Improving Failure Prediction in Aircraft Fastener Assembly Using Synthetic Data in Imbalanced Datasets

TL;DR

This work tackles failure prediction in aircraft fastener collar assembly under severe data imbalance by evaluating deep learning models directly on raw multivariate time-series from force/torque and rotation data. Through hyperparameter optimization and 10-fold cross-validation, it compares MLP, CNN, LSTM, Transformer, and Vision Transformer architectures across original, balanced, and synthetic datasets generated via SMOTE, with and without class weighting. The results show that SMOTE-based oversampling and balanced data improve jammed recall while maintaining high mounted precision, with ViT on balanced data delivering the best jammed detection (0.92±0.08), albeit with high compute needs; a lighter Transformer without rotation on synthetic data also offers strong performance for constrained environments. The study advances a task-focused evaluation framework that moves beyond accuracy, highlights data-imbalance strategies, and outlines practical directions for synthetic data generation and anomaly detection to enhance safety and efficiency in aircraft assembly.

Abstract

Automating aircraft manufacturing still relies heavily on human labor due to the complexity of the assembly processes and customization requirements. One key challenge is achieving precise positioning, especially for large aircraft structures, where errors can lead to substantial maintenance costs or part rejection. Existing solutions often require costly hardware or lack flexibility. Used in aircraft by the thousands, threaded fasteners, e.g., screws, bolts, and collars, are traditionally executed by fixed-base robots and usually have problems in being deployed in the mentioned manufacturing sites. This paper emphasizes the importance of error detection and classification for efficient and safe assembly of threaded fasteners, especially aeronautical collars. Safe assembly of threaded fasteners is paramount since acquiring sufficient data for training deep learning models poses challenges due to the rarity of failure cases and imbalanced datasets. The paper addresses this by proposing techniques like class weighting and data augmentation, specifically tailored for temporal series data, to improve classification performance. Furthermore, the paper introduces a novel problem-modeling approach, emphasizing metrics relevant to collar assembly rather than solely focusing on accuracy. This tailored approach enhances the models' capability to handle the challenges of threaded fastener assembly effectively.

Figures (5)

• Figure 1: The collected data from the three phases (Alignment, Backspin, and Screwing) are pre-processed and used as input for several deep-learning models. The outputs are split between the three possible outcomes: Mounted, Not mounted, and Jammed.
• Figure 2: TP are used to account for accuracy
• Figure 3: General architecture of the deep learning models evaluated in this work. MLP processes flattened time-series input through dense layers; CNN applies 1D convolutions to extract local temporal patterns; LSTM captures sequential dependencies across time steps; Transformer leverages self-attention for long-range dependencies; and ViT uses a temporal patch-based approach with convolutional token embedding and Transformer blocks.
• Figure 4: Hyperparameter optimization pipeline to obtain the tuned models. The test data is split from the train right after Data Normalization. The model optimization block runs for each model in different processes.
• Figure 5: Mean precision and standard error for the mounted class obtained during hyperparameter optimization of each model for (a) without rotation and (b) with rotation data. They are evaluated across three dataset variants: original, balanced, and synthetic, as well as a class weighing technique. Results reflect 10-fold cross-validation averaged over 100 optimization trials.