Multi model machine learning approach for automated data analysis of carbon fiber reinforced polymer composites

The aerospace industry is increasingly shifting towards using automated solutions for sensor delivery and data acquisition for Ultrasonic Testing (UT) of Carbon Fibre Reinforced Polymers (CFRPs). While this transition has enabled faster and more reliable inspections, it generates large volumes of data in a short time, which are still analysed and interpreted manually, making the process lengthy and prone to human error [1]. This creates a bottleneck in the manufacturing and Non-Destructive Evaluation (NDE) workflows, particularly given the increasing use of CFRPs in flagship aircraft models by Airbus and Boeing, currently accounting for up to 50% of the total material weight [2, 3]. The manual NDE data analysis is sometimes paired with simple traditional rule-based tools such as signal thresholding. However, these tools often struggle to effectively manage complex data patterns or high noise levels, leading to unreliable defect detection. Additionally, they require frequent manual adjustments to set appropriate parameters for varying inspection conditions, which can be inefficient and error-prone in dynamic or fast-paced environments. In contrast, Artificial Intelligence (AI) analysis tools have demonstrated improvements over traditional methods, offering greater accuracy in defect detection and adaptability to higher variability within captured signals. However, their adoption in industrial settings remains limited due to challenges associated with model trust and their “black box” nature. Additionally, practical guidelines for implementing AI tools into NDE workflow are rarely discussed, motivating this work to explore various integration strategies across different automation levels. Three levels of automation were explored, ranging from basic AI-assisted workflows, where tools provide suggestions, to advanced applications where multiple AI models simultaneously process data in a comprehensive analysis, shifting human operators to a supervisory role.

Proposed strategies of AI integration into the NDE automation workflow were evaluated on inspection of two defective complex-geometry CFRP components, commonly used in aerospace and energy sectors for safety-critical structures such as aircraft fuselages and wind turbine blades. The experimental scans were conducted using a phased array UT roller probe mounted on an industrial manipulator, closely replicating industrial practices, and successfully identifying 36 manufactured defects through a combination of supervised object detection on ultrasonic amplitude C-scans, unsupervised anomaly detection on ultrasonic B-scans, and a self-supervised AI model for processing full volumetric ultrasonic data. Specifically, a Faster Region-based Convolutional Neural Network was used for object detection, trained exclusively on simulated data to mitigate data scarcity issues. Meanwhile, the anomaly detection model, implemented as a convolutional autoencoder, and the self-supervised AI model, designed as a forecasting model for time-series data, were both trained on pristine CFRP samples. This inclusion of multiple AI models led to an improvement of up to 17.2% in the F1 score compared to single-model approaches. Additionally, this framework enables integration into existing NDE workflows by incorporating a human-in-the-loop mechanism, improving trust in automation and allowing process customisation depending on specific application requirements. Unlike manual data analysis, which take hours for larger components, the proposed approach completes the analysis in 94.03 and 57.01 seconds for the two inspected samples, respectively.