Table of Content
- The Dawn of Digital Technologies in NDE
- Types of Digital Technologies in NDE
- The Key Components of AI in NDE
- The Role of AI in Data Analysis and Interpretation in NDE
- Improved Inspection as A Result of AI Integration
- AI Integration with Robotics in NDT
- The Path Ahead
The Dawn of Digital Technologies in NDE
Non-destructive Evaluation dates back to when simple visual and tactile inspections were the norm. Over the years, NDE has been marked by milestones like the development of Industrial Radiography in the 1920s and the introduction of ultrasonic testing in the 1940s.
These advancements represented significant leaps in detecting internal and hidden material flaws. Traditional NDE methods, such as dye penetrant testing, Magnetic Particle Testing, and visual inspection, have been the backbone of quality assurance for decades.
However, these methods often rely heavily on the skill and experience of the operator, leading to the potential for human error. Traditional methods can be time-consuming and less effective in analysing complex data or inspecting large-scale structures, prompting a need for more advanced, efficient techniques.
The emergence of digital technologies opened new possibilities for more precise, efficient, and comprehensive evaluations. Digital technologies not only improved the accuracy but also significantly reduced the time required for inspections, leading to cost-effective and scalable solutions.
Types of Digital Technologies in NDE
The different digital technologies that are generally applied in NDT Inspections include:
1. Advanced Sensors
Modern NDE extensively uses advanced sensors offering higher precision and sensitivity that can detect flaws or changes in materials that were previously undetectable with traditional methods.
2. Internet of Things (IoT)
IoT connects various inspection devices and sensors, enabling the collection and transmission of real-time data. AI enables a continuous data flow and real-time fault detection when integrated with IoT devices.
3. Cloud Computing
Cloud platforms are instrumental in storing and managing the vast amounts of data generated by NDE Processes. They provide scalable storage solutions and powerful computing capabilities for data analysis.
4. Artificial Intelligence (AI) and Machine Learning (ML)
AI and ML algorithms analyse NDE data, identifying patterns and anomalies in predictive maintenance and automated defect recognition.
5. Digital Imaging and 3D Modelling
Techniques like Digital Radiography and 3D modelling offer detailed visualisations of internal structures.
Image Credit: Visiconsult
6. Drones and Robotics
These are used for remote NDE, allowing access to difficult or hazardous locations through its various sensors and cameras.
7. Augmented Reality (AR) and Virtual Reality (VR)
Augmented Reality (AR) and VR are used for training purposes and to simulate inspection scenarios, providing inspectors with an immersive experience.
8. Digital Data in NDE Processes
Digital data and NDE integration involve data collection, analysis ,and interpretation providing accurate and instantaneous insights.
The Key Components of AI in NDE
Artificial Intelligence (AI) has become a cornerstone in modern Non-destructive Evaluation, a field critical for ensuring the integrity and safety of structures and materials in various industries. AI encompasses several components, each contributing uniquely to enhancing NDE processes. These components include:
1. Machine Learning (ML):
ML algorithms and technologies have been incorporated into NDT in numerous ways, including:
I. Supervised Learning
Algorithms like Support Vector Machines (SVMs) and Random Forests can classify defects like cracks or corrosion in datasets labelled by NDT inspectors. These algorithms use pre-labelled training data to map features like signal amplitudes from Ultrasonic Testing, to defect types, improving classification accuracy in weld inspections or composite material assessments.
II. Unsupervised Learning
Techniques like k-means clustering identify hidden patterns in unlabelled data from ultrasonic testing (UT) or Eddy Current Testing (ECT). By segmenting raw sensor data into clusters, this method detects anomalies in pipelines or pressure vessels without prior knowledge of defect signatures.
II. Reinforcement Learning:
Used for adaptive inspection planning, and optimising sensor paths in robotic NDT systems. AI agents learn optimal scanning trajectories for automated crawlers in complex geometries, reducing inspection time in aerospace or energy sectors.
ML models process time-series data from Phased Array Ultrasonic Testing (PAUT) or acoustic emission (AE) sensors, detecting sub-surface flaws in real-time. Algorithms analyse waveform features like time-of-flight and amplitude to identify cracks or voids in steel welds or turbine blades during live inspections.
Anomaly detection algorithms flag deviations in baseline datasets from defect-free materials, which are critical for aerospace component inspections. Statistical models can compare live eddy current signals with historical data to detect early-stage fatigue cracks in aircraft fuselages.
2. Deep Learning (DL):
Neural network architecture has been involved in NDT in different forms, such as:
I. Convolutional Neural Networks (CNNs)
CNNs process greyscale images layer-by-layer, identifying micron-level voids in welded joints more accurately than manual radiography interpretation.
II. Autoencoders
Autoencoders compress and reconstruct thermal imaging data from infrared cameras, highlighting delaminations in composite materials.
These networks reduce noise in thermograms, isolating temperature anomalies caused by disbonds in carbon-fiber-reinforced polymers (CFRPs).
DL models are trained on annotated datasets from Digital Radiography (DR) systems or laser shearography, achieving sub-millimetre defect resolution. The training for this involves thousands of labelled images of cracks, corrosion, or inclusions, enabling precise defect sizing in pipeline girth welds.
For low-latency inference, frameworks like TensorFlow Lite can deploy pre-trained models on edge devices like FLIR thermal cameras. Optimised DL models run locally on industrial cameras in high-speed manufacturing lines.
In the automated interpretation of B-scan and C-scan data in PAUT for pipeline inspections, DL algorithms convert ultrasonic A-scan signals into 2D/3D visualisations, which help pinpoint crack depths in Oil and Gas Transmission pipelines.
3. Computer Vision
Here, image processing techniques are employed, as follows:
I. Semantic Segmentation
Tools like OpenCV and MATLAB can partition digital radiography images into defect or non-defect regions using U-Net architectures.
The pixel-wise classification can isolate corrosion patches or cracks in the X-ray images of castings.
II. Optical Flow Analysis
This helps track surface deformation in digital image correlation (DIC) systems for fatigue testing. This technique quantifies strain distribution in aluminium alloys under cyclic loading by monitoring pixel displacement sequences.
High-resolution cameras paired with structured light projection capture 3D surface topography in Turbine Blade Inspection. The projected fringe patterns analysed via phase-shifting algorithms can reveal sub-surface defects in Jet Engine Inspection.
Automated optical inspection (AOI) systems integrate YOLO (You Only Look Once) algorithms to classify corrosion in real-time drone-based inspections.
Drones with 4K cameras and YOLOv5 detect rust spots on offshore wind turbines, generating geotagged defect maps for maintenance crews.
Robotic crawlers with embedded vision systems perform autonomous Visual Testing (VT) in confined spaces. Crawlers use stereo cameras and LED illumination to conduct 360° visual inspections of boiler tubes or storage tanks, transmitting HD video to inspectors.
4. Natural Language Processing (NLP):
The applications of NLP in NDT are:
I. Automated Report Generation
Transformer models like BERT extract key findings from inspection logs, streamlining compliance documentation. NLP algorithms parse unstructured text from magnetic particle inspection (MPI) reports, auto-populating fields like defect location and severity in PDF templates.
II. Semantic Search
NLP-powered databases retrieve historical defect records linked to specific NDT Methods e.g., magnetic particle testing. Engineers can search natural language terms for relevant testing records from decade-old reports.
Named Entity Recognition (NER) identifies material grades or defect types in PDF reports from Penetrant Testing (PT) or magnetic particle inspection (MPI). NER models tag entities in scanned documents, structuring data for predictive maintenance analytics.
5. Robotics & Autonomous Systems
The applications of AI-driven NDT robots include:
I. Path Planning Algorithms
Simultaneous Localisation and Mapping (SLAM) guides ultrasonic crawlers through complex geometries like in those of pressure vessels. SLAM-enabled robots can be used to build 3D maps of confined spaces, avoiding obstacles while maintaining consistent probe coupling for UT Thickness Measurements.
II. Collaborative Robots (Cobots)
These can be used to perform automated tap testing on aircraft fuselages as they are equipped with LiDAR and force-torque sensors. Cobots use acoustic feedback from controlled impacts to detect disbonds in composite panels, which is similar to manual hammer testing, but with repeatable precision.
Omron TM Series Cobots are integrated with AI-based acoustic emission sensors for real-time composite material assessment. These cobots can help analyse stress-wave signals during load tests, identifying fibre fractures in wind turbine blade prototypes.
6. Data Fusion & Multi-Modal Analysis:
The technologies are used in NDT as follows:
I. Sensor Fusion
Combines data from electromagnetic acoustic transducers (EMATs), thermography, and UT probes using Kalman filters for holistic defect characterisation. The fusion algorithms can correlate thermal anomalies with ultrasonic echo patterns to distinguish between corrosion and slag inclusions in petrochemical tanks.
II. Digital Twin Integration
AI correlates NDT data with finite element analysis (FEA) simulations to predict remaining asset life. Live inspection data from PAUT scans keeps digital twins of offshore platforms updated, forecasting crack propagation rates under their operational loads.
The fusion of eddy current and ultrasonic data enhances crack sizing accuracy in aerospace alloy inspections. Eddy current detects surface cracks, while UT measures depth. This provides combined sizing for FAA-compliant airframe assessments.
7. Edge AI for Real-Time Processing:
The technologies used here include:
NVIDIA Jetson or Google Coral boards process Time-of-flight Diffraction (TOFD) data on-device, minimising cloud latency. Edge AI chips perform real-time TOFD signal analysis in pipeline inspection gauges (PIGs), enabling immediate weld defect alerts during inline inspections.
Image Credit: Wikipedia
Quantised Neural Networks (QNNs) reduce the computational load for portable thickness gauges performing real-time corrosion mapping. 8-bit quantised models run efficiently on low-power microcontrollers, allowing handheld devices to generate 2D corrosion profiles in refinery pipelines.
These technologies can provide instant defect detection in automated inline inspection (ILI) systems for oil and gas pipelines. Edge AI can analyse Magnetic Flux Leakage (MFL) data from ILI tools, identifying pitting corrosion or dents without delaying pipeline operations.
When incorporated into NDT workflows, AI technology can help industries achieve higher accuracy, faster inspections, and predictive maintenance significantly benefitting asset integrity management.
The Role of AI in Data Analysis and Interpretation in NDE
When it comes to AI, one of the first applications we could think of is to optimise data interpretation and analysis in NDT processes. It has been extensively implemented across the industry for its many merits, which include:
1. Enhanced Accuracy
I. Phased Array Ultrasonic Testing (PAUT)
3D convolutional neural networks (CNNs) can process full matrix capture (FMC) data from phased array systems to detect micron-scale fatigue cracks in aluminium airframe rivets. This reduces false positives from grain noise compared to manual A-scan analysis.
II. Digital Radiography (DR)
U-Net architectures segment reconstructed CT scans which can isolate gas porosity in investment-cast turbine blades with high pixel accuracy. e.g. Nikon XT H 225 ST industrial CT scanners.
III. Acoustic Emission (AE)
Wavelet scattering transforms + SVM classifiers can differentiate between benign weld slag inclusions and critical hydrogen cracks in AE datasets. e.g. PAC Micro-II
2. Efficiency Gains
I. Automated Magnetic Flux Leakage (MFL)
Federated learning trains YOLOv7 models on crawler data from multiple oil companies to classify corrosion patches in API 5L pipelines, cutting analysis time from hours to seconds.
II. Laser Shearography
Optical flow algorithms in Electronic Speckle Pattern Interferometry (ESPI) systems like the Dantec Dynamics Q-450 can quantify strain anomalies in CFRP aircraft panels approximately ten times faster than manual fringe pattern interpretation.
3. Predictive Maintenance
I. Eddy Current Testing (ECT)
Long Short Term Memory (LSTM) networks can predict tube plugging schedules by analysing tube testing impedance trends in ASTM A179 condenser tubes under cyclic thermal stress. e.g. Evident NORTEC 600. This drastically reduces unplanned downtime in nuclear power plants.
AI detecting the flaw
Image Credit: Sentin
I. Time-of-Flight Diffraction (TOFD)
Physics-informed neural networks (PINNs) correlate PAUT diffraction signals with ASME BPVC crack growth models to forecast SCC propagation rates in X70 pipeline girth welds. e.g. Zetec TOPAZ
4. Workflow Integration:
I. AI-Assisted Reporting
Bidirectional Encoder Representations from Transformers (BERT)-based NLP like in Zeiss INSPECT™ can auto-generate ASME Section V-compliant reports from PAUT datasets, considerably reducing inspector admin time.
II. Human-in-the-Loop Validation
Level III inspectors can refine random forest defect classifiers using Label Studio platforms.
Improved Inspection as A Result of AI Integration
Combining AI and NDT automates and simplifies an otherwise repetitive and iterative process. This includes merits like the following:
1. Real-Time Data Analysis
Digital technologies enable real-time data collection and analysis, providing immediate insights during inspections. AI systems analyse historical and real-time data from sensors and inspection tools to monitor the condition of equipment.
2. Enhanced Data Accuracy and Reliability
Digital tools reduce human error, enhancing the accuracy and reliability of NDE data. Machine Learning models identify patterns that deviate from the norm, helping signal potential problems.
3. Data Management and Accessibility
Cloud computing and IoT ensure that data is easily accessible and well-managed. AI can assess the risk of future failures by prioritising maintenance tasks based on the severity and likelihood of potential issues.
4. Image Recognition and Analysis
Image recognition using AI is used in Non-destructive Testing in techniques like radiography, ultrasonics, and visual inspection to generate images. Analysing these for anomalies or defects requires high precision, especially in complex or noisy environments. AI algorithms can be trained on vast datasets to recognise patterns associated with defects.
AI in UT can help analyse waveforms to detect and characterise flaws or material properties. By training AI algorithms on a diverse range of radiographic images, they can learn to recognise and evaluate radiographic images. In Computed Tomography, AI aids in reconstructing and analysing complex 3D models, offering a non-invasive yet comprehensive evaluation of internal features. AI algorithms can also process thermal images to identify irregularities in thermography. Beyond detection, AI evaluates the severity of defects, aiding in decision-making about necessary repairs or replacements.
AI Integration with Robotics in NDT
Advanced NDT robotics, merged with AI has upgraded the NDT experience globally. This is mainly because of intensive technological research and development, that have led to the following technologies:
I. Crawler Robots with Phased Array Ultrasonic Testing (PAUT)
Tracked robotic platforms like the Inuktun Versatrax 300 integrate PAUT probes and EMATs to perform automated weld inspections in confined spaces like pressure vessels. AI-driven path-planning algorithms like Rapidly-exploring Random Trees (RRT) optimise scan paths around complex geometries, providing full coverage in circumferential weld inspection.
II. Aerial Drones with Thermography
DJI Matrice 300 RTK drones equipped with FLIR T1K thermal cameras use AI-based image stitching to create three-dimensional heat maps of refinery columns. Embedded YOLOv7 models detect hotspots indicative of Corrosion Under Insulation (CUI) in real time.
III. Underwater ROVs for Subsea Inspections
Underwater ROVs deploy pulsed eddy current (PEC) arrays and Laser Shearography to assess subsea pipeline integrity. AI processes time-domain reflectometry (TDR) data to differentiate between marine growth and wall loss defects.
IV. SLAM (Simultaneous Localisation and Mapping) with LiDAR-IMU Fusion
SLAM Robots combining these technologies can generate 3D maps of boiler internals.
V. Edge AI for Real-Time Data Processing
Inspection crawlers can run quantised neural networks (QNNs) to analyse full matrix capture (FMC) ultrasonic data, instantly flagging stress corrosion cracking (SCC) in petrochemical assets.
1. Acoustic Emission Testing (AET)
I. Piezoelectric Sensor Arrays with Deep Learning
Digital AE Systems with wideband sensors capture high-frequency stress waves. Convolutional Neural Networks (CNNs) trained on Mel-frequency cepstral coefficients (MFCCs) classify emission patterns into fracture modes in nuclear reactor pressure vessels.
II. Time-Frequency Analysis with Wavelet Transforms
Continuous Wavelet Transform (CWT) algorithms decompose AE signals into time-frequency domains, enabling AI to distinguish between benign plastic deformation and critical hydrogen-induced cracking in offshore wind turbine foundations.
III. Data Acquisition Software with Autoencoders
Unsupervised variational autoencoders (VAEs) compress AE waveforms, isolating anomalies in composite overwrapped pressure vessels (COPVs) by comparing latent-space distributions against baseline datasets.
2. Guided Wave Testing (GWT)
I. EMAT-Based Long-Range Ultrasonic Testing (LRUT)
AI analyses dispersion curves and mode-converted echoes in GWT using finite element method (FEM)-augmented datasets to detect circumferential cracks in austenitic stainless-steel pipes.
II. Multi-Modal Signal Fusion
Recurrent Neural Networks (RNNs) process fused data from piezoelectric wafer active sensors (PWAS) and fiber Bragg grating (FBG) arrays to suppress false calls from weld caps in buried pipelines.
III. Defect Sizing with Inverse Problem Solvers
Physics-informed neural networks (PINNs) iteratively solve wave propagation equations to estimate crack depth in pipelines from reflected wave amplitudes.
3. AI-Optimised Heat Exchanger Tube Inspection
I. Multi-Frequency Eddy Current (MFEC)
ECT flaw detector systems use double driver probes to excite multiple frequencies (10 Hz–10 MHz). Support Vector Machines (SVMs) classify impedance plane trajectories, differentiating between pitting corrosion and baffle cuts in titanium condenser tubes.
II. Remote Field Testing (RFT)
Magnetic saturation RFT probes coupled with Isolation Forest algorithms can detect ID/OD (inner diameter- outer diameter) defects in ferritic heat exchanger tubes, even through non-magnetic deposits.
III. Digital Twins
Digital Twin platforms correlate eddy current data with computational fluid dynamics (CFD) simulations to predict erosion-corrosion rates in U-tube bundles under multiphase flow conditions.
4. AI-Driven Root Cause Analysis (RCA)
I. SHAP (SHapley Additive exPlanations):
Extreme Gadget Boosting (XGBoost) models trained on historical PAUT, ECT, and AET data use SHAP values to quantify feature importance, identifying contributing factors like residual stress or chloride exposure to intergranular stress corrosion cracking (IGSCC) in stainless steel.
II. Fractography with Computer Vision:
Scanning Electron Microscope (SEM) images of fracture surfaces are analysed via Mask R-CNN architectures to classify failure mechanisms e.g., fatigue striations, cleavage facets etc. in aircraft landing gear components.
III. Automated Inspection Software:
These can graph neural networks (GNNs) and map defect morphology to material certificates and process histories, tracing forging defects back to specific heat-treatment batches.
RVI of Solar Panels
The Path Ahead
While the benefits of integrating AI into practical NDE Applications are substantial, the challenge is navigating investment, training, and data security.
The role of research and development in advancing AI applications within NDE is fundamental to achieving leaps in innovation and addressing current challenges. R&D efforts are also driving the development of new inspection methodologies that are precise and aligned with global sustainability goals. By focusing on creating eco-friendly & sustainable practices researchers are setting new benchmarks for responsible NDE.
AI has the potential to revolutionise NDE data analysis by automating the interpretation of large datasets, which traditionally required intensive human labour. These advancements have been crucial in industries where the margin for error is minimal, and the cost of failure can be significant.
