-21 2017-07-14 东北大学 一种基于卷积神经网络的管道缺陷深度的反演方法
CN107842713A (zh) * 2017-11-03 2018-03-27 东北大学 基于knn‑svr的海底管道漏磁数据缺失插补方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LIU, JINHAI ET AL.: "MFL inner detection based defect recognition method", CHINESE JOURNAL OF SCIENTIFIC INSTRUMENT, vol. 11, no. 37, 30 November 2016 (2016-11-30), DOI: 20190905122417A *

Cited By (68)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112150410A (zh) * 2020-08-24 2020-12-29 浙江工商大学 一种焊缝缺陷的自动检测方法及其系统
CN112150410B (zh) * 2020-08-24 2023-12-12 浙江工商大学 一种焊缝缺陷的自动检测方法及其系统
CN112130100B (zh) * 2020-09-09 2024-05-28 中国石油天然气股份有限公司 管道端面剩磁信号确定方法、装置、计算机设备及介质
CN112130100A (zh) * 2020-09-09 2020-12-25 中国石油天然气股份有限公司 管道端面剩磁信号确定方法、装置、计算机设备及介质
CN112329590A (zh) * 2020-10-30 2021-02-05 中海石油(中国)有限公司 一种管道组件检测系统及检测方法
CN112329590B (zh) * 2020-10-30 2024-05-10 中海石油(中国)有限公司 一种管道组件检测系统及检测方法
CN112336310A (zh) * 2020-11-04 2021-02-09 吾征智能技术(北京)有限公司 一种基于fcbf和svm融合的心脏疾病诊断系统
CN112336310B (zh) * 2020-11-04 2024-03-08 吾征智能技术(北京)有限公司 一种基于fcbf和svm融合的心脏疾病诊断系统
CN112417738A (zh) * 2020-12-08 2021-02-26 江苏科技大学 一种含随机坑点蚀缺陷的球形耐压壳的数值计算方法
CN112417738B (zh) * 2020-12-08 2024-03-19 江苏科技大学 一种含随机坑点蚀缺陷的球形耐压壳的数值计算方法
CN112304255A (zh) * 2020-12-14 2021-02-02 核工业理化工程研究院 长管类组件端部跳动值检测方法及检测装置
CN112881514A (zh) * 2021-01-22 2021-06-01 北京埃彼咨能源科技有限公司 一种油气管道用漏磁成像内检测仪
CN112881514B (zh) * 2021-01-22 2023-08-29 北京埃彼咨能源科技有限公司 一种油气管道用漏磁成像内检测仪
CN113237951A (zh) * 2021-05-11 2021-08-10 重庆大学 一种基于形状上下文动态时间规整的金属板疲劳损伤超声导波检测方法
CN113391239A (zh) * 2021-06-10 2021-09-14 国网四川省电力公司营销服务中心 一种基于边缘计算的互感器异常监测方法与系统
CN113391239B (zh) * 2021-06-10 2024-01-23 国网四川省电力公司营销服务中心 一种基于边缘计算的互感器异常监测方法与系统
CN113390956A (zh) * 2021-06-18 2021-09-14 西安建筑科技大学 双磁敏元件探头及基于探头的漏磁检测缺陷定量评价方法
CN113390956B (zh) * 2021-06-18 2024-02-20 西安建筑科技大学 双磁敏元件探头及基于探头的漏磁检测缺陷定量评价方法
CN113723210A (zh) * 2021-08-05 2021-11-30 中冶南方城市建设工程技术有限公司 一种应用于智慧园区给水管泄露边缘设备检测方法和系统
CN113506440A (zh) * 2021-09-08 2021-10-15 四川国蓝中天环境科技集团有限公司 一种拉格朗日坐标系下多源数据融合的交通状态估计方法
CN113506440B (zh) * 2021-09-08 2021-11-30 四川国蓝中天环境科技集团有限公司 一种拉格朗日坐标系下多源数据融合的交通状态估计方法
CN114002327B (zh) * 2021-11-05 2024-02-23 湘潭大学 一种用于钢绞线完整程度的检测方法
CN114002327A (zh) * 2021-11-05 2022-02-01 湘潭大学 一种用于钢绞线完整程度的检测方法
CN114370984B (zh) * 2022-01-18 2024-01-05 中水东北勘测设计研究有限责任公司 一种长距离供水管道水压试验渗漏位置的智能识别方法
CN114370984A (zh) * 2022-01-18 2022-04-19 中水东北勘测设计研究有限责任公司 一种长距离供水管道水压试验渗漏位置的智能识别方法
CN114518203B (zh) * 2022-01-26 2024-02-20 中交隧道工程局有限公司 一种盾构隧道管片渗漏点识别方法
CN114518203A (zh) * 2022-01-26 2022-05-20 中交隧道工程局有限公司 一种盾构隧道管片渗漏点识别方法
CN114371009A (zh) * 2022-02-28 2022-04-19 西安理工大学 基于改进随机森林的高速列车轴承故障诊断方法
CN114811452A (zh) * 2022-04-08 2022-07-29 国网山东省电力公司建设公司 一种变电站埋地管道泄漏监测定位方法、装置和系统
CN114994172B (zh) * 2022-05-06 2024-06-04 北京工业大学 一种基于贝叶斯理论的超声c扫描路径优化方法
CN114994172A (zh) * 2022-05-06 2022-09-02 北京工业大学 一种基于贝叶斯理论的超声c扫描路径优化方法
CN114926692B (zh) * 2022-06-01 2023-09-05 国家石油天然气管网集团有限公司 管道斜接角度检测方法及装置
CN114926692A (zh) * 2022-06-01 2022-08-19 国家石油天然气管网集团有限公司 管道斜接角度检测方法及装置
CN115096987B (zh) * 2022-06-24 2024-04-26 合肥工业大学 一种基于漏磁信号特征的管道缺陷量化方法
CN115096987A (zh) * 2022-06-24 2022-09-23 合肥工业大学 一种基于漏磁信号特征的管道缺陷量化方法
CN115099151B (zh) * 2022-07-01 2024-06-11 西南石油大学 基于gwo-bp修正机理模型的管道腐蚀速率预测方法
CN115099151A (zh) * 2022-07-01 2022-09-23 西南石油大学 基于gwo-bp修正机理模型的管道腐蚀速率预测方法
CN115180364A (zh) * 2022-08-01 2022-10-14 宁夏广天夏科技股份有限公司 基于gmi磁传感器的矿用输送带异物监测装置及方法
CN115180364B (zh) * 2022-08-01 2024-02-27 宁夏广天夏科技股份有限公司 基于gmi磁传感器的矿用输送带异物监测装置及方法
CN115654381A (zh) * 2022-10-24 2023-01-31 电子科技大学 一种基于图神经网络的供水管道漏损检测方法
CN116011183B (zh) * 2022-12-08 2023-09-15 中国石油大学(北京) 在役油气管道检测方法、装置、设备和存储介质
CN116011183A (zh) * 2022-12-08 2023-04-25 中国石油大学(北京) 在役油气管道检测方法、装置、设备和存储介质
CN116384253B (zh) * 2023-04-20 2024-04-05 中国石油大学(北京) 管道漏磁检测缺陷深度预测模型建立、预测方法及装置
CN116384253A (zh) * 2023-04-20 2023-07-04 中国石油大学(北京) 管道漏磁检测缺陷深度预测模型建立、预测方法及装置
CN116825243A (zh) * 2023-05-09 2023-09-29 安徽工程大学 一种基于多源数据的热障涂层服役寿命预测方法及系统
CN116825243B (zh) * 2023-05-09 2024-01-16 安徽工程大学 一种基于多源数据的热障涂层服役寿命预测方法及系统
CN116429911A (zh) * 2023-06-13 2023-07-14 中国科学院合肥物质科学研究院 一种基于缺陷脉冲信号与图像相融合的智能化识别方法
CN116429911B (zh) * 2023-06-13 2023-09-01 中国科学院合肥物质科学研究院 一种基于缺陷脉冲信号与图像相融合的智能化识别方法
CN117454236A (zh) * 2023-06-30 2024-01-26 哈尔滨工业大学 基于重叠信号的航天继电器多余物检测与材质识别方法
CN116628435A (zh) * 2023-07-21 2023-08-22 山东高速股份有限公司 一种路网交通流量数据修复方法、装置、设备及介质
CN116628435B (zh) * 2023-07-21 2023-09-29 山东高速股份有限公司 一种路网交通流量数据修复方法、装置、设备及介质
CN116658741A (zh) * 2023-07-28 2023-08-29 东北大学 一种管道不停输内修复装置
CN116658741B (zh) * 2023-07-28 2023-09-26 东北大学 一种管道不停输内修复装置
CN116721704A (zh) * 2023-08-03 2023-09-08 北京星云医学检验实验室有限公司 一种分级防护的生物信息数据库的更新方法及系统
CN116721704B (zh) * 2023-08-03 2023-10-20 北京星云医学检验实验室有限公司 一种分级防护的生物信息数据库的更新方法及系统
CN116956754B (zh) * 2023-09-21 2023-12-15 中石化经纬有限公司 一种结合深度学习的裂缝型漏失压力计算方法
CN116956754A (zh) * 2023-09-21 2023-10-27 中石化经纬有限公司 一种结合深度学习的裂缝型漏失压力计算方法
CN117635543B (zh) * 2023-11-09 2024-04-30 西安交通大学 一种金属内部缺陷正向追踪和故障反向定位方法及系统
CN117635543A (zh) * 2023-11-09 2024-03-01 西安交通大学 一种金属内部缺陷正向追踪和故障反向定位方法及系统
CN117571815A (zh) * 2023-11-15 2024-02-20 中磁数智(北京)科技有限公司 一种供热直埋管道的预留支管位置的弱磁检测方法
CN117607243A (zh) * 2023-11-23 2024-02-27 中磁数智(北京)科技有限公司 一种用于交叉管道焊缝的磁记忆检测系统和方法
CN117490810A (zh) * 2023-12-29 2024-02-02 浙江黄氏建设科技股份有限公司 一种基于定位信息的水表远程监控方法及监控系统
CN117490810B (zh) * 2023-12-29 2024-03-22 浙江黄氏建设科技股份有限公司 一种基于定位信息的水表远程监控方法及监控系统
CN117909886A (zh) * 2024-03-18 2024-04-19 南京海关工业产品检测中心 一种基于优化随机森林模型的锯齿棉品级分类方法及系统
CN117908497A (zh) * 2024-03-18 2024-04-19 山东汉峰新材料科技有限公司 基于物联网的甘宝素产线安全监测系统
CN117909886B (zh) * 2024-03-18 2024-05-24 南京海关工业产品检测中心 一种基于优化随机森林模型的锯齿棉品级分类方法及系统
CN117908497B (zh) * 2024-03-18 2024-06-07 山东汉峰新材料科技有限公司 基于物联网的甘宝素产线安全监测系统
CN117952983A (zh) * 2024-03-27 2024-04-30 中电科大数据研究院有限公司 一种基于人工智能的智能制造生产过程监控方法和系统

Also Published As

Publication number Publication date
CN109783906A (zh) 2019-05-21
CN109783906B (zh) 2023-07-07

Similar Documents

Publication Publication Date Title
WO2020133639A1 (zh) 一种管道内检测漏磁数据智能分析系统
US11488010B2 (en) Intelligent analysis system using magnetic flux leakage data in pipeline inner inspection
CN113239504B (zh) 一种基于优化神经网络的管道腐蚀缺陷预测方法
Liu et al. XGBoost algorithm-based prediction of safety assessment for pipelines
CN110470729B (zh) 一种基于电涡流的油田套管缺陷无损检测方法
Ma et al. Advances in corrosion growth modeling for oil and gas pipelines: A review
Xu et al. The research progress and prospect of data mining methods on corrosion prediction of oil and gas pipelines
CN114169374B (zh) 一种斜拉桥斜拉索损伤识别方法及电子设备
CN116123988B (zh) 一种基于数据处理的管道几何变形检测方法及系统
CN116258086B (zh) 一种燃气管道风险评估方法及系统
Zhang et al. Data recovery of magnetic flux leakage data gaps using multifeature conditional risk
CN105975678B (zh) 一种基于参数化模型的油气管道剩余强度预测方法
Zhang et al. Deep emulational semi-supervised knowledge probability imaging method for plate structural health monitoring using guided waves
CN116777223A (zh) 一种城市地下管网安全综合风险评估方法及系统
CN113343402B (zh) 基于多层卷积稀疏编码的管道腐蚀等级评估方法
CN115791969A (zh) 基于声发射信号的导管架水下裂纹检测系统及其方法
Li et al. A data-driven methodology for predicting residual strength of subsea pipeline with double corrosion defects
CN113111923A (zh) 一种基于一维迁移学习卷积神经网络集成模型的供水管网泄漏检测与定位方法
Jarmulak et al. Case-based reasoning for interpretation of data from non-destructive testing
CN116050564A (zh) 基于尾矿库坝体位移预测的算法流程
CN115512717A (zh) 一种基于集成学习的二阶段单传感器管道泄漏定位方法
CN114065624A (zh) 一种炼化装置管网完整性管理水平智能评估方法
Ferreira et al. Corrosion-like defect severity estimation in pipelines using convolutional neural networks
Huang et al. Residual swin transformer-based weld crack leakage monitoring of pressure pipeline
Lingyun et al. A novel procedure to evaluate the performance of failure assessment models

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19903978

Country of ref document: EP

Kind code of ref document: A1