[2C6] Multi-model machine learning approach for automated data analysis of carbon fibre-reinforced polymer composites
E Mohseni¹, V Tunukovic¹, S McKnight¹, A Hifi¹, S G Pierce¹, R Vithanage¹, G Dobie¹, C MacLeod¹,
S Cochran¹, G Munro² and T O’Hare²
¹University of Strathclyde, UK
²Spirit AeroSystems, UK
The aerospace sector is increasingly embracing automation for sensor deployment and data collection in phased array ultrasonic testing (PAUT) of carbon fibre-reinforced polymers (CFRPs). While this shift enhances inspection speed and reliability, it also produces vast amounts of data that still require manual analysis, leading to inefficiencies and susceptibility to human error. This bottleneck is particularly pronounced in the production of newer aircraft, with CFRPs now comprising up to 50% of the total material weight. Traditional rule-based tools, such as signal thresholding, are sometimes used to assist manual analysis, but they often struggle with complex data patterns and high noise levels, requiring frequent manual adjustments that introduce inefficiencies. Artificial intelligence (AI)-based analysis tools offer improved defect detection and adaptability but face challenges in industrial adoption due to concerns about model transparency and trust. This study explores AI integration across three automation levels, from AI-assisted workflows to fully AI-driven supervisory roles. AI integration strategies for automation of PAUT inspection data interpretation were tested on two defective CFRP components used in aerospace and energy sectors. The experimental scans utilised a PAUT roller probe mounted on an industrial manipulator, mirroring real-world inspection processes. This set-up successfully detected 36 manufactured defects by combining supervised object detection on ultrasonic amplitude C-scans, unsupervised anomaly detection on ultrasonic B-scans and a self-supervised AI model for analysing full volumetric ultrasonic data. A faster region-based convolutional neural network, trained solely on simulated data to address data scarcity, was employed for object detection. Additionally, the anomaly detection model, implemented as a convolutional autoencoder, and the self-supervised AI model, designed as a forecasting tool for time-series data, were trained on pristine CFRP samples. Incorporating multiple AI models improved the F1 score by up to 17.2% over single-model approaches. The framework also integrates a human-in-the-loop mechanism, enhancing trust and adaptability within NDE workflows. Unlike manual analysis, which takes hours, this method processes data in just 94.03 s and 57.01 s for the two samples.
