[2A4] Ultrasound B-scan defect detection in carbon fibre-reinforced plastic composites with NDT machine learning algorithms
V Tunukovic¹,³, S McKnight¹, R Pyle¹, E Duernberger¹, M Vasilev¹, C Loukas¹, E Mohseni¹, S Pierce¹, G Dobiev¹, C MacLeod¹, S Cochran³ and T O’Hare²
¹University of Strathclyde, UK
²Spirit AeroSystems, UK
³FUSE CDT, UK
Carbon fibre-reinforced polymers (CFRPs) are lightweight materials that make up 50% of the structural weight in aircraft such as the Airbus A350 XWB and the Boeing 787 Dreamliner. Non-destructive evaluation (NDE) procedures are crucial during and after manufacturing to ensure the highest quality of important structures made from CFRPs, such as wing covers, stabilisers, fuselage and engine covers. Common NDE techniques include radiographic and eddy current testing, visual inspection and ultrasonic testing (UT). Manual UT inspections using single ultrasonic transducers or phased array ultrasonic testing (PAUT) arrays are laborious and heavily influenced by human operators. Automated NDE using industrial manipulators has accelerated the acquisition of ultrasonic data, but manual data interpretation still presents a bottleneck. An automated approach to data processing and interpretation would be beneficial, as data acquisition can take a few hours while data interpretation and quality report generation can take up to 6-8 hours. This paper proposes an automated deep learning (DL) algorithm based on autoencoder architectures that processes ultrasonic B-scan images. An unsupervised training approach was adopted to train the model to recognise anomalies in the data in the form of surface and subsurface defects. Data was captured from a series of CFRP coupons of various thicknesses, ply orientation configurations and surface finishes. Flat-bottomed holes that were 3 mm, 4 mm, 6 mm, 7 mm and 9 mm in size at depths of 1.5 mm, 3 mm, 4.5 mm, 6 mm and 7.5 mm were introduced in some of the CFRP samples.
Training/testing data was acquired with an experimental set-up that mimics the industrial process, with a KUKA KR90 industrial manipulator delivering an Olympus RollerFORM-5L64 to the surface of the test-piece. The array was excited in linear mode with an unfocused four-element sub-aperture, operating at a frequency of 5 MHz. The roller probe was controlled by a Peak NDT MP6 ultrasonic controller, while vertical movement of the KUKA KR90 was controlled with a force-torque unit. This resulted in the acquisition of 7500 individual B-scans. Several convolutional autoencoder models with varying latent space sizes were tested. Due to the nature of B-scans and the domain representation, the hyperbolic tangent function was used as activation to retain pixel values between –1 and 1. Furthermore, different dimensionality reductions were explored, such us max pooling with various square kernel sizes, max pooling with rectangular kernels and standard convolutional layers. The combination of L1 and L2 losses was used to evaluate the image reconstruction performance of the decoder. To prevent overfitting and improve convergence, the authors used the combination of batch normalisation layers, learning rate schedulers, momentum and weight decays for the Adam optimiser. This automated non-destructive testing scanning combined with a DL algorithm enabled swift and accurate analysis of ultrasonic B-scans. The algorithm completes analysis of a single B-scan frame of 1024 × 64 pixels in 13.6 ms. This approach shows a major improvement in the data interpretation time and defect detection success rate.
Training/testing data was acquired with an experimental set-up that mimics the industrial process, with a KUKA KR90 industrial manipulator delivering an Olympus RollerFORM-5L64 to the surface of the test-piece. The array was excited in linear mode with an unfocused four-element sub-aperture, operating at a frequency of 5 MHz. The roller probe was controlled by a Peak NDT MP6 ultrasonic controller, while vertical movement of the KUKA KR90 was controlled with a force-torque unit. This resulted in the acquisition of 7500 individual B-scans. Several convolutional autoencoder models with varying latent space sizes were tested. Due to the nature of B-scans and the domain representation, the hyperbolic tangent function was used as activation to retain pixel values between –1 and 1. Furthermore, different dimensionality reductions were explored, such us max pooling with various square kernel sizes, max pooling with rectangular kernels and standard convolutional layers. The combination of L1 and L2 losses was used to evaluate the image reconstruction performance of the decoder. To prevent overfitting and improve convergence, the authors used the combination of batch normalisation layers, learning rate schedulers, momentum and weight decays for the Adam optimiser. This automated non-destructive testing scanning combined with a DL algorithm enabled swift and accurate analysis of ultrasonic B-scans. The algorithm completes analysis of a single B-scan frame of 1024 × 64 pixels in 13.6 ms. This approach shows a major improvement in the data interpretation time and defect detection success rate.
