Increasing the Fidelity of Synthetic Training Data improves Crack Characterizations using Machine Learning
Tracks
Energy
| Thursday, October 24, 2024 |
| 8:30 AM - 9:00 AM |
| 209/210 - Technical Session |
Details
To validate the structural integrity of safety critical component, the nuclear and chemical sectors require improved crack detection via ultrasonic non-destructive evaluation (NDE), for example using in-situ A-scans. The application of machine and deep learning (ML/DL) methods to NDE is rapidly growing as computing power and parallelisation increase. However, the training datasets have a relative paucity of real defects and realistic inputs which, in turn, are causing major obstruction for in-situ crack characterisation. Finite element modelling, informed by real defect data (obtained from samples provided by industrial partners) and realistic transducer characteristics are used to generate a synthetic dataset suitable for application to ML and DL in the field. Throughout, the improved ML/DL is used to successfully characterise the length, angle and roughness of surface -breaking thermal fatigue cracks, inspected via pulse-echo transduction. Using this improved dataset, the crack charactersiation is improved by an order of magnitude compared to established methods. This presentation falls into two sections: finite element modelling of the cracks and ML for crack characterisation.
The datasets are generated using the finite element method (FEM) software Pogo. In Pogo, the standard transducer temporal input is improved by importing a measured transducer response. The transducer geometry is optimised to include mode conversion from the wedge (used for an angled beam shear wave), but also to be accurate and efficient for the repetitive training data simulations. The crack geometries are realistic for industrial applications (through consultation with industrial project partners) so ranges in length from 2mm to 6mm, and tilt angles up to 20 degrees. The crack roughness is taken from measured thermally fatigued cracks in stainless steel. Noise from the experiments has been characterised and added to the simulated A-scan time histories to enhance fidelity. A structured dataset of around 9000 A-scans was used to train the ML algorithm. Pogo utilises the GPU so can generate A-scans quickly and a large dataset can be generated is a reasonable timeframe (under a week to generate the scans).
The simulation results from Pogo are interpreted via a neural network which has been implemented in the python library Tensorflow. Modelled A-scan timeseries data is used for training, validation and testing of multiple neural network types. Specifically, 1D-Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and transformers have been used to estimate the relevant quantities. Currently, we note best test scores using a 1D-CNN for Mean Absolute Error (MAE) and Mean Square Error (MSE).
The dataset used is structured, paramters vary between the previously discussed limits of defect length and angle at constant intervals. An investigation into the random variation of these quantities is be done to assess model performance on unstructured data.
It has been found that concatenating A-scans of opposite angle but same length significantly increases neural network performance when estimating defect length. Hence, two individual neural networks are required for estimating the quantities investigated this far.
Eventually, when using the algorithms developed here, an NDE practitioner would be able to successfully characterise the length and angle of a surface- breaking crack using just A-scans.
The datasets are generated using the finite element method (FEM) software Pogo. In Pogo, the standard transducer temporal input is improved by importing a measured transducer response. The transducer geometry is optimised to include mode conversion from the wedge (used for an angled beam shear wave), but also to be accurate and efficient for the repetitive training data simulations. The crack geometries are realistic for industrial applications (through consultation with industrial project partners) so ranges in length from 2mm to 6mm, and tilt angles up to 20 degrees. The crack roughness is taken from measured thermally fatigued cracks in stainless steel. Noise from the experiments has been characterised and added to the simulated A-scan time histories to enhance fidelity. A structured dataset of around 9000 A-scans was used to train the ML algorithm. Pogo utilises the GPU so can generate A-scans quickly and a large dataset can be generated is a reasonable timeframe (under a week to generate the scans).
The simulation results from Pogo are interpreted via a neural network which has been implemented in the python library Tensorflow. Modelled A-scan timeseries data is used for training, validation and testing of multiple neural network types. Specifically, 1D-Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and transformers have been used to estimate the relevant quantities. Currently, we note best test scores using a 1D-CNN for Mean Absolute Error (MAE) and Mean Square Error (MSE).
The dataset used is structured, paramters vary between the previously discussed limits of defect length and angle at constant intervals. An investigation into the random variation of these quantities is be done to assess model performance on unstructured data.
It has been found that concatenating A-scans of opposite angle but same length significantly increases neural network performance when estimating defect length. Hence, two individual neural networks are required for estimating the quantities investigated this far.
Eventually, when using the algorithms developed here, an NDE practitioner would be able to successfully characterise the length and angle of a surface- breaking crack using just A-scans.
Speaker
James Gaffney
Academic
RCNDE
Increasing the Fidelity of Synthetic Training Data improves Crack Characterizations using Machine Learning
Presentation Description
To validate the structural integrity of safety critical component, the nuclear and chemical sectors require improved crack detection via ultrasonic non-destructive evaluation (NDE), for example using in-situ A-scans. The application of machine and deep learning (ML/DL) methods to NDE is rapidly growing as computing power and parallelisation increase. However, the training datasets have a relative paucity of real defects and realistic inputs which, in turn, are causing major obstruction for in-situ crack characterisation. Finite element modelling, informed by real defect data (obtained from samples provided by industrial partners) and realistic transducer characteristics are used to generate a synthetic dataset suitable for application to ML and DL in the field. Throughout, the improved ML/DL is used to successfully characterise the length, angle and roughness of surface -breaking thermal fatigue cracks, inspected via pulse-echo transduction. Using this improved dataset, the crack charactersiation is improved by an order of magnitude compared to established methods. This presentation falls into two sections: finite element modelling of the cracks and ML for crack characterisation.
The datasets are generated using the finite element method (FEM) software Pogo. In Pogo, the standard transducer temporal input is improved by importing a measured transducer response. The transducer geometry is optimised to include mode conversion from the wedge (used for an angled beam shear wave), but also to be accurate and efficient for the repetitive training data simulations. The crack geometries are realistic for industrial applications (through consultation with industrial project partners) so ranges in length from 2mm to 6mm, and tilt angles up to 20 degrees. The crack roughness is taken from measured thermally fatigued cracks in stainless steel. Noise from the experiments has been characterised and added to the simulated A-scan time histories to enhance fidelity. A structured dataset of around 9000 A-scans was used to train the ML algorithm. Pogo utilises the GPU so can generate A-scans quickly and a large dataset can be generated is a reasonable timeframe (under a week to generate the scans).
The simulation results from Pogo are interpreted via a neural network which has been implemented in the python library Tensorflow. Modelled A-scan timeseries data is used for training, validation and testing of multiple neural network types. Specifically, 1D-Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and transformers have been used to estimate the relevant quantities. Currently, we note best test scores using a 1D-CNN for Mean Absolute Error (MAE) and Mean Square Error (MSE).
The dataset used is structured, parameters vary between the previously discussed limits of defect length and angle at constant intervals. An investigation into the random variation of these quantities is be done to assess model performance on unstructured data.
It has been found that concatenating A-scans of opposite angle but same length significantly increases neural network performance when estimating defect length. Hence, two individual neural networks are required for estimating the quantities investigated this far.
Eventually, when using the algorithms developed here, an NDE practitioner would be able to successfully characterise the length and angle of a surface- breaking crack using just A-scans.
The datasets are generated using the finite element method (FEM) software Pogo. In Pogo, the standard transducer temporal input is improved by importing a measured transducer response. The transducer geometry is optimised to include mode conversion from the wedge (used for an angled beam shear wave), but also to be accurate and efficient for the repetitive training data simulations. The crack geometries are realistic for industrial applications (through consultation with industrial project partners) so ranges in length from 2mm to 6mm, and tilt angles up to 20 degrees. The crack roughness is taken from measured thermally fatigued cracks in stainless steel. Noise from the experiments has been characterised and added to the simulated A-scan time histories to enhance fidelity. A structured dataset of around 9000 A-scans was used to train the ML algorithm. Pogo utilises the GPU so can generate A-scans quickly and a large dataset can be generated is a reasonable timeframe (under a week to generate the scans).
The simulation results from Pogo are interpreted via a neural network which has been implemented in the python library Tensorflow. Modelled A-scan timeseries data is used for training, validation and testing of multiple neural network types. Specifically, 1D-Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and transformers have been used to estimate the relevant quantities. Currently, we note best test scores using a 1D-CNN for Mean Absolute Error (MAE) and Mean Square Error (MSE).
The dataset used is structured, parameters vary between the previously discussed limits of defect length and angle at constant intervals. An investigation into the random variation of these quantities is be done to assess model performance on unstructured data.
It has been found that concatenating A-scans of opposite angle but same length significantly increases neural network performance when estimating defect length. Hence, two individual neural networks are required for estimating the quantities investigated this far.
Eventually, when using the algorithms developed here, an NDE practitioner would be able to successfully characterise the length and angle of a surface- breaking crack using just A-scans.
Biography
James Gaffney completed his undergraduate and doctorate in acoustics. He workied for numerical method software company COMSOL, he simulated the acoustics of insects and designed sound absorbers. He works at the University of Liverpool training machine learning algorithms to characterise cracks for in-situ NDE.
Tom Beckingham studied Electrical Engineering as an undergraduate before writing his doctoral thesis in sensor technology, which focused on the application of machine learning to health data. Currently, he is a postdoctoral at the University of Liverpool where he works on simulations on computation fluid dynamics and machine learning analysis of crack data.
