Learning Defects From Aircraft NDT Data

Non-destructive evaluation of aircraft production is optimized and digitalized with Industry 4.0. The aircraft structures produced using fiber metal laminate are traditionally inspected using water-coupled ultrasound scans and manually evaluated.

October 1, 2023

German Aerospace Center (DLR), Augsburg, Germany

Ultrasonic Testing (UT) is a typical Non-destructive testing (NDT) method for examining the structural components for aircraft production. Manufacturing aircraft made of fiber metal laminates (FML) includes cascaded steps such as placement of aluminum, glass prepreg, adhesive, doublers, stringers, vacuum bagging and curing in an autoclave. Quality control (QC) is performed first at the layup of the component (without stringers) after curing and the quality assessment is visually evaluated. The manually performed examination of anomalies is very time-consuming. In addition, conducted NDT inspection using a manual UT phased array for Glass Reinforced (GLARE^®) FML of A380, it lacked the high capacity of data and additionally an evaluation software.

So, non-destructive evaluation (NDE) 4.0 helps streamline processes, increase quality and lower costs in aircraft production with an automated quality assurance (QA). Traditionally, the quality control of FML is performed by an experienced examiner after the final production of an aircraft structure. But, with the implementation of machine learning (ML) techniques, defects can be identified instantaneously to help the examiner. So, the primary motivation was to develop an automated QA in aircraft production by implementing a machine learning algorithm. The quality analysis process in the proposed method consists of pre analyzing the sensor data acquisition to classify the features according to the defects and good qualities. The proposed approach reduces the examiner’s workload, expensive repairs and manufacturing waste.

The proposed work can be vital in the automated offline-QA to scrutinize FML aircraft production and adapt to other aircraft materials like aluminum, thermoplastic fiber and carbon fiber reinforced plastic (CFRP). The proposed research aims: to understand and prepare ultrasonic scans of aircraft FML (raw data) provided by the aircraft industry and pre-process data (convert raw data to images) to gain feasibility for the proposed method. Additionally, implement embedded machine learning classifiers with image feature extraction techniques to achieve the best defect detection rate and further interpret industry-based certification process to evaluate this approach.

Machine learning is a subset of artificial intelligence (AI), dealing with data acquired from sensors for learning the data-generating distribution. There are three primary techniques: supervised learning - data needs to be labelled (each data point tagged to belong to a particular class) for training, mainly used for classification (predicts discrete labels) and regression (predicts a continuous quantity). Next, unsupervised learning - requires no data labels for training; dimensionality reduction and clustering are the two significant methodologies. Following is reinforcement learning - the agent (training) sends an action (a move causing change) to the environment (real or virtual world) and in-return environment sends the state and its reward (evaluation of the action, either positive or negative) for the agent; real-time decisions and gaming models are its prototypes. Additionally, semi supervised learning is a combination of supervised and unsupervised learning methodologies.

Supervised learning examples include support vector machine (SVM), Decision Trees, Random Forest (RF), κ-Nearest Neighbor (κ-NN), Naïve Bayes, Linear Discriminant Analysis (LDA) and Logistic Regression. The Fuzzy C-means (FCM) κ-means and principal component analysis (PCA) are a few state-of-the-art unsupervised learning techniques. SVM predicts classes based on an optimal hyperplane creating margins to find similar features from each class and classifies them together. Decision Trees predict a class by learning the decision rules from the data features of that class. Random Forest combines the outcome of multiple Decision Trees into a prediction. κ-Nearest Neighbor predicts using the proximity of κ nearest data points for classification.

Naïve Bayes classifies based on the probability of data points applying Bayes’ theorem. Using Fisher’s algorithm, LDA finds a linear combination of data features to characterize different classes. Logistic regression finds the probability of an event occurring, such as voted or no vote, based on the data variables. Fuzzy C-means is similar to κ-means but is a soft clustering where a data point can belong to one or more clusters. κ-means is a hard clustering that partitions data points into κ clusters, each belonging to one cluster with the nearest mean value. PCA reduces data dimensionality and increases its interpretability with less information loss.

This work was performed by Navya Prakash, Dorothea Nieberl, Monika Mayer and Alfons Schuster for the German Aerospace Center. For more information, download the Technical Support Package (free white paper) below.