US20230314373A9 - System and method for automated acquisition and analysis of electromagnetic testing data - Google Patents

System and method for automated acquisition and analysis of electromagnetic testing data Download PDF

Info

Publication number
US20230314373A9
US20230314373A9 US17/975,314 US202217975314A US2023314373A9 US 20230314373 A9 US20230314373 A9 US 20230314373A9 US 202217975314 A US202217975314 A US 202217975314A US 2023314373 A9 US2023314373 A9 US 2023314373A9
Authority
US
United States
Prior art keywords
test data
indications
landmarks
data
machine learning
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
US17/975,314
Other languages
English (en)
Other versions
US20230138975A1 (en
Inventor
Philippe MACKAY
Vincent GAUDREAULT
Florian Hardy
Marco Michele Sisto
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Eddyfi Canada Inc
Original Assignee
Eddyfi Canada Inc
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 Eddyfi Canada Inc filed Critical Eddyfi Canada Inc
Priority to US17/975,314 priority Critical patent/US20230314373A9/en
Assigned to EDDYFI CANADA INC. reassignment EDDYFI CANADA INC. ASSIGNMENT OF ASSIGNORS INTEREST (SEE DOCUMENT FOR DETAILS). Assignors: EDDYFI NDT INC.
Assigned to EDDYFI NDT INC. reassignment EDDYFI NDT INC. ASSIGNMENT OF ASSIGNORS INTEREST (SEE DOCUMENT FOR DETAILS). Assignors: MACKAY, PHILIPPE, GAUDREAULT, VINCENT, HARDY, FLORIAN, SISTO, MARCO MICHELE
Publication of US20230138975A1 publication Critical patent/US20230138975A1/en
Publication of US20230314373A9 publication Critical patent/US20230314373A9/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/72—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables
    • G01N27/82—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws
    • G01N27/90—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws using eddy currents
    • G01N27/9046—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws using eddy currents by analysing electrical signals
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/72—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables
    • G01N27/82—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N27/00—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means
    • G01N27/72—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables
    • G01N27/82—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws
    • G01N27/90—Investigating or analysing materials by the use of electric, electrochemical, or magnetic means by investigating magnetic variables for investigating the presence of flaws using eddy currents
    • G01N27/9073—Recording measured data
    • G01N27/9086—Calibrating of recording device

Definitions

  • the technical field generally relates to non-destructive testing, and more specifically to systems and methods for testing ferromagnetic and non-ferromagnetic tubes, which are often inspected by inserting probes in the tube.
  • NDT Non-destructive testing
  • Eddy Current Testing ECT
  • ECA Eddy Current Arrays
  • RFT Remote Field Testing
  • RFA Remote Field Arrays
  • IRIS Internal Rotary Inspection Systems
  • PSECT Partially Saturated Eddy Current Testing
  • FSECT Fully Saturated Eddy Current Testing
  • MFL Magnetic Flux Leakage
  • NFT Near Field Testing
  • NFA Near Field Arrays
  • PEC Pulsed Eddy Current
  • the inspection is performed by capturing electromagnetic/acoustic “raw” data with probes and instruments.
  • the data is then normalized and analyzed by a human analyst to identify all interesting indications. Both the acquisition and the analysis processes can be long and error prone. There is therefore a need in the industry for solutions to assist probe operators in the acquisition of good quality data and assist analysts in the detection and classification of indications.
  • a method for identifying landmarks in an object from signals captured by a probe during non-destructive testing of the object.
  • the method includes: processing the signals using a machine learning algorithm to identify a plurality of predetermined landmark types, the algorithm being configured to output a sequence of connected regions in the test data corresponding to sequences of the plurality of data points likely corresponding to a same type of the plurality of predetermined landmark types; generating a sequence of identified landmarks by concatenating each of the connected regions; comparing the sequence of identified landmarks to a model defining an expected sequence of landmarks in the object, and identifying a subset of identified landmarks in the sequence of identified landmarks that fits best with the model; and outputting a landmark sequence corresponding to the subset of identified landmarks.
  • a method for identifying indications in an object via non-destructive testing includes: recording test data from a probe operated to scan the object, the test data comprising a plurality of data points, each data point corresponding to a signal measurement acquired by the probe; processing the test data using a first machine learning algorithm trained to identify a plurality of predetermined landmark types, the first algorithm being configured to output a sequence of identified landmarks and to identify regions in the object based on the landmarks; processing the test data and the identified object regions using a second machine learning algorithm trained to segment indications in the identified regions of the object from the test data, the second neural network being configured to output a list of detected indications; for each indication in the list of detected indications, applying a predefined decision tree to either discard the indication or classify the indication according to one of a plurality of predefined indication types; and outputting the classified indications in a report specifying positions of the classified indications relative to at least some of the identified landmarks.
  • a method for identifying landmarks in an object from signals captured by a probe during non-destructive testing of the object.
  • the method includes: processing the signals using a neural network trained to identify a plurality of predetermined landmark types, the algorithm being configured to output a sequence of connected regions in the test data corresponding to sequences of the plurality of data points likely corresponding to a same type of the plurality of predetermined landmark types; generating a sequence of identified landmarks by concatenating each of the connected regions; comparing the sequence of identified landmarks to a model defining an expected sequence of landmarks in the object, and identifying a subset of identified landmarks in the sequence of identified landmarks that fits best with the model; and outputting a landmark sequence corresponding to the subset of identified landmarks.
  • a method for identifying indications in an object via non-destructive testing includes: recording test data from a probe operated to scan the object, the test data comprising a plurality of data points, each data point corresponding to a signal measurement acquired by the probe; processing the test data using a first machine learning algorithm trained to identify a plurality of predetermined landmark types, the first algorithm being configured to output a sequence of identified landmarks and to identify regions in the object based on the landmarks; processing the test data and the identified object regions using a second machine learning algorithm trained to segment indications in the identified regions of the object from the test data, the second neural network being configured to output a list of detected indications; for each indication in the list of detected indications, applying a predefined decision tree to either discard the indication or classify the indication according to one of a plurality of predefined indication types; and outputting the classified indications in a report specifying positions of the classified indications relative to at least some of the identified landmarks.
  • a method for identifying indications in an object via non-destructive testing using inspection equipment comprising a probe.
  • the method includes: recording test data from the probe while the probe is operated to scan the object, the test data comprising a plurality of data points, each data point corresponding to a signal measurement acquired by the probe; processing the test data using a first analysis machine learning algorithm trained to detect landmarks in the test data and output a list of detected landmarks; processing the list of detected landmarks to identify regions in the object based on the landmarks; processing the test data and the identified regions using a second analysis machine learning algorithm trained to detect indications in the test data and output a list of detected indications; processing the list of indications to automatically classify each indication according to one of a plurality of predefined indication types; and outputting the classified indications in a report specifying positions of the classified indications in the object.
  • a system for identifying indications in an object via non-destructive testing includes: a probe; a recording device in operative communication with the probe, the recording device comprising a storage module configured to record and store test data from the probe while the probe is operated to scan the object, the test data comprising a plurality of data points, each data point corresponding to a signal measurement acquired by the probe; and an analysis device configured to access the test data stored by the recording device, the analysis device comprising a test data analysis module configured to: process the test data using a first analysis machine learning algorithm trained to detect landmarks in the test data and output a list of detected landmarks; process the list of detected landmarks to identify regions in the object based on the landmarks; process the test data and the identified regions using a second analysis machine learning algorithm trained to detect indications in the test data and output a list of detected indications; process the list of indications to automatically classify each indication according to one of a plurality of predefined indication types; and output the classified indications in a report
  • a non-transitory computer-readable medium has instructions stored thereon to identify indications in an object via non-destructive testing using inspection equipment comprising a probe.
  • the instructions when executed by one or more processors, cause the one or more processors to: record test data from the probe while the probe is operated to scan the object, the test data comprising a plurality of data points, each data point corresponding to a signal measurement acquired by the probe; process the test data using a first analysis machine learning algorithm trained to detect landmarks in the test data and output a list of detected landmarks; process the list of detected landmarks to identify regions in the object based on the landmarks; process the test data and the identified regions using a second analysis machine learning algorithm trained to detect indications in the test data and output a list of detected indications; process the list of indications to automatically classify each indication according to one of a plurality of predefined indication types; and output the classified indications in a report specifying positions of the classified indications in the object.
  • FIGS. 1 A and 1 B illustrate a heat exchanger having tubes that can be inspected using NDT.
  • FIG. 2 is a flowchart illustrating a method for acquiring and analyzing test data during NDT, according to an embodiment.
  • FIG. 3 is a schematic illustrating inspection equipment that can be used to conduct NDT, according to an embodiment.
  • FIG. 4 is a flowchart illustrating subprocesses for acquiring and validating test data in the method of FIG. 2 .
  • FIG. 5 is a flowchart illustrating a subprocess for assisted calibration of inspection equipment, according to an embodiment.
  • FIG. 6 is a flowchart illustrating subprocesses for analyzing test data in the method of FIG. 2 .
  • FIG. 7 is a flowchart illustrating a method for automatically predicting normalization coefficients, according to an embodiment.
  • FIG. 8 is a flowchart illustrating a subprocess for assisted landmark detection, according to an embodiment.
  • FIG. 9 is a schematic illustrating an exemplary directed graph that can be output by the subprocess for assisted landmark detection of FIG. 8 .
  • FIG. 10 is a flowchart illustrating a method for automatically detecting indications using a neural network, according to an embodiment.
  • FIG. 11 is a schematic illustrating an exemplary decision tree for labelling indications detected according to the method of FIG. 10 .
  • FIG. 12 is a flowchart illustrating subprocesses for preparing for inspection in the method of FIG. 2 .
  • FIGS. 13 A- 13 E are examples of extracted ECT signals represented on a complex plane.
  • an exemplary object 100 (also referred to as an asset) that can be inspected using NDT is shown.
  • the object 100 corresponds to a heat exchanger used during the processing of oil and gas.
  • the heat exchanger comprises a plurality of metallic tubes 101 .
  • the interior of the metallic tubes 101 can be inspected to identify indications, such as defects including corrosion or pitting, among others.
  • electromagnetic probes 103 are inserted into the tubes 101 , and electromagnetic signals measured by the probes 103 can be recorded as raw test data as the probes 103 are moved along the length of the tube.
  • the raw test data can be processed and analyzed in order to identify, characterize and locate indications in the metallic tubes 101 .
  • a report can subsequently be generated and delivered to the asset owner.
  • the report can specify the health status of the heat exchanger as well as any indications revealed by the inspection.
  • Such indications can include, for example: landmarks (i.e. locations of predefined structural elements within the heat exchanger, such as baffles 105 and tube sheets 107 ); defects with loss of metal (ex: corrosion, pitting, etc.); defects without loss of metal (ex: dents, bulges, etc.); presence of magnetic deposit in exchanger tubes; etc.
  • the location of defects identified in the tubes 101 can be indicated. In some embodiments, the locations of defects can be indicated according to their position relative to landmarks.
  • the method 200 can include a first step 202 of providing inspection equipment at an inspection site.
  • the inspection equipment that is provided can vary depending on the nature of the object that is to be inspected and the type of inspection to be carried out. For example, an inspection contract can be agreed upon with a client and the appropriate inspection equipment can be selected. Inspection staff and equipment can then be scheduled before eventually travelling and shipping the equipment to the site where the inspection is to be carried out, and where the equipment can be prepared to conduct the inspection.
  • inspection equipment 300 can include at least an inspection instrument 302 in operative communication with a recording device 304 .
  • the inspection instrument 302 comprises an electromagnetic probe
  • the recording device 304 comprises a portable computing system coupled to the probe.
  • the portable computing system can comprise various software and/or hardware modules to assist in the acquisition of test data and the analysis thereof.
  • the computing device comprises a processor and memory storing computer-executable instructions which, when executed by the processor, cause the processor to implement a calibration module 306 for calibrating the inspection equipment 300 , a storage module 308 for storing signals measured via recording device 304 , and a validation module 310 for validating recorded data.
  • the computing device further includes a user interface module, for example comprising a touchscreen display 312 and buttons 314 , allowing a user to interact with the computing device and control the inspection equipment 300 as needed. It is appreciated, however, that other hardware and/or software modules can be provided.
  • subsequent steps in the method 200 can comprise preparing for inspection 204 , and operating the inspection equipment 300 to acquire and validate test data 206 .
  • the acquired testing data can then be analyzed 208 in order to identify, characterize and locate indications in the metallic tubes of the heat exchanger.
  • the analysis can be subject to review and validation 210 before a report is delivered to the client 212 .
  • a first subprocess 1200 can comprise preparing inspection setup. This can include guiding a user (ex: a probe operator) in the setup definition by requesting user input on some basic parameters and suggesting the appropriate instrument settings and setup. In an embodiment, this can be done via rule-based algorithms.
  • a configuration wizard can be provided for allowing selection of parameters such as inspection frequencies from tube material and thickness, using known-in-the-art rules and algorithms.
  • a second subprocess 1202 can comprise assisting the preparation of the inspection folder and tube list by creating a tube list automatically. In an embodiment, this can be accomplished by analyzing a photo of the exchanger tube sheet, with visual recognition of each tube. By way of example, machine learning or deep learning methods can be applied for image analysis and tube recognition.
  • the output of the analysis can comprise a tube list with tube rows and columns numbered to represent the exchanger and tube layout.
  • a first subprocess of the data acquisition and validation 206 step can comprise calibrating the inspection equipment 400 .
  • This subprocess can, for example, be carried out via calibration module 306 of the inspection equipment 300 .
  • the inspection equipment 300 can be calibrated by using the inspection instrument 302 (i.e. the electromagnetic probe) to record measurements from a reference object, such as a reference tube having known defects or other indications. More specifically, in the present embodiment, calibration data can be recorded from electromagnetic signals measured by the probe as the probe is moved along the interior of the reference tube. A calibration algorithm can then be applied to the captured calibration data to extract calibration parameters therefrom. In order for such algorithms to be applicable, specific signatures in the calibration data need to be identified, such as signatures corresponding to known indications in the reference tube (for example, a signature corresponding to a hole or other reference defect in the reference tube).
  • the calibration subprocess 400 can be substantially automated by applying machine learning.
  • the recorded calibration data can be analyzed via a calibration machine learning algorithm, such as a deep learning neural network trained on historical calibration data with signatures labelled according to a plurality of possible indication types.
  • the neural network can thus segment signatures in the calibration data and classify the segmented signatures according to one of the plurality of indication types.
  • the classified reference signatures can then be provided to the calibration algorithms to allow those algorithms to extract the required calibration parameters.
  • a neural network is described, it is appreciated that other machine learning algorithms can also be used.
  • the neural network can be trained to output a status code which can assist in indicating whether the process of identifying reference signatures has succeeded or failed.
  • One or more different methods can be applied to output the status code, such as: applying a rule-based validation of the segmented and classified signatures to identify predetermined error conditions (for example, if the segmented signature does not match a predefined mask, a status error can be generated); estimating the confidence level of the neural network classification/prediction using any suitable technique, such as Bayesian confidence estimation, estimation of confidence via ensemble networks, etc.
  • the calibration subprocess 400 can further include other suitable steps, such as: probe balancing or nulling; array balancing or normalization; and material properties compensation, such as adjusting an offset, scale and rotation in remote field testing to compensate for the difference in material properties between a reference tube and the tube to be inspected.
  • a subsequent subprocess 402 can comprise operating the inspection equipment 300 to capture test data.
  • the test data collection process can be carried out in any suitable manner depending on the object being expected.
  • the heat exchanger being inspected comprises a plurality of straight tubes with two open ends opposite one another, referred to as a proximal end and a distal end.
  • a probe operator can insert the probe in the tube via the opening in the proximal end and push the probe along the tube without recording data, until the distal end of the tube is reached.
  • the data recording can be started, and the probe can be pulled along the tube at constant speed back towards the proximal end.
  • the data recording can be stopped when the probe exits the tube again at the proximal end.
  • the process can be repeated for each tube in the heat exchanger, such that test data is captured for each tube.
  • the captured test data can be subject to a validation subprocess 404 in order to ensure the quality thereof.
  • the validation subprocess 404 can, for example, be carried out via the validation module 310 of the inspection equipment 300 .
  • the validation subprocess 404 comprises a combination of one or more rules applied on the raw test data to identify potential problems or errors in the capture process.
  • data quality validation (DQV) rules can verify if the inspection instrument returns invalid or saturated raw data.
  • a DQV rule can fail if the expected landmarks are not found in the data, which could indicate that the tube was not fully inspected.
  • this DQV rule can take advantage of the analysis capabilities of the AI-based landmark detection subprocess described below.
  • a DQV rule can measure the length of the data array captured in a tube and compare it with a theoretical expected length. In a time-encoded data recording system, this rule could fail if the probe was moved too quickly in the tube.
  • the setup and calibration of the inspection system can also be subject to a validation subprocess 405 to confirm the quality of the calibration and ensure that the inspection setup is properly adjusted.
  • the validation subprocess 405 can comprise a plurality of checks, such as: validating that all required inspection frequencies are activated; validating that the inspection frequencies are properly arranged in view of the setup settings, such as tube thickness and materials; validating that the calibration operation does not cause the system to saturate internal gain; validating that all setup parameters are congruent and valid; etc.
  • the step of data analysis 208 comprises seven subprocesses, namely calibration adjustment for analysis 600 , data enhancement 602 , assisted landmark detection 604 , identification of tube regions 606 , assisted indication detection 608 , assisted indication classification 610 and assisted indication sizing 612 .
  • these subprocesses can be carried out by an analyst, for example on an analysis device comprising a processor and memory storing computer executable instruction which, when executed, cause the processor to implement a test data analysis module configured to carry out one or more of the subprocesses.
  • the analysis subprocesses can be performed on the data recording device 304 used to record the test data and as such, the analysis device can comprise the recording device 304 . In other embodiments, the analysis subprocesses can be performed on a separate device.
  • the test data can be transferred to and/or accessed from an analysis device that is separate and distinct from the recording device 304 (such as an analyst workstation or other computing device), and the test data can be processed via the analysis device.
  • the calibration adjustment for analysis 600 subprocess can comprise calibrating and adjusting other software settings prior to analysis to optimize the analysis environment.
  • simple adjustments can be made, such as adjusting analysis views in the user interface, adjusting color palettes or contrast in 2D views, etc.
  • more complex adjustments may be required. Such adjustments may need to be applied to every tube or globally for the whole inspection.
  • the test data may need to be subject to a form of normalization to account for small variations of material properties (ex: conductivity and permeability) from one tube to another.
  • RFT Remote-Field Testing
  • Such normalization can correct signal offsets, scaling and rotation of test data acquired from RFT signals as compared to calibration data acquired from reference tubes when the inspection equipment was calibrated.
  • normalization coefficients must be calculated (such as offsets, scaling and rotation), and those coefficients can then be applied to the test data acquired form measured RFT signals.
  • the normalization coefficients can be calculated using algorithms that operate on defined normalization features.
  • the test data acquired from RFT signals can be manually analyzed to identify such normalization features, which can include, for example: a region of nominal tube wall thickness; the signal phase rotation associated with a landmark such as a support plate or a tube sheet; etc.
  • These normalization features can then be fed to a normalization algorithm that calculates the normalization coefficients (such as offsets, scaling and rotation) therefrom.
  • the normalization coefficients can be automatically predicted or estimated without explicitly identifying normalization features.
  • a process 700 for automatically predicting normalization coefficients is shown in FIG. 7 .
  • raw test data acquired from RFT signals is provided to a normalization machine learning algorithm, such as a deep learning neural network.
  • the neural network is trained to predict normalization coefficients from the raw data, for example using regression algorithms trained on historical test data and with corresponding normalization coefficients.
  • the predicted normalization coefficients can then be applied to the test data in the same manner described above.
  • a neural network is described, it is appreciated that other machine learning algorithms can also be used.
  • the neural network can also provide a status code, for example to provide an indication on the success or failure of the prediction of coefficients.
  • a status code can be implemented via a combination of the following methods that estimate uncertainty of regression and output a status code of failure if the uncertainty is above a predetermined threshold: rule-based coefficients validity check, where too large scaling or offset could either indicate large magnetic property variations or low confidence regression; estimation of the confidence level of the neural network prediction via any suitable technique, such as bayesian confidence estimation, estimation of confidence via ensemble networks, analysis of the prediction entropy, etc.
  • remedial action can be taken in response to outputting a failure status code. For example, large magnetic property variations may require data recapturing with adjustment of inspection frequency.
  • the data enhancement 602 subprocess can comprise processing the raw test data with filters to improve the quality of the raw data and enhance interesting features in the data. Any suitable techniques can be applied, such as: noise filtering; low, median or high pass filtering; dynamic range adjustments, etc.
  • test data can subsequently be analyzed to identify tube landmarks in the test data.
  • This can include, for example, identifying signal signatures corresponding to mechanical support structures in the tubes.
  • the identified landmarks can later be used to estimate an absolute positioning of defects and other indications in the tubes.
  • landmarks are expected tube features and can thus be distinguished from defects.
  • an assisted landmark detection 604 subprocess is provided to substantially automate the detection and identification of landmarks in the test data.
  • manual detection and segmentation landmarks in the data is time consuming and error prone.
  • Automatic landmark segmentation can thus reduce the analysis time and allow an analyst to focus on identifying defects.
  • the list of detected landmarks may need to be filtered.
  • Such a filtering operation aims to select a proper sequence of landmarks, which may be a subset of the detected landmarks.
  • the filtered landmarks sequence can isolate a portion of the signal corresponding to a region of the tube that should be analyzed for defects and exclude other portions of signal that are not pertinent for the analysis.
  • the recorded data for a straight tube should start with a short section of air (measured at the distal end of the tube), followed by a first landmark called a tube sheet (TS), a series or landmarks called support plates (SP), another tube sheet and air measured at the proximate end of the tube.
  • TS tube sheet
  • SP support plates
  • the landmark sequence after detection would be AIR-TS-SP-SP . . . -SP-TS-AIR.
  • the operator may start the recording while still pushing the probe toward the distal end of the tube.
  • the landmark sequence would show some SP detected (while pushing) before the TS: SP-SP- . . . -SP-TS-AIR-TS-SP-SP . . . -SP-TS-AIR.
  • the underlined portion of this landmark sequence corresponds to the section where the probe is being pulled out of the tube, and the portion of recorded data that will need to be analyzed.
  • Landmark filtering thus aims to isolate landmark sequences corresponding to data capture while pulling the probe out of the tube. Other types of recording errors are possible, requiring a similar landmark filtering operation.
  • the subprocess 604 implements a two-stage method to automatically detect and identify landmarks from signals captured during non-destructive testing of an object. In a first stage, the position and type of the landmarks is detected, and in a second stage sequences of detected landmarks are filtered.
  • a first step of the method can comprise receiving raw test data.
  • the object under inspection corresponds to a metallic tube.
  • the raw test data corresponds to electromagnetic signals captured by a probe moving along the interior of the metallic tube during non-destructive testing thereof.
  • the input test data may need to be calibrated and/or normalized. Accordingly, initial conditioning of the input signals can be carried out in step 800 . This can include, for example, scaling and offset operations to normalize the signals statistical mean and variance values.
  • the test data can be processed by a first analysis machine learning algorithm 802 trained to identify a plurality of predetermined landmark types.
  • the first analysis machine learning algorithm 802 is a neural network which can, for example, be trained on historical or previously acquired test data labelled according to landmarks identified therein.
  • the neural network 802 can be configured to segment landmarks 804 in the test data and output, for each of a plurality of data points in the test data, a classification of at least one of the plurality of predetermined landmarks types to which the data point corresponds.
  • the output of 804 can comprise curves expressing, for each landmark type, a probability that each data point in the test data is classified as that type of landmarks.
  • the neural network structure can be particularly adapted for segmenting landmarks.
  • the neural network can comprise a convolutional neural network architecture in a U-net multi-class classifier configuration.
  • a neural network is described, it is appreciated that other machine learning algorithms can be used as first machine learning algorithm.
  • the classified test data can subsequently be grouped in order to identify start-stop ranges in the test data corresponding to identified landmarks. More specifically, connected regions in the classified test data can be identified, with such connected regions corresponding to sequences of data points likely corresponding to a same landmark. A first data point in each connected region can indicate a start of a range of data points corresponding to a landmark, and a last data point of each connected region can indicate an end of the range of data points corresponding to that landmark.
  • the defined ranges can subsequently be labelled according to the landmark to which they correspond.
  • any suitable process can be applied to group and label sequences of data points in the test data classified by the neural network.
  • the curves output by 804 can be transformed into a sequence of start-stop ranges for all detected landmarks using a connected component labelling (CCL) algorithm 808 .
  • CCL connected component labelling
  • the first neural network 802 is configured to identify landmarks via semantic segmentation 804 , it is appreciated that other configurations are possible.
  • the first neural network 802 can comprise an object detection machine learning algorithm trained to build an object detection model such that the neural network 802 is configured to directly output an array of detected landmarks, i.e. comprising an array of start positions and end positions, corresponding to the start and end of each landmark.
  • the neural network 802 can be a convolutional neural network trained on historical or previously acquired test data having groups of data points labelled according to landmarks identified therein.
  • the neural network 802 can be configured to also extract additional features for each landmark 806 , which can be used to further qualify and validate the landmarks.
  • additional feature can comprise a level of confidence associated to the prediction of each landmark by the neural network 802 .
  • additional features can be calculated outside the neural network using conventional algorithms 807 . Such algorithms include statistical values such as variance and mean value of the signal associated to a landmark.
  • the neural network 802 can be configured to output a status code 818 to provide an indication of whether the landmark segmentation 804 and/or extraction of landmark validation features 806 is successful.
  • One or more different methods can be applied to output the status code, such as: applying a rule-based validation of the segmented and classified signatures to identify predetermined error conditions (for example, if the segmented signature does not match a predefined mask, a status error can be generated); estimating the confidence level of the neural network prediction using any suitable technique, such as Bayesian confidence estimation, estimation of confidence via ensemble networks, etc. (for example if the confidence level is below a predetermined threshold, a status error can be generated); analyzing the prediction entropy; and/or analyzing prediction curves using techniques such as those described in Devries et al.
  • a sequence of landmarks can be extracted from the grouped test data. More specifically, the grouped test data can be modelled as a chain of identified landmarks, and pattern recognition techniques can be applied in order to identify patterns within the chain corresponding to an expected sequence of landmarks (ex: a sequence of AIR-TS-SP-SP . . . -SP-TS-AIR which corresponds to the expected pattern when the probe is being pulled out of the tube). Landmarks identified in the test data that match the pattern can be retained, while landmarks that do not match the pattern can be discarded. In this fashion, an isolated landmark sequence is extracted, corresponding to data captured while pulling the probe out of the tube.
  • a linear-chain conditional random field (CRF) modelling method can be used. More specifically, the start-stop ranges and the additional features calculated for each landmark can be concatenated and provided as an input to the linear-chain CRF method, and the CRF can be configured to model the landmark sequences using a directed graph. Once provided with the detected landmark sequence, the CRF can find a subset of landmarks that best fit a target graph (ex: a graph representing a sequence pattern MR-TS-SP-SP . . . -SP-TS-MR) while considering both the landmark position and the sequence.
  • a target graph ex: a graph representing a sequence pattern MR-TS-SP-SP . . . -SP-TS-MR
  • the subset of landmarks can be retained while the other landmarks and their corresponding test data can be rejected.
  • the subset of landmarks typically contains consecutive landmarks. This step can allow identifying the group of consecutive landmarks that best identify the inspected tube, leaving out any set of additional landmarks captured before and after the tube. These additional landmarks may stem from bad data capture practices, for example if the recording starts too early or ends too late. This step can also remove isolated landmarks in the middle of the good sequence, because those landmarks are identified as errors.
  • the pattern recognition can be followed by a validation step 812 to validate the extracted sequence of landmarks.
  • a validation step 812 to validate the extracted sequence of landmarks.
  • one or more checks can be carried out to confirm the pattern recognition behaves and/or performs as expected.
  • the check can comprise: validating that the pattern recognition has rejected/discarded a small fraction of the detected landmarks (ex: that the number of rejects landmarks is below a predetermined threshold and/or percentage relative to the total number of identified landmarks); validating that the test data rejected/discarded by the pattern recognition corresponds to a small fraction of the input test data (ex: that the number of data points rejects is below a predetermined threshold and/or percentage relative to a total number of data points in the input test data); etc.
  • the validation step 812 can output a corresponding status code 820 . For example, if one or more validations fail, an error status code can be output to inform a user/analyst that the automatic analysis should be manually validated.
  • the validated landmark sequence 816 can subsequently be output, for example modelled as a directed graph.
  • An exemplary directed graph for an exchanger tube 900 is shown in FIG. 9 .
  • the landmark sequence 816 can be output as an array of validated landmarks with start stop positions of data points in the test data, such as:
  • a subsequent subprocess 606 can comprise identifying or defining analysis regions in the tube using the validated landmark sequence.
  • an “in-tube” region can be defined to identify the whole tube
  • a “near landmark” region can be identified to define parts of the tube that are in close proximity to landmarks (ex: within a predetermine distance relative to landmarks)
  • a “free span” region can be defined to identify parts of the tube far from landmarks (ex: further away than a predetermined distance relative to landmarks)
  • the defined regions can subsequently be used to fine tune automated detection of indications.
  • indications can be detected using machine learning employing classification algorithms, and such classification algorithms can use different rules or settings depending on the region in which the indication is to be detected.
  • an assisted indication detection 608 subprocess is provided to substantially automate the detection of indications.
  • a neural network is provided to automatically detect possible indications via segmentation, although it is appreciated that other indication detection techniques are possible.
  • manual detection of indications can be time consuming and error prone.
  • rule-based assisted indication detection can be difficult to configure, complex to operate, and not well adapted to the data quality and signals typically encountered in oil and gas inspections. Neural network-assisted automatic indication detection can thus reduce analysis time and provide more accurate results when inspection objects such as an oil and gas heat exchanger.
  • a neural network is described, it is appreciated that other machine learning algorithms can also be used.
  • FIG. 10 An exemplary method for automatically detecting indications using a neural network is shown in FIG. 10 .
  • the previously defined tube regions and the calibrated test data are provided as an input to a second analysis machine learning algorithm 1002 .
  • the calibrated tube signals may be conditioned 1000 prior to being provided as an input to the machine learning algorithm 1002 to normalize statistical properties such as mean and variance.
  • the second analysis machine learning algorithm 1002 is a neural network is trained to detect possible indications in the test data, while accounting for the tube regions to which data points in the test data correspond.
  • the neural network 1002 can, for example, be a convolutional neural network trained on historical or previously acquired test data having data points labelled according to whether or not they correspond to an indication, and labelled according to a tube region to which the data point corresponds.
  • the neural network 1002 can be configured to segment indications 1004 in the test data and output, for each of a plurality of data points in the test data, a classification corresponding to whether the data point does or does not correspond to an indication.
  • the output of 1004 can comprise a detection probability curve that provides, for each data point in the test data, a probability that the data point belongs to an indication.
  • the neural network can also be configured to output a status code 1006 indicating whether or not the segmentation was successful.
  • the segmented test data can subsequently be grouped in order to identify start-stop ranges in the test data corresponding to detected indications. More specifically, connected regions in the segmented test data can be identified, with such connected regions corresponding to sequences of data points likely corresponding to a same indication. A first data point in each connected region can indicate a start of a range of data points corresponding to an indication, and a last data point of each connected region can indicate an end of the range of data points corresponding to that indication. As can be appreciated, any suitable process can be applied to group sequences of data points in the test data segmented by the neural network.
  • the curves output by 1004 can be transformed into a sequence of start-stop ranges for all detected indications using a CCL algorithm 1008 .
  • the start-stop ranges can be fine-tuned and validated using measurements on the test data 1010 .
  • the result is a list of indications 1012 which includes detected indications and their corresponding data points.
  • the second neural network 1002 is configured to detect indications via semantic segmentation 1004 , it is appreciated that other configurations are possible.
  • the second neural network 1002 can comprise an object detection machine learning algorithm trained to build an object detection model such that the neural network 1002 is configured to directly output an array of detected indications, i.e. comprising an array of start positions and end positions, corresponding to the start and end of each indication.
  • the neural network 1002 can be a convolutional neural network trained on historical or previously acquired test data having groups of data points labelled according to an indication to which they correspond. This approach can have the advantage of enable the prediction of superposed indications.
  • An example of superposed indications can be a large corrosion (one large indication) with a small pit inside (another short indication that should be marked as superposed to the corrosion).
  • One possible embodiment of fine-tuning applicable to ECT signals can be based on the analysis of ECT differential channels extracted at positions corresponding to the segmented indication.
  • the extracted signals are complex (meaning, characterized by an amplitude and phase at each point). When pictured on a complex plane, the extracted signals should approximately form a FIG. 8 . Two examples are shown in FIGS. 13 A and 13 B .
  • the folds are identified by the head and tail of the vector.
  • An appropriate start-stop range allows to identify the folds.
  • a misplaced range or a range too small would not allow to identify the folds.
  • Examples of bad ranges can include misplaced ranges, as shown in the example of FIG. 13 D , and ranges that are too small, as shown in the Example of FIG. 13 E .
  • Misplaced or small start-stop ranges can be fine-tuned (enlarged or replaced) until the folds are properly detected in the signal. It is appreciated, however, that other embodiments of fine-tuning start-stop ranges are also possible.
  • an assisted indication classification 610 subprocess is provided to substantially automate the labelling and/or discarding of detected indications.
  • different methods can be applied in order to carry out the labelling and/or discarding of indications.
  • a rule-based method is applied in order to label each indication as one of a plurality of predefined indication types or as an indication to be discard, based on a plurality of configurable rules.
  • a decision tree is used to label indications based on rules relating to measurements in the test data associated with an indication.
  • the decision tree can be user-defined in that the user/analyst can configure the different rules and/or the structure of the decision tree based on requirements of the inspection being carried out.
  • FIG. 11 An exemplary user-configured decision tree 1100 for labelling an indication is shown in FIG. 11 .
  • the decision tree is configured to label the indication according to one of a plurality of predefined indications including a type 1 indication, a type 2 indication, a type 3 indication, or an indication to be discarded.
  • a position of the indication is evaluated.
  • the indication can be considered as being near a landmark, for example, if it is within a specified distance (distance in space, or distance in time between timecoded measurements) relative to a landmark. If the indication is near a landmark, it can be labelled or tagged as a type 1 indication. If the indication is not near a landmark, the indication can be evaluated at a second node in the decision tree.
  • a phase of the signals recorded in test data corresponding to the indication are evaluated. In particular, it is determined whether the phase is within a first defined range or a second defined range. If the phase is within a first define range of 0° to 40°, the indication can be labelled or tagged as a type 2 indication. If the phase is within a second defined range of 40° to 100°, the indication can be labelled or tagged as a type 3 indication. Otherwise, the indication can be discarded.
  • the decision tree can be configured by a user/analyst as needed depending on the inspection being carried out and the types of indications to be identified. For example, a user can define different thresholds/ranges when evaluating the position and phase of indications. A user can also define rules to evaluate different characteristics of the test data, such as amplitude and/or define additional rules or conditions to allow labelling additional types of indications.
  • subprocess 612 allows a user to assign one or more sizing curves to different types (or classes) of indications. Then, all indications classified according to that type can be automatically sized with these sizing curves.
  • step 208 of analyzing the testing data can be subject to a step of review and analysis 210 .
  • the subprocesses of step 210 can substantially correspond to those of step 208 , but can be conducted by a different analyst.
  • step 208 can be conducted by a first analyst via a first analysis device (such as a first analysis workstation), and step 210 can be conducted by a second analyst via a second analysis device (such as a second analysis workstation).
  • the second analyst device can be a different device than the first analyst device.
  • the analysis conducted by the first analyst can be validated and/or confirmed by the second analyst.
  • the analysis step 208 can be performed again and/or parameters thereof can be adjusted if there is a difference in the analysis between the first and second analysts.
  • an inspection report 212 can be generated and output.
  • the inspection report can list the indications located and identified in the testing data.
  • the positions of the indications can be specified relative to locations of the landmarks identified in the testing data.

Landscapes

  • Engineering & Computer Science (AREA)
  • Chemical & Material Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Engineering & Computer Science (AREA)
  • Immunology (AREA)
  • Biochemistry (AREA)
  • General Health & Medical Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Pathology (AREA)
  • Health & Medical Sciences (AREA)
  • Electrochemistry (AREA)
  • Chemical Kinetics & Catalysis (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Investigating Or Analyzing Materials By The Use Of Magnetic Means (AREA)
US17/975,314 2021-10-29 2022-10-27 System and method for automated acquisition and analysis of electromagnetic testing data Pending US20230314373A9 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/975,314 US20230314373A9 (en) 2021-10-29 2022-10-27 System and method for automated acquisition and analysis of electromagnetic testing data

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US202163263269P 2021-10-29 2021-10-29
US17/975,314 US20230314373A9 (en) 2021-10-29 2022-10-27 System and method for automated acquisition and analysis of electromagnetic testing data

Publications (2)

Publication Number Publication Date
US20230138975A1 US20230138975A1 (en) 2023-05-04
US20230314373A9 true US20230314373A9 (en) 2023-10-05

Family

ID=84044816

Family Applications (1)

Application Number Title Priority Date Filing Date
US17/975,314 Pending US20230314373A9 (en) 2021-10-29 2022-10-27 System and method for automated acquisition and analysis of electromagnetic testing data

Country Status (3)

Country Link
US (1) US20230314373A9 (de)
EP (1) EP4174486A1 (de)
CA (1) CA3180653A1 (de)

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9945814B2 (en) * 2014-07-18 2018-04-17 Westinghouse Electric Company Llc Total integrated tube analysis

Also Published As

Publication number Publication date
EP4174486A1 (de) 2023-05-03
US20230138975A1 (en) 2023-05-04
CA3180653A1 (en) 2023-04-29

Similar Documents

Publication Publication Date Title
JP6922539B2 (ja) 表面欠陥判定方法および表面欠陥検査装置
US6424151B2 (en) Method and apparatus for evaluation of eddy current testing signal
US20100312494A1 (en) Process and apparatus for testing a component using an omni-directional eddy current probe
KR20150140349A (ko) 반도체 검사 레시피 생성, 결함 리뷰 및 계측을 위한 적응적 샘플링
RU2591584C1 (ru) Способ оценки геометрических размеров дефектов стенки трубной секции и сварных швов по данным магнитного внутритрубного дефектоскопа с помощью универсальной нейросетевой модели, подходящей для дефектоскопов с различными диаметрами и магнитными системами
CN113344886A (zh) 晶圆表面缺陷检测方法和设备
US8387444B2 (en) Method of modeling steam generator and processing steam generator tube data of nuclear power plant
US9335296B2 (en) Systems and methods for steam generator tube analysis for detection of tube degradation
CN112528975A (zh) 工业质检方法、装置和计算机可读存储介质
Dong et al. Automatic inspection of aerospace welds using x-ray images
US9245399B2 (en) Media authentication
JP4616695B2 (ja) マルチセンサ信号異常検知装置および方法
JP2012533752A (ja) 原子力発電所の蒸気発生器細管の処理方法
CN116596889A (zh) 基于深度学习的x射线工业检测缺陷识别方法及系统
CN113554645B (zh) 基于wgan的工业异常检测方法和装置
Tian et al. Signal processing schemes for Eddy Current Testing of steam generator tubes of nuclear power plants
US20230314373A9 (en) System and method for automated acquisition and analysis of electromagnetic testing data
CN106932469A (zh) 一种基于涡流信号特征的热交换管缺陷检测方法
JP2004163288A (ja) 非破壊検査データ抽出装置及びこの装置を用いた収集システム
JP2023081845A (ja) 自動検査方法及び装置
Banerjee et al. Confidence metric for signal classification in non destructive evaluation
CN112801110B (zh) 轨道列车线阵相机图像畸变校正的目标检测方法及装置
Udpa et al. Automated analysis of eddy current signals in steam generator tube inspection
US20220253735A1 (en) Method for simulation assisted data generation and deep learning intelligence creation in non-destructive evaluation systems
JP4948238B2 (ja) 分布解析方法および装置、異常設備推定方法および装置、上記分布解析方法または異常設備推定方法をコンピュータに実行させるためのプログラム、並びに上記プログラムを記録したコンピュータ読み取り可能な記録媒体

Legal Events

Date Code Title Description
STPP Information on status: patent application and granting procedure in general

Free format text: DOCKETED NEW CASE - READY FOR EXAMINATION

AS Assignment

Owner name: EDDYFI CANADA INC., CANADA

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:EDDYFI NDT INC.;REEL/FRAME:062374/0990

Effective date: 20220301

Owner name: EDDYFI NDT INC., CANADA

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:MACKAY, PHILIPPE;GAUDREAULT, VINCENT;HARDY, FLORIAN;AND OTHERS;SIGNING DATES FROM 20211111 TO 20211201;REEL/FRAME:062368/0430